Datasets:
Dataset Viewer
id
stringclasses 179
values | question
stringlengths 8.75k
85.9k
| answer
dict |
|---|---|---|
1909.00694 | Please extract a concise answer without any additional explanation for the following question based on the given text.
Question: What is the seed lexicon?
Context: <<<Title>>>
Minimally Supervised Learning of Affective Events Using Discourse Relations
<<<Abstract>>>
Recognizing affective events that trigger positive or negative sentiment has a wide range of natural language processing applications but remains a challenging problem mainly because the polarity of an event is not necessarily predictable from its constituent words. In this paper, we propose to propagate affective polarity using discourse relations. Our method is simple and only requires a very small seed lexicon and a large raw corpus. Our experiments using Japanese data show that our method learns affective events effectively without manually labeled data. It also improves supervised learning results when labeled data are small.
<<</Abstract>>>
<<<Introduction>>>
Affective events BIBREF0 are events that typically affect people in positive or negative ways. For example, getting money and playing sports are usually positive to the experiencers; catching cold and losing one's wallet are negative. Understanding affective events is important to various natural language processing (NLP) applications such as dialogue systems BIBREF1, question-answering systems BIBREF2, and humor recognition BIBREF3. In this paper, we work on recognizing the polarity of an affective event that is represented by a score ranging from $-1$ (negative) to 1 (positive).
Learning affective events is challenging because, as the examples above suggest, the polarity of an event is not necessarily predictable from its constituent words. Combined with the unbounded combinatorial nature of language, the non-compositionality of affective polarity entails the need for large amounts of world knowledge, which can hardly be learned from small annotated data.
In this paper, we propose a simple and effective method for learning affective events that only requires a very small seed lexicon and a large raw corpus. As illustrated in Figure FIGREF1, our key idea is that we can exploit discourse relations BIBREF4 to efficiently propagate polarity from seed predicates that directly report one's emotions (e.g., “to be glad” is positive). Suppose that events $x_1$ are $x_2$ are in the discourse relation of Cause (i.e., $x_1$ causes $x_2$). If the seed lexicon suggests $x_2$ is positive, $x_1$ is also likely to be positive because it triggers the positive emotion. The fact that $x_2$ is known to be negative indicates the negative polarity of $x_1$. Similarly, if $x_1$ and $x_2$ are in the discourse relation of Concession (i.e., $x_2$ in spite of $x_1$), the reverse of $x_2$'s polarity can be propagated to $x_1$. Even if $x_2$'s polarity is not known in advance, we can exploit the tendency of $x_1$ and $x_2$ to be of the same polarity (for Cause) or of the reverse polarity (for Concession) although the heuristic is not exempt from counterexamples. We transform this idea into objective functions and train neural network models that predict the polarity of a given event.
We trained the models using a Japanese web corpus. Given the minimum amount of supervision, they performed well. In addition, the combination of annotated and unannotated data yielded a gain over a purely supervised baseline when labeled data were small.
<<</Introduction>>>
<<<Related Work>>>
Learning affective events is closely related to sentiment analysis. Whereas sentiment analysis usually focuses on the polarity of what are described (e.g., movies), we work on how people are typically affected by events. In sentiment analysis, much attention has been paid to compositionality. Word-level polarity BIBREF5, BIBREF6, BIBREF7 and the roles of negation and intensification BIBREF8, BIBREF6, BIBREF9 are among the most important topics. In contrast, we are more interested in recognizing the sentiment polarity of an event that pertains to commonsense knowledge (e.g., getting money and catching cold).
Label propagation from seed instances is a common approach to inducing sentiment polarities. While BIBREF5 and BIBREF10 worked on word- and phrase-level polarities, BIBREF0 dealt with event-level polarities. BIBREF5 and BIBREF10 linked instances using co-occurrence information and/or phrase-level coordinations (e.g., “$A$ and $B$” and “$A$ but $B$”). We shift our scope to event pairs that are more complex than phrase pairs, and consequently exploit discourse connectives as event-level counterparts of phrase-level conjunctions.
BIBREF0 constructed a network of events using word embedding-derived similarities. Compared with this method, our discourse relation-based linking of events is much simpler and more intuitive.
Some previous studies made use of document structure to understand the sentiment. BIBREF11 proposed a sentiment-specific pre-training strategy using unlabeled dialog data (tweet-reply pairs). BIBREF12 proposed a method of building a polarity-tagged corpus (ACP Corpus). They automatically gathered sentences that had positive or negative opinions utilizing HTML layout structures in addition to linguistic patterns. Our method depends only on raw texts and thus has wider applicability.
<<</Related Work>>>
<<<Proposed Method>>>
<<<Polarity Function>>>
Our goal is to learn the polarity function $p(x)$, which predicts the sentiment polarity score of an event $x$. We approximate $p(x)$ by a neural network with the following form:
${\rm Encoder}$ outputs a vector representation of the event $x$. ${\rm Linear}$ is a fully-connected layer and transforms the representation into a scalar. ${\rm tanh}$ is the hyperbolic tangent and transforms the scalar into a score ranging from $-1$ to 1. In Section SECREF21, we consider two specific implementations of ${\rm Encoder}$.
<<</Polarity Function>>>
<<<Discourse Relation-Based Event Pairs>>>
Our method requires a very small seed lexicon and a large raw corpus. We assume that we can automatically extract discourse-tagged event pairs, $(x_{i1}, x_{i2})$ ($i=1, \cdots $) from the raw corpus. We refer to $x_{i1}$ and $x_{i2}$ as former and latter events, respectively. As shown in Figure FIGREF1, we limit our scope to two discourse relations: Cause and Concession.
The seed lexicon consists of positive and negative predicates. If the predicate of an extracted event is in the seed lexicon and does not involve complex phenomena like negation, we assign the corresponding polarity score ($+1$ for positive events and $-1$ for negative events) to the event. We expect the model to automatically learn complex phenomena through label propagation. Based on the availability of scores and the types of discourse relations, we classify the extracted event pairs into the following three types.
<<<AL (Automatically Labeled Pairs)>>>
The seed lexicon matches (1) the latter event but (2) not the former event, and (3) their discourse relation type is Cause or Concession. If the discourse relation type is Cause, the former event is given the same score as the latter. Likewise, if the discourse relation type is Concession, the former event is given the opposite of the latter's score. They are used as reference scores during training.
<<</AL (Automatically Labeled Pairs)>>>
<<<CA (Cause Pairs)>>>
The seed lexicon matches neither the former nor the latter event, and their discourse relation type is Cause. We assume the two events have the same polarities.
<<</CA (Cause Pairs)>>>
<<<CO (Concession Pairs)>>>
The seed lexicon matches neither the former nor the latter event, and their discourse relation type is Concession. We assume the two events have the reversed polarities.
<<</CO (Concession Pairs)>>>
<<</Discourse Relation-Based Event Pairs>>>
<<<Loss Functions>>>
Using AL, CA, and CO data, we optimize the parameters of the polarity function $p(x)$. We define a loss function for each of the three types of event pairs and sum up the multiple loss functions.
We use mean squared error to construct loss functions. For the AL data, the loss function is defined as:
where $x_{i1}$ and $x_{i2}$ are the $i$-th pair of the AL data. $r_{i1}$ and $r_{i2}$ are the automatically-assigned scores of $x_{i1}$ and $x_{i2}$, respectively. $N_{\rm AL}$ is the total number of AL pairs, and $\lambda _{\rm AL}$ is a hyperparameter.
For the CA data, the loss function is defined as:
$y_{i1}$ and $y_{i2}$ are the $i$-th pair of the CA pairs. $N_{\rm CA}$ is the total number of CA pairs. $\lambda _{\rm CA}$ and $\mu $ are hyperparameters. The first term makes the scores of the two events closer while the second term prevents the scores from shrinking to zero.
The loss function for the CO data is defined analogously:
The difference is that the first term makes the scores of the two events distant from each other.
<<</Loss Functions>>>
<<</Proposed Method>>>
<<<Experiments>>>
<<<Dataset>>>
<<<AL, CA, and CO>>>
As a raw corpus, we used a Japanese web corpus that was compiled through the procedures proposed by BIBREF13. To extract event pairs tagged with discourse relations, we used the Japanese dependency parser KNP and in-house postprocessing scripts BIBREF14. KNP used hand-written rules to segment each sentence into what we conventionally called clauses (mostly consecutive text chunks), each of which contained one main predicate. KNP also identified the discourse relations of event pairs if explicit discourse connectives BIBREF4 such as “ので” (because) and “のに” (in spite of) were present. We treated Cause/Reason (原因・理由) and Condition (条件) in the original tagset BIBREF15 as Cause and Concession (逆接) as Concession, respectively. Here is an example of event pair extraction.
. 重大な失敗を犯したので、仕事をクビになった。
Because [I] made a serious mistake, [I] got fired.
From this sentence, we extracted the event pair of “重大な失敗を犯す” ([I] make a serious mistake) and “仕事をクビになる” ([I] get fired), and tagged it with Cause.
We constructed our seed lexicon consisting of 15 positive words and 15 negative words, as shown in Section SECREF27. From the corpus of about 100 million sentences, we obtained 1.4 millions event pairs for AL, 41 millions for CA, and 6 millions for CO. We randomly selected subsets of AL event pairs such that positive and negative latter events were equal in size. We also sampled event pairs for each of CA and CO such that it was five times larger than AL. The results are shown in Table TABREF16.
<<</AL, CA, and CO>>>
<<<ACP (ACP Corpus)>>>
We used the latest version of the ACP Corpus BIBREF12 for evaluation. It was used for (semi-)supervised training as well. Extracted from Japanese websites using HTML layouts and linguistic patterns, the dataset covered various genres. For example, the following two sentences were labeled positive and negative, respectively:
. 作業が楽だ。
The work is easy.
. 駐車場がない。
There is no parking lot.
Although the ACP corpus was originally constructed in the context of sentiment analysis, we found that it could roughly be regarded as a collection of affective events. We parsed each sentence and extracted the last clause in it. The train/dev/test split of the data is shown in Table TABREF19.
The objective function for supervised training is:
where $v_i$ is the $i$-th event, $R_i$ is the reference score of $v_i$, and $N_{\rm ACP}$ is the number of the events of the ACP Corpus.
To optimize the hyperparameters, we used the dev set of the ACP Corpus. For the evaluation, we used the test set of the ACP Corpus. The model output was classified as positive if $p(x) > 0$ and negative if $p(x) \le 0$.
<<</ACP (ACP Corpus)>>>
<<</Dataset>>>
<<<Model Configurations>>>
As for ${\rm Encoder}$, we compared two types of neural networks: BiGRU and BERT. GRU BIBREF16 is a recurrent neural network sequence encoder. BiGRU reads an input sequence forward and backward and the output is the concatenation of the final forward and backward hidden states.
BERT BIBREF17 is a pre-trained multi-layer bidirectional Transformer BIBREF18 encoder. Its output is the final hidden state corresponding to the special classification tag ([CLS]). For the details of ${\rm Encoder}$, see Sections SECREF30.
We trained the model with the following four combinations of the datasets: AL, AL+CA+CO (two proposed models), ACP (supervised), and ACP+AL+CA+CO (semi-supervised). The corresponding objective functions were: $\mathcal {L}_{\rm AL}$, $\mathcal {L}_{\rm AL} + \mathcal {L}_{\rm CA} + \mathcal {L}_{\rm CO}$, $\mathcal {L}_{\rm ACP}$, and $\mathcal {L}_{\rm ACP} + \mathcal {L}_{\rm AL} + \mathcal {L}_{\rm CA} + \mathcal {L}_{\rm CO}$.
<<</Model Configurations>>>
<<<Results and Discussion>>>
Table TABREF23 shows accuracy. As the Random baseline suggests, positive and negative labels were distributed evenly. The Random+Seed baseline made use of the seed lexicon and output the corresponding label (or the reverse of it for negation) if the event's predicate is in the seed lexicon. We can see that the seed lexicon itself had practically no impact on prediction.
The models in the top block performed considerably better than the random baselines. The performance gaps with their (semi-)supervised counterparts, shown in the middle block, were less than 7%. This demonstrates the effectiveness of discourse relation-based label propagation.
Comparing the model variants, we obtained the highest score with the BiGRU encoder trained with the AL+CA+CO dataset. BERT was competitive but its performance went down if CA and CO were used in addition to AL. We conjecture that BERT was more sensitive to noises found more frequently in CA and CO.
Contrary to our expectations, supervised models (ACP) outperformed semi-supervised models (ACP+AL+CA+CO). This suggests that the training set of 0.6 million events is sufficiently large for training the models. For comparison, we trained the models with a subset (6,000 events) of the ACP dataset. As the results shown in Table TABREF24 demonstrate, our method is effective when labeled data are small.
The result of hyperparameter optimization for the BiGRU encoder was as follows:
As the CA and CO pairs were equal in size (Table TABREF16), $\lambda _{\rm CA}$ and $\lambda _{\rm CO}$ were comparable values. $\lambda _{\rm CA}$ was about one-third of $\lambda _{\rm CO}$, and this indicated that the CA pairs were noisier than the CO pairs. A major type of CA pairs that violates our assumption was in the form of “$\textit {problem}_{\text{negative}}$ causes $\textit {solution}_{\text{positive}}$”:
. (悪いところがある, よくなるように努力する)
(there is a bad point, [I] try to improve [it])
The polarities of the two events were reversed in spite of the Cause relation, and this lowered the value of $\lambda _{\rm CA}$.
Some examples of model outputs are shown in Table TABREF26. The first two examples suggest that our model successfully learned negation without explicit supervision. Similarly, the next two examples differ only in voice but the model correctly recognized that they had opposite polarities. The last two examples share the predicate “落とす" (drop) and only the objects are different. The second event “肩を落とす" (lit. drop one's shoulders) is an idiom that expresses a disappointed feeling. The examples demonstrate that our model correctly learned non-compositional expressions.
<<</Results and Discussion>>>
<<</Experiments>>>
<<<Conclusion>>>
In this paper, we proposed to use discourse relations to effectively propagate polarities of affective events from seeds. Experiments show that, even with a minimal amount of supervision, the proposed method performed well.
Although event pairs linked by discourse analysis are shown to be useful, they nevertheless contain noises. Adding linguistically-motivated filtering rules would help improve the performance.
<<</Conclusion>>>
<<</Title>>>
| {
"references": [
"seed lexicon consists of positive and negative predicates"
],
"type": "extractive"
} |
1909.00694 | Please extract a concise answer without any additional explanation for the following question based on the given text.
Question: What are labels available in dataset for supervision?
Context: <<<Title>>>
Minimally Supervised Learning of Affective Events Using Discourse Relations
<<<Abstract>>>
Recognizing affective events that trigger positive or negative sentiment has a wide range of natural language processing applications but remains a challenging problem mainly because the polarity of an event is not necessarily predictable from its constituent words. In this paper, we propose to propagate affective polarity using discourse relations. Our method is simple and only requires a very small seed lexicon and a large raw corpus. Our experiments using Japanese data show that our method learns affective events effectively without manually labeled data. It also improves supervised learning results when labeled data are small.
<<</Abstract>>>
<<<Introduction>>>
Affective events BIBREF0 are events that typically affect people in positive or negative ways. For example, getting money and playing sports are usually positive to the experiencers; catching cold and losing one's wallet are negative. Understanding affective events is important to various natural language processing (NLP) applications such as dialogue systems BIBREF1, question-answering systems BIBREF2, and humor recognition BIBREF3. In this paper, we work on recognizing the polarity of an affective event that is represented by a score ranging from $-1$ (negative) to 1 (positive).
Learning affective events is challenging because, as the examples above suggest, the polarity of an event is not necessarily predictable from its constituent words. Combined with the unbounded combinatorial nature of language, the non-compositionality of affective polarity entails the need for large amounts of world knowledge, which can hardly be learned from small annotated data.
In this paper, we propose a simple and effective method for learning affective events that only requires a very small seed lexicon and a large raw corpus. As illustrated in Figure FIGREF1, our key idea is that we can exploit discourse relations BIBREF4 to efficiently propagate polarity from seed predicates that directly report one's emotions (e.g., “to be glad” is positive). Suppose that events $x_1$ are $x_2$ are in the discourse relation of Cause (i.e., $x_1$ causes $x_2$). If the seed lexicon suggests $x_2$ is positive, $x_1$ is also likely to be positive because it triggers the positive emotion. The fact that $x_2$ is known to be negative indicates the negative polarity of $x_1$. Similarly, if $x_1$ and $x_2$ are in the discourse relation of Concession (i.e., $x_2$ in spite of $x_1$), the reverse of $x_2$'s polarity can be propagated to $x_1$. Even if $x_2$'s polarity is not known in advance, we can exploit the tendency of $x_1$ and $x_2$ to be of the same polarity (for Cause) or of the reverse polarity (for Concession) although the heuristic is not exempt from counterexamples. We transform this idea into objective functions and train neural network models that predict the polarity of a given event.
We trained the models using a Japanese web corpus. Given the minimum amount of supervision, they performed well. In addition, the combination of annotated and unannotated data yielded a gain over a purely supervised baseline when labeled data were small.
<<</Introduction>>>
<<<Related Work>>>
Learning affective events is closely related to sentiment analysis. Whereas sentiment analysis usually focuses on the polarity of what are described (e.g., movies), we work on how people are typically affected by events. In sentiment analysis, much attention has been paid to compositionality. Word-level polarity BIBREF5, BIBREF6, BIBREF7 and the roles of negation and intensification BIBREF8, BIBREF6, BIBREF9 are among the most important topics. In contrast, we are more interested in recognizing the sentiment polarity of an event that pertains to commonsense knowledge (e.g., getting money and catching cold).
Label propagation from seed instances is a common approach to inducing sentiment polarities. While BIBREF5 and BIBREF10 worked on word- and phrase-level polarities, BIBREF0 dealt with event-level polarities. BIBREF5 and BIBREF10 linked instances using co-occurrence information and/or phrase-level coordinations (e.g., “$A$ and $B$” and “$A$ but $B$”). We shift our scope to event pairs that are more complex than phrase pairs, and consequently exploit discourse connectives as event-level counterparts of phrase-level conjunctions.
BIBREF0 constructed a network of events using word embedding-derived similarities. Compared with this method, our discourse relation-based linking of events is much simpler and more intuitive.
Some previous studies made use of document structure to understand the sentiment. BIBREF11 proposed a sentiment-specific pre-training strategy using unlabeled dialog data (tweet-reply pairs). BIBREF12 proposed a method of building a polarity-tagged corpus (ACP Corpus). They automatically gathered sentences that had positive or negative opinions utilizing HTML layout structures in addition to linguistic patterns. Our method depends only on raw texts and thus has wider applicability.
<<</Related Work>>>
<<<Proposed Method>>>
<<<Polarity Function>>>
Our goal is to learn the polarity function $p(x)$, which predicts the sentiment polarity score of an event $x$. We approximate $p(x)$ by a neural network with the following form:
${\rm Encoder}$ outputs a vector representation of the event $x$. ${\rm Linear}$ is a fully-connected layer and transforms the representation into a scalar. ${\rm tanh}$ is the hyperbolic tangent and transforms the scalar into a score ranging from $-1$ to 1. In Section SECREF21, we consider two specific implementations of ${\rm Encoder}$.
<<</Polarity Function>>>
<<<Discourse Relation-Based Event Pairs>>>
Our method requires a very small seed lexicon and a large raw corpus. We assume that we can automatically extract discourse-tagged event pairs, $(x_{i1}, x_{i2})$ ($i=1, \cdots $) from the raw corpus. We refer to $x_{i1}$ and $x_{i2}$ as former and latter events, respectively. As shown in Figure FIGREF1, we limit our scope to two discourse relations: Cause and Concession.
The seed lexicon consists of positive and negative predicates. If the predicate of an extracted event is in the seed lexicon and does not involve complex phenomena like negation, we assign the corresponding polarity score ($+1$ for positive events and $-1$ for negative events) to the event. We expect the model to automatically learn complex phenomena through label propagation. Based on the availability of scores and the types of discourse relations, we classify the extracted event pairs into the following three types.
<<<AL (Automatically Labeled Pairs)>>>
The seed lexicon matches (1) the latter event but (2) not the former event, and (3) their discourse relation type is Cause or Concession. If the discourse relation type is Cause, the former event is given the same score as the latter. Likewise, if the discourse relation type is Concession, the former event is given the opposite of the latter's score. They are used as reference scores during training.
<<</AL (Automatically Labeled Pairs)>>>
<<<CA (Cause Pairs)>>>
The seed lexicon matches neither the former nor the latter event, and their discourse relation type is Cause. We assume the two events have the same polarities.
<<</CA (Cause Pairs)>>>
<<<CO (Concession Pairs)>>>
The seed lexicon matches neither the former nor the latter event, and their discourse relation type is Concession. We assume the two events have the reversed polarities.
<<</CO (Concession Pairs)>>>
<<</Discourse Relation-Based Event Pairs>>>
<<<Loss Functions>>>
Using AL, CA, and CO data, we optimize the parameters of the polarity function $p(x)$. We define a loss function for each of the three types of event pairs and sum up the multiple loss functions.
We use mean squared error to construct loss functions. For the AL data, the loss function is defined as:
where $x_{i1}$ and $x_{i2}$ are the $i$-th pair of the AL data. $r_{i1}$ and $r_{i2}$ are the automatically-assigned scores of $x_{i1}$ and $x_{i2}$, respectively. $N_{\rm AL}$ is the total number of AL pairs, and $\lambda _{\rm AL}$ is a hyperparameter.
For the CA data, the loss function is defined as:
$y_{i1}$ and $y_{i2}$ are the $i$-th pair of the CA pairs. $N_{\rm CA}$ is the total number of CA pairs. $\lambda _{\rm CA}$ and $\mu $ are hyperparameters. The first term makes the scores of the two events closer while the second term prevents the scores from shrinking to zero.
The loss function for the CO data is defined analogously:
The difference is that the first term makes the scores of the two events distant from each other.
<<</Loss Functions>>>
<<</Proposed Method>>>
<<<Experiments>>>
<<<Dataset>>>
<<<AL, CA, and CO>>>
As a raw corpus, we used a Japanese web corpus that was compiled through the procedures proposed by BIBREF13. To extract event pairs tagged with discourse relations, we used the Japanese dependency parser KNP and in-house postprocessing scripts BIBREF14. KNP used hand-written rules to segment each sentence into what we conventionally called clauses (mostly consecutive text chunks), each of which contained one main predicate. KNP also identified the discourse relations of event pairs if explicit discourse connectives BIBREF4 such as “ので” (because) and “のに” (in spite of) were present. We treated Cause/Reason (原因・理由) and Condition (条件) in the original tagset BIBREF15 as Cause and Concession (逆接) as Concession, respectively. Here is an example of event pair extraction.
. 重大な失敗を犯したので、仕事をクビになった。
Because [I] made a serious mistake, [I] got fired.
From this sentence, we extracted the event pair of “重大な失敗を犯す” ([I] make a serious mistake) and “仕事をクビになる” ([I] get fired), and tagged it with Cause.
We constructed our seed lexicon consisting of 15 positive words and 15 negative words, as shown in Section SECREF27. From the corpus of about 100 million sentences, we obtained 1.4 millions event pairs for AL, 41 millions for CA, and 6 millions for CO. We randomly selected subsets of AL event pairs such that positive and negative latter events were equal in size. We also sampled event pairs for each of CA and CO such that it was five times larger than AL. The results are shown in Table TABREF16.
<<</AL, CA, and CO>>>
<<<ACP (ACP Corpus)>>>
We used the latest version of the ACP Corpus BIBREF12 for evaluation. It was used for (semi-)supervised training as well. Extracted from Japanese websites using HTML layouts and linguistic patterns, the dataset covered various genres. For example, the following two sentences were labeled positive and negative, respectively:
. 作業が楽だ。
The work is easy.
. 駐車場がない。
There is no parking lot.
Although the ACP corpus was originally constructed in the context of sentiment analysis, we found that it could roughly be regarded as a collection of affective events. We parsed each sentence and extracted the last clause in it. The train/dev/test split of the data is shown in Table TABREF19.
The objective function for supervised training is:
where $v_i$ is the $i$-th event, $R_i$ is the reference score of $v_i$, and $N_{\rm ACP}$ is the number of the events of the ACP Corpus.
To optimize the hyperparameters, we used the dev set of the ACP Corpus. For the evaluation, we used the test set of the ACP Corpus. The model output was classified as positive if $p(x) > 0$ and negative if $p(x) \le 0$.
<<</ACP (ACP Corpus)>>>
<<</Dataset>>>
<<<Model Configurations>>>
As for ${\rm Encoder}$, we compared two types of neural networks: BiGRU and BERT. GRU BIBREF16 is a recurrent neural network sequence encoder. BiGRU reads an input sequence forward and backward and the output is the concatenation of the final forward and backward hidden states.
BERT BIBREF17 is a pre-trained multi-layer bidirectional Transformer BIBREF18 encoder. Its output is the final hidden state corresponding to the special classification tag ([CLS]). For the details of ${\rm Encoder}$, see Sections SECREF30.
We trained the model with the following four combinations of the datasets: AL, AL+CA+CO (two proposed models), ACP (supervised), and ACP+AL+CA+CO (semi-supervised). The corresponding objective functions were: $\mathcal {L}_{\rm AL}$, $\mathcal {L}_{\rm AL} + \mathcal {L}_{\rm CA} + \mathcal {L}_{\rm CO}$, $\mathcal {L}_{\rm ACP}$, and $\mathcal {L}_{\rm ACP} + \mathcal {L}_{\rm AL} + \mathcal {L}_{\rm CA} + \mathcal {L}_{\rm CO}$.
<<</Model Configurations>>>
<<<Results and Discussion>>>
Table TABREF23 shows accuracy. As the Random baseline suggests, positive and negative labels were distributed evenly. The Random+Seed baseline made use of the seed lexicon and output the corresponding label (or the reverse of it for negation) if the event's predicate is in the seed lexicon. We can see that the seed lexicon itself had practically no impact on prediction.
The models in the top block performed considerably better than the random baselines. The performance gaps with their (semi-)supervised counterparts, shown in the middle block, were less than 7%. This demonstrates the effectiveness of discourse relation-based label propagation.
Comparing the model variants, we obtained the highest score with the BiGRU encoder trained with the AL+CA+CO dataset. BERT was competitive but its performance went down if CA and CO were used in addition to AL. We conjecture that BERT was more sensitive to noises found more frequently in CA and CO.
Contrary to our expectations, supervised models (ACP) outperformed semi-supervised models (ACP+AL+CA+CO). This suggests that the training set of 0.6 million events is sufficiently large for training the models. For comparison, we trained the models with a subset (6,000 events) of the ACP dataset. As the results shown in Table TABREF24 demonstrate, our method is effective when labeled data are small.
The result of hyperparameter optimization for the BiGRU encoder was as follows:
As the CA and CO pairs were equal in size (Table TABREF16), $\lambda _{\rm CA}$ and $\lambda _{\rm CO}$ were comparable values. $\lambda _{\rm CA}$ was about one-third of $\lambda _{\rm CO}$, and this indicated that the CA pairs were noisier than the CO pairs. A major type of CA pairs that violates our assumption was in the form of “$\textit {problem}_{\text{negative}}$ causes $\textit {solution}_{\text{positive}}$”:
. (悪いところがある, よくなるように努力する)
(there is a bad point, [I] try to improve [it])
The polarities of the two events were reversed in spite of the Cause relation, and this lowered the value of $\lambda _{\rm CA}$.
Some examples of model outputs are shown in Table TABREF26. The first two examples suggest that our model successfully learned negation without explicit supervision. Similarly, the next two examples differ only in voice but the model correctly recognized that they had opposite polarities. The last two examples share the predicate “落とす" (drop) and only the objects are different. The second event “肩を落とす" (lit. drop one's shoulders) is an idiom that expresses a disappointed feeling. The examples demonstrate that our model correctly learned non-compositional expressions.
<<</Results and Discussion>>>
<<</Experiments>>>
<<<Conclusion>>>
In this paper, we proposed to use discourse relations to effectively propagate polarities of affective events from seeds. Experiments show that, even with a minimal amount of supervision, the proposed method performed well.
Although event pairs linked by discourse analysis are shown to be useful, they nevertheless contain noises. Adding linguistically-motivated filtering rules would help improve the performance.
<<</Conclusion>>>
<<</Title>>>
| {
"references": [
"negative,positive"
],
"type": "extractive"
} |
1909.00694 | Please extract a concise answer without any additional explanation for the following question based on the given text.
Question: How large is raw corpus used for training?
Context: <<<Title>>>
Minimally Supervised Learning of Affective Events Using Discourse Relations
<<<Abstract>>>
Recognizing affective events that trigger positive or negative sentiment has a wide range of natural language processing applications but remains a challenging problem mainly because the polarity of an event is not necessarily predictable from its constituent words. In this paper, we propose to propagate affective polarity using discourse relations. Our method is simple and only requires a very small seed lexicon and a large raw corpus. Our experiments using Japanese data show that our method learns affective events effectively without manually labeled data. It also improves supervised learning results when labeled data are small.
<<</Abstract>>>
<<<Introduction>>>
Affective events BIBREF0 are events that typically affect people in positive or negative ways. For example, getting money and playing sports are usually positive to the experiencers; catching cold and losing one's wallet are negative. Understanding affective events is important to various natural language processing (NLP) applications such as dialogue systems BIBREF1, question-answering systems BIBREF2, and humor recognition BIBREF3. In this paper, we work on recognizing the polarity of an affective event that is represented by a score ranging from $-1$ (negative) to 1 (positive).
Learning affective events is challenging because, as the examples above suggest, the polarity of an event is not necessarily predictable from its constituent words. Combined with the unbounded combinatorial nature of language, the non-compositionality of affective polarity entails the need for large amounts of world knowledge, which can hardly be learned from small annotated data.
In this paper, we propose a simple and effective method for learning affective events that only requires a very small seed lexicon and a large raw corpus. As illustrated in Figure FIGREF1, our key idea is that we can exploit discourse relations BIBREF4 to efficiently propagate polarity from seed predicates that directly report one's emotions (e.g., “to be glad” is positive). Suppose that events $x_1$ are $x_2$ are in the discourse relation of Cause (i.e., $x_1$ causes $x_2$). If the seed lexicon suggests $x_2$ is positive, $x_1$ is also likely to be positive because it triggers the positive emotion. The fact that $x_2$ is known to be negative indicates the negative polarity of $x_1$. Similarly, if $x_1$ and $x_2$ are in the discourse relation of Concession (i.e., $x_2$ in spite of $x_1$), the reverse of $x_2$'s polarity can be propagated to $x_1$. Even if $x_2$'s polarity is not known in advance, we can exploit the tendency of $x_1$ and $x_2$ to be of the same polarity (for Cause) or of the reverse polarity (for Concession) although the heuristic is not exempt from counterexamples. We transform this idea into objective functions and train neural network models that predict the polarity of a given event.
We trained the models using a Japanese web corpus. Given the minimum amount of supervision, they performed well. In addition, the combination of annotated and unannotated data yielded a gain over a purely supervised baseline when labeled data were small.
<<</Introduction>>>
<<<Related Work>>>
Learning affective events is closely related to sentiment analysis. Whereas sentiment analysis usually focuses on the polarity of what are described (e.g., movies), we work on how people are typically affected by events. In sentiment analysis, much attention has been paid to compositionality. Word-level polarity BIBREF5, BIBREF6, BIBREF7 and the roles of negation and intensification BIBREF8, BIBREF6, BIBREF9 are among the most important topics. In contrast, we are more interested in recognizing the sentiment polarity of an event that pertains to commonsense knowledge (e.g., getting money and catching cold).
Label propagation from seed instances is a common approach to inducing sentiment polarities. While BIBREF5 and BIBREF10 worked on word- and phrase-level polarities, BIBREF0 dealt with event-level polarities. BIBREF5 and BIBREF10 linked instances using co-occurrence information and/or phrase-level coordinations (e.g., “$A$ and $B$” and “$A$ but $B$”). We shift our scope to event pairs that are more complex than phrase pairs, and consequently exploit discourse connectives as event-level counterparts of phrase-level conjunctions.
BIBREF0 constructed a network of events using word embedding-derived similarities. Compared with this method, our discourse relation-based linking of events is much simpler and more intuitive.
Some previous studies made use of document structure to understand the sentiment. BIBREF11 proposed a sentiment-specific pre-training strategy using unlabeled dialog data (tweet-reply pairs). BIBREF12 proposed a method of building a polarity-tagged corpus (ACP Corpus). They automatically gathered sentences that had positive or negative opinions utilizing HTML layout structures in addition to linguistic patterns. Our method depends only on raw texts and thus has wider applicability.
<<</Related Work>>>
<<<Proposed Method>>>
<<<Polarity Function>>>
Our goal is to learn the polarity function $p(x)$, which predicts the sentiment polarity score of an event $x$. We approximate $p(x)$ by a neural network with the following form:
${\rm Encoder}$ outputs a vector representation of the event $x$. ${\rm Linear}$ is a fully-connected layer and transforms the representation into a scalar. ${\rm tanh}$ is the hyperbolic tangent and transforms the scalar into a score ranging from $-1$ to 1. In Section SECREF21, we consider two specific implementations of ${\rm Encoder}$.
<<</Polarity Function>>>
<<<Discourse Relation-Based Event Pairs>>>
Our method requires a very small seed lexicon and a large raw corpus. We assume that we can automatically extract discourse-tagged event pairs, $(x_{i1}, x_{i2})$ ($i=1, \cdots $) from the raw corpus. We refer to $x_{i1}$ and $x_{i2}$ as former and latter events, respectively. As shown in Figure FIGREF1, we limit our scope to two discourse relations: Cause and Concession.
The seed lexicon consists of positive and negative predicates. If the predicate of an extracted event is in the seed lexicon and does not involve complex phenomena like negation, we assign the corresponding polarity score ($+1$ for positive events and $-1$ for negative events) to the event. We expect the model to automatically learn complex phenomena through label propagation. Based on the availability of scores and the types of discourse relations, we classify the extracted event pairs into the following three types.
<<<AL (Automatically Labeled Pairs)>>>
The seed lexicon matches (1) the latter event but (2) not the former event, and (3) their discourse relation type is Cause or Concession. If the discourse relation type is Cause, the former event is given the same score as the latter. Likewise, if the discourse relation type is Concession, the former event is given the opposite of the latter's score. They are used as reference scores during training.
<<</AL (Automatically Labeled Pairs)>>>
<<<CA (Cause Pairs)>>>
The seed lexicon matches neither the former nor the latter event, and their discourse relation type is Cause. We assume the two events have the same polarities.
<<</CA (Cause Pairs)>>>
<<<CO (Concession Pairs)>>>
The seed lexicon matches neither the former nor the latter event, and their discourse relation type is Concession. We assume the two events have the reversed polarities.
<<</CO (Concession Pairs)>>>
<<</Discourse Relation-Based Event Pairs>>>
<<<Loss Functions>>>
Using AL, CA, and CO data, we optimize the parameters of the polarity function $p(x)$. We define a loss function for each of the three types of event pairs and sum up the multiple loss functions.
We use mean squared error to construct loss functions. For the AL data, the loss function is defined as:
where $x_{i1}$ and $x_{i2}$ are the $i$-th pair of the AL data. $r_{i1}$ and $r_{i2}$ are the automatically-assigned scores of $x_{i1}$ and $x_{i2}$, respectively. $N_{\rm AL}$ is the total number of AL pairs, and $\lambda _{\rm AL}$ is a hyperparameter.
For the CA data, the loss function is defined as:
$y_{i1}$ and $y_{i2}$ are the $i$-th pair of the CA pairs. $N_{\rm CA}$ is the total number of CA pairs. $\lambda _{\rm CA}$ and $\mu $ are hyperparameters. The first term makes the scores of the two events closer while the second term prevents the scores from shrinking to zero.
The loss function for the CO data is defined analogously:
The difference is that the first term makes the scores of the two events distant from each other.
<<</Loss Functions>>>
<<</Proposed Method>>>
<<<Experiments>>>
<<<Dataset>>>
<<<AL, CA, and CO>>>
As a raw corpus, we used a Japanese web corpus that was compiled through the procedures proposed by BIBREF13. To extract event pairs tagged with discourse relations, we used the Japanese dependency parser KNP and in-house postprocessing scripts BIBREF14. KNP used hand-written rules to segment each sentence into what we conventionally called clauses (mostly consecutive text chunks), each of which contained one main predicate. KNP also identified the discourse relations of event pairs if explicit discourse connectives BIBREF4 such as “ので” (because) and “のに” (in spite of) were present. We treated Cause/Reason (原因・理由) and Condition (条件) in the original tagset BIBREF15 as Cause and Concession (逆接) as Concession, respectively. Here is an example of event pair extraction.
. 重大な失敗を犯したので、仕事をクビになった。
Because [I] made a serious mistake, [I] got fired.
From this sentence, we extracted the event pair of “重大な失敗を犯す” ([I] make a serious mistake) and “仕事をクビになる” ([I] get fired), and tagged it with Cause.
We constructed our seed lexicon consisting of 15 positive words and 15 negative words, as shown in Section SECREF27. From the corpus of about 100 million sentences, we obtained 1.4 millions event pairs for AL, 41 millions for CA, and 6 millions for CO. We randomly selected subsets of AL event pairs such that positive and negative latter events were equal in size. We also sampled event pairs for each of CA and CO such that it was five times larger than AL. The results are shown in Table TABREF16.
<<</AL, CA, and CO>>>
<<<ACP (ACP Corpus)>>>
We used the latest version of the ACP Corpus BIBREF12 for evaluation. It was used for (semi-)supervised training as well. Extracted from Japanese websites using HTML layouts and linguistic patterns, the dataset covered various genres. For example, the following two sentences were labeled positive and negative, respectively:
. 作業が楽だ。
The work is easy.
. 駐車場がない。
There is no parking lot.
Although the ACP corpus was originally constructed in the context of sentiment analysis, we found that it could roughly be regarded as a collection of affective events. We parsed each sentence and extracted the last clause in it. The train/dev/test split of the data is shown in Table TABREF19.
The objective function for supervised training is:
where $v_i$ is the $i$-th event, $R_i$ is the reference score of $v_i$, and $N_{\rm ACP}$ is the number of the events of the ACP Corpus.
To optimize the hyperparameters, we used the dev set of the ACP Corpus. For the evaluation, we used the test set of the ACP Corpus. The model output was classified as positive if $p(x) > 0$ and negative if $p(x) \le 0$.
<<</ACP (ACP Corpus)>>>
<<</Dataset>>>
<<<Model Configurations>>>
As for ${\rm Encoder}$, we compared two types of neural networks: BiGRU and BERT. GRU BIBREF16 is a recurrent neural network sequence encoder. BiGRU reads an input sequence forward and backward and the output is the concatenation of the final forward and backward hidden states.
BERT BIBREF17 is a pre-trained multi-layer bidirectional Transformer BIBREF18 encoder. Its output is the final hidden state corresponding to the special classification tag ([CLS]). For the details of ${\rm Encoder}$, see Sections SECREF30.
We trained the model with the following four combinations of the datasets: AL, AL+CA+CO (two proposed models), ACP (supervised), and ACP+AL+CA+CO (semi-supervised). The corresponding objective functions were: $\mathcal {L}_{\rm AL}$, $\mathcal {L}_{\rm AL} + \mathcal {L}_{\rm CA} + \mathcal {L}_{\rm CO}$, $\mathcal {L}_{\rm ACP}$, and $\mathcal {L}_{\rm ACP} + \mathcal {L}_{\rm AL} + \mathcal {L}_{\rm CA} + \mathcal {L}_{\rm CO}$.
<<</Model Configurations>>>
<<<Results and Discussion>>>
Table TABREF23 shows accuracy. As the Random baseline suggests, positive and negative labels were distributed evenly. The Random+Seed baseline made use of the seed lexicon and output the corresponding label (or the reverse of it for negation) if the event's predicate is in the seed lexicon. We can see that the seed lexicon itself had practically no impact on prediction.
The models in the top block performed considerably better than the random baselines. The performance gaps with their (semi-)supervised counterparts, shown in the middle block, were less than 7%. This demonstrates the effectiveness of discourse relation-based label propagation.
Comparing the model variants, we obtained the highest score with the BiGRU encoder trained with the AL+CA+CO dataset. BERT was competitive but its performance went down if CA and CO were used in addition to AL. We conjecture that BERT was more sensitive to noises found more frequently in CA and CO.
Contrary to our expectations, supervised models (ACP) outperformed semi-supervised models (ACP+AL+CA+CO). This suggests that the training set of 0.6 million events is sufficiently large for training the models. For comparison, we trained the models with a subset (6,000 events) of the ACP dataset. As the results shown in Table TABREF24 demonstrate, our method is effective when labeled data are small.
The result of hyperparameter optimization for the BiGRU encoder was as follows:
As the CA and CO pairs were equal in size (Table TABREF16), $\lambda _{\rm CA}$ and $\lambda _{\rm CO}$ were comparable values. $\lambda _{\rm CA}$ was about one-third of $\lambda _{\rm CO}$, and this indicated that the CA pairs were noisier than the CO pairs. A major type of CA pairs that violates our assumption was in the form of “$\textit {problem}_{\text{negative}}$ causes $\textit {solution}_{\text{positive}}$”:
. (悪いところがある, よくなるように努力する)
(there is a bad point, [I] try to improve [it])
The polarities of the two events were reversed in spite of the Cause relation, and this lowered the value of $\lambda _{\rm CA}$.
Some examples of model outputs are shown in Table TABREF26. The first two examples suggest that our model successfully learned negation without explicit supervision. Similarly, the next two examples differ only in voice but the model correctly recognized that they had opposite polarities. The last two examples share the predicate “落とす" (drop) and only the objects are different. The second event “肩を落とす" (lit. drop one's shoulders) is an idiom that expresses a disappointed feeling. The examples demonstrate that our model correctly learned non-compositional expressions.
<<</Results and Discussion>>>
<<</Experiments>>>
<<<Conclusion>>>
In this paper, we proposed to use discourse relations to effectively propagate polarities of affective events from seeds. Experiments show that, even with a minimal amount of supervision, the proposed method performed well.
Although event pairs linked by discourse analysis are shown to be useful, they nevertheless contain noises. Adding linguistically-motivated filtering rules would help improve the performance.
<<</Conclusion>>>
<<</Title>>>
| {
"references": [
"100 million sentences"
],
"type": "extractive"
} |
1910.14497 | Please extract a concise answer without any additional explanation for the following question based on the given text.
Question: How is embedding quality assessed?
Context: <<<Title>>>
Probabilistic Bias Mitigation in Word Embeddings
<<<Abstract>>>
It has been shown that word embeddings derived from large corpora tend to incorporate biases present in their training data. Various methods for mitigating these biases have been proposed, but recent work has demonstrated that these methods hide but fail to truly remove the biases, which can still be observed in word nearest-neighbor statistics. In this work we propose a probabilistic view of word embedding bias. We leverage this framework to present a novel method for mitigating bias which relies on probabilistic observations to yield a more robust bias mitigation algorithm. We demonstrate that this method effectively reduces bias according to three separate measures of bias while maintaining embedding quality across various popular benchmark semantic tasks
<<</Abstract>>>
<<<Introduction>>>
Word embeddings, or vector representations of words, are an important component of Natural Language Processing (NLP) models and necessary for many downstream tasks. However, word embeddings, including embeddings commonly deployed for public use, have been shown to exhibit unwanted societal stereotypes and biases, raising concerns about disparate impact on axes of gender, race, ethnicity, and religion BIBREF0, BIBREF1. The impact of this bias has manifested in a range of downstream tasks, ranging from autocomplete suggestions BIBREF2 to advertisement delivery BIBREF3, increasing the likelihood of amplifying harmful biases through the use of these models.
The most well-established method thus far for mitigating bias relies on projecting target words onto a bias subspace (such as a gender subspace) and subtracting out the difference between the resulting distances BIBREF0. On the other hand, the most popular metric for measuring bias is the WEAT statistic BIBREF1, which compares the cosine similarities between groups of words. However, WEAT has been recently shown to overestimate bias as a result of implicitly relying on similar frequencies for the target words BIBREF4, and BIBREF5 demonstrated that evidence of bias can still be recovered after geometric bias mitigation by examining the neighborhood of a target word among socially-biased words.
In response to this, we propose an alternative framework for bias mitigation in word embeddings that approaches this problem from a probabilistic perspective. The motivation for this approach is two-fold. First, most popular word embedding algorithms are probabilistic at their core – i.e., they are trained (explicitly or implicitly BIBREF6) to minimize some form of word co-occurrence probabilities. Thus, we argue that a framework for measuring and treating bias in these embeddings should take into account, in addition to their geometric aspect, their probabilistic nature too. On the other hand, the issue of bias has also been approached (albeit in different contexts) in the fairness literature, where various intuitive notions of equity such as equalized odds have been formalized through probabilistic criteria. By considering analogous criteria for the word embedding setting, we seek to draw connections between these two bodies of work.
We present experiments on various bias mitigation benchmarks and show that our framework is comparable to state-of-the-art alternatives according to measures of geometric bias mitigation and that it performs far better according to measures of neighborhood bias. For fair comparison, we focus on mitigating a binary gender bias in pre-trained word embeddings using SGNS (skip-gram with negative-sampling), though we note that this framework and methods could be extended to other types of bias and word embedding algorithms.
<<</Introduction>>>
<<<Background>>>
<<<Geometric Bias Mitigation>>>
Geometric bias mitigation uses the cosine distances between words to both measure and remove gender bias BIBREF0. This method implicitly defines bias as a geometric asymmetry between words when projected onto a subspace, such as the gender subspace constructed from a set of gender pairs such as $\mathcal {P} = \lbrace (he,she),(man,woman),(king,queen)...\rbrace $. The projection of a vector $v$ onto $B$ (the subspace) is defined by $v_B = \sum _{j=1}^{k} (v \cdot b_j) b_j$ where a subspace $B$ is defined by k orthogonal unit vectors $B = {b_1,...,b_k}$.
<<<WEAT>>>
The WEAT statistic BIBREF1 demonstrates the presence of biases in word embeddings with an effect size defined as the mean test statistic across the two word sets:
Where $s$, the test statistic, is defined as: $s(w,A,B) = mean_{a \in A} cos(w,a) - mean_{b \in B} cos(w,a)$, and $X$,$Y$,$A$, and $B$ are groups of words for which the association is measured. Possible values range from $-2$ to 2 depending on the association of the words groups, and a value of zero indicates $X$ and $Y$ are equally associated with $A$ and $B$. See BIBREF4 for further details on WEAT.
<<</WEAT>>>
<<<RIPA>>>
The RIPA (relational inner product association) metric was developed as an alternative to WEAT, with the critique that WEAT is likely to overestimate the bias of a target attribute BIBREF4. The RIPA metric formalizes the measure of bias used in geometric bias mitigation as the inner product association of a word vector $v$ with respect to a relation vector $b$. The relation vector is constructed from the first principal component of the differences between gender word pairs. We report the absolute value of the RIPA metric as the value can be positive or negative according to the direction of the bias. A value of zero indicates a lack of bias, and the value is bound by $[-||w||,||w||]$.
<<</RIPA>>>
<<<Neighborhood Metric>>>
The neighborhood bias metric proposed by BIBREF5 quantifies bias as the proportion of male socially-biased words among the $k$ nearest socially-biased male and female neighboring words, whereby biased words are obtained by projecting neutral words onto a gender relation vector. As we only examine the target word among the 1000 most socially-biased words in the vocabulary (500 male and 500 female), a word’s bias is measured as the ratio of its neighborhood of socially-biased male and socially-biased female words, so that a value of 0.5 in this metric would indicate a perfectly unbiased word, and values closer to 0 and 1 indicate stronger bias.
<<</Neighborhood Metric>>>
<<</Geometric Bias Mitigation>>>
<<</Background>>>
<<<A Probabilistic Framework for Bias Mitigation>>>
Our objective here is to extend and complement the geometric notions of word embedding bias described in the previous section with an alternative, probabilistic, approach. Intuitively, we seek a notion of equality akin to that of demographic parity in the fairness literature, which requires that a decision or outcome be independent of a protected attribute such as gender. BIBREF7. Similarly, when considering a probabilistic definition of unbiased in word embeddings, we can consider the conditional probabilities of word pairs, ensuring for example that $p(doctor|man) \approx p(doctor|woman)$, and can extend this probabilistic framework to include the neighborhood of a target word, addressing the potential pitfalls of geometric bias mitigation.
Conveniently, most word embedding frameworks allow for immediate computation of the conditional probabilities $P(w|c)$. Here, we focus our attention on the Skip-Gram method with Negative Sampling (SGNS) of BIBREF8, although our framework can be equivalently instantiated for most other popular embedding methods, owing to their core similarities BIBREF6, BIBREF9. Leveraging this probabilistic nature, we construct a bias mitigation method in two steps, and examine each step as an independent method as well as the resulting composite method.
<<<Probabilistic Bias Mitigation>>>
This component of our bias mitigation framework seeks to enforce that the probability of prediction or outcome cannot depend on a protected class such as gender. We can formalize this intuitive goal through a loss function that penalizes the discrepancy between the conditional probabilities of a target word (i.e., one that should not be affected by the protected attribute) conditioned on two words describing the protected attribute (e.g., man and woman in the case of gender). That is, for every target word we seek to minimize:
where $\mathcal {P} = \lbrace (he,she),(man,woman),(king,queen), \dots \rbrace $ is a set of word pairs characterizing the protected attribute, akin to that used in previous work BIBREF0.
At this point, the specific form of the objective will depend on the type of word embeddings used. For our expample of SGNS, recall that this algorithm models the conditional probability of a target word given a context word as a function of the inner product of their representations. Though an exact method for calculating the conditional probability includes summing over conditional probability of all the words in the vocabulary, we can use the estimation of log conditional probability proposed by BIBREF8, i.e., $ \log p(w_O|w_I) \approx \log \sigma ({v^{\prime }_{wo}}^T v_{wI}) + \sum _{i=1}^{k} [\log {\sigma ({{-v^{\prime }_{wi}}^T v_{wI}})}] $.
<<</Probabilistic Bias Mitigation>>>
<<<Nearest Neighbor Bias Mitigation>>>
Based on observations by BIBREF5, we extend our method to consider the composition of the neighborhood of socially-gendered words of a target word. We note that bias in a word embedding depends not only on the relationship between a target word and explicitly gendered words like man and woman, but also between a target word and socially-biased male or female words. Bolukbasi et al BIBREF0 proposed a method for eliminating this kind of indirect bias through geometric bias mitigation, but it is shown to be ineffective by the neighborhood metric BIBREF5.
Instead, we extend our method of bias mitigation to account for this neighborhood effect. Specifically, we examine the conditional probabilities of a target word given the $k/2$ nearest neighbors from the male socially-biased words as well as given the $k/2$ female socially-biased words (in sorted order, from smallest to largest). The groups of socially-biased words are constructed as described in the neighborhood metric. If the word is unbiased according to the neighborhood metric, these probabilities should be comparable. We then use the following as our loss function:
where $m$ and $f$ represent the male and female neighbors sorted by distance to the target word $t$ (we use $L1$ distance).
<<</Nearest Neighbor Bias Mitigation>>>
<<</A Probabilistic Framework for Bias Mitigation>>>
<<<Experiments>>>
We evaluate our framework on fastText embeddings trained on Wikipedia (2017), UMBC webbase corpus and statmt.org news dataset (16B tokens) BIBREF11. For simplicity, only the first 22000 words are used in all embeddings, though preliminary results indicate the findings extend to the full corpus. For our novel methods of mitigating bias, a shallow neural network is used to adjust the embedding. The single layer of the model is an embedding layer with weights initialized to those of the original embedding. For the composite method, these weights are initialized to those of the embedding after probabilistic bias mitigation. A batch of word indices is fed into the model, which are then embedded and for which a loss value is calculated, allowing back-propagation to adjust the embeddings. For each of the models, a fixed number of iterations is used to prevent overfitting, which can eventually hurt performance on the embedding benchmarks (See Figure FIGREF12). We evaluated the embedding after 1000 iterations, and stopped training if performance on a benchmark decreased significantly.
We construct a list of candidate words to debias, taken from the words used in the WEAT gender bias statistics. Words in this list should be gender neutral, and are related to the topics of career, arts, science, math, family and professions (see appendix). We note that this list can easily be expanded to include a greater proportion of words in the corpus. For example, BIBREF4 suggested a method for identifying inappropriately gendered words using unsupervised learning.
We compare this method of bias mitigation with the no bias mitigation ("Orig"), geometric bias mitigation ("Geo"), the two pieces of our method alone ("Prob" and "KNN") and the composite method ("KNN+Prob"). We note that the composite method performs reasonably well according the the RIPA metric, and much better than traditional geometric bias mitigation according to the neighborhood metric, without significant performance loss according to the accepted benchmarks. To our knowledge this is the first bias mitigation method to perform reasonably both on both metrics.
<<</Experiments>>>
<<<Discussion>>>
We proposed a simple method of bias mitigation based on this probabilistic notions of fairness, and showed that it leads to promising results in various benchmark bias mitigation tasks. Future work should include considering a more rigorous definition and non-binary of bias and experimenting with various embedding algorithms and network architectures.
<<<Acknowledgements>>>
The authors would like to thank Tommi Jaakkola for stimulating discussions during the initial stages of this work.
<<</Acknowledgements>>>
<<</Discussion>>>
<<<Experiment Notes>>>
For Equation 4, as described in the original work, in regards to the k sample words $w_i$ is drawn from the corpus using the Unigram distribution raised to the 3/4 power.
For reference, the most male socially-biased words include words such as:’john’, ’jr’, ’mlb’, ’dick’, ’nfl’, ’cfl’, ’sgt’, ’abbot’, ’halfback’, ’jock’, ’mike’, ’joseph’,while the most female socially-biased words include words such as:’feminine’, ’marital’, ’tatiana’, ’pregnancy’, ’eva’, ’pageant’, ’distress’, ’cristina’, ’ida’, ’beauty’, ’sexuality’,’fertility’
<<</Experiment Notes>>>
<<<Professions>>>
'accountant', 'acquaintance', 'actor', 'actress', 'administrator', 'adventurer', 'advocate', 'aide', 'alderman', 'ambassador', 'analyst', 'anthropologist', 'archaeologist', 'archbishop', 'architect', 'artist', 'assassin', 'astronaut', 'astronomer', 'athlete', 'attorney', 'author', 'baker', 'banker', 'barber', 'baron', 'barrister', 'bartender', 'biologist', 'bishop', 'bodyguard', 'boss', 'boxer', 'broadcaster', 'broker', 'businessman', 'butcher', 'butler', 'captain', 'caretaker', 'carpenter', 'cartoonist', 'cellist', 'chancellor', 'chaplain', 'character', 'chef', 'chemist', 'choreographer', 'cinematographer', 'citizen', 'cleric', 'clerk', 'coach', 'collector', 'colonel', 'columnist', 'comedian', 'comic', 'commander', 'commentator', 'commissioner', 'composer', 'conductor', 'confesses', 'congressman', 'constable', 'consultant', 'cop', 'correspondent', 'counselor', 'critic', 'crusader', 'curator', 'dad', 'dancer', 'dean', 'dentist', 'deputy', 'detective', 'diplomat', 'director', 'doctor', 'drummer', 'economist', 'editor', 'educator', 'employee', 'entertainer', 'entrepreneur', 'envoy', 'evangelist', 'farmer', 'filmmaker', 'financier', 'fisherman', 'footballer', 'foreman', 'gangster', 'gardener', 'geologist', 'goalkeeper', 'guitarist', 'headmaster', 'historian', 'hooker', 'illustrator', 'industrialist', 'inspector', 'instructor', 'inventor', 'investigator', 'journalist', 'judge', 'jurist', 'landlord', 'lawyer', 'lecturer', 'legislator', 'librarian', 'lieutenant', 'lyricist', 'maestro', 'magician', 'magistrate', 'maid', 'manager', 'marshal', 'mathematician', 'mechanic', 'midfielder', 'minister', 'missionary', 'monk', 'musician', 'nanny', 'narrator', 'naturalist', 'novelist', 'nun', 'nurse', 'observer', 'officer', 'organist', 'painter', 'pastor', 'performer', 'philanthropist', 'philosopher', 'photographer', 'physician', 'physicist', 'pianist', 'planner', 'playwright', 'poet', 'policeman', 'politician', 'preacher', 'president', 'priest', 'principal', 'prisoner', 'professor', 'programmer', 'promoter', 'proprietor', 'prosecutor', 'protagonist', 'provost', 'psychiatrist', 'psychologist', 'rabbi', 'ranger', 'researcher', 'sailor', 'saint', 'salesman', 'saxophonist', 'scholar', 'scientist', 'screenwriter', 'sculptor', 'secretary', 'senator', 'sergeant', 'servant', 'singer', 'skipper', 'sociologist', 'soldier', 'solicitor', 'soloist', 'sportsman', 'statesman', 'steward', 'student', 'substitute', 'superintendent', 'surgeon', 'surveyor', 'swimmer', 'teacher', 'technician', 'teenager', 'therapist', 'trader', 'treasurer', 'trooper', 'trumpeter', 'tutor', 'tycoon', 'violinist', 'vocalist', 'waiter', 'waitress', 'warden', 'warrior', 'worker', 'wrestler', 'writer'
<<</Professions>>>
<<<WEAT Word Sets>>>
Words used for WEAT statistic, consisting of baseline bias tests and gender bias tests in the format X vs Y / A vs B
Flowers vs Insects / Pleasant vs Unpleasant
X: "aster", "clover", "hyacinth", "marigold", "poppy", "azalea", "crocus", "iris", "orchid", "rose", "bluebell", "daffodil", "lilac", "pansy", "tulip", "buttercup", "daisy", "lily", "peony", "violet", "carnation", "gladiola", "magnolia", "petunia", "zinnia"
Y: "ant", "caterpillar", "flea", "locust", "spider", "bedbug", "centipede", "fly", "maggot", "tarantula", "bee", "cockroach", "gnat", "mosquito", "termite", "beetle", "cricket", "hornet", "moth", "wasp", "blackfly", "dragonfly", "horsefly", "roach", "weevil"
A: "caress", "freedom", "health", "love", "peace", "cheer", "friend", "heaven", "loyal", "pleasure", "diamond", "gentle", "honest", "lucky", "rainbow", "diploma", "gift", "honor", "miracle", "sunrise", "family", "happy", "laughter", "paradise", "vacation"
B: "abuse", "crash", "filth", "murder", "sickness", "accident", "death", "grief", "poison", "stink", "assault", "disaster", "hatred", "pollute", "tragedy", "divorce", "jail", "poverty", "ugly", "cancer", "kill", "rotten", "vomit", "agony", "prison"
Instruments vs Weapons / Pleasant vs Unpleasant:
X: "bagpipe", "cello", "guitar", "lute", "trombone", "banjo", "clarinet", "harmonica", "mandolin", "trumpet", "bassoon", "drum", "harp", "oboe", "tuba", "bell", "fiddle", "harpsichord", "piano", "viola", "bongo", "flute", "horn", "saxophone", "violin"
Y: "arrow", "club", "gun", "missile", "spear", "ax", "dagger", "harpoon", "pistol", "sword", "blade", "dynamite", "hatchet", "rifle", "tank", "bomb", "firearm", "knife", "shotgun", "teargas", "cannon", "grenade", "mace", "slingshot", "whip"
A: "caress", "freedom", "health", "love", "peace", "cheer", "friend", "heaven", "loyal", "pleasure", "diamond", "gentle", "honest", "lucky", "rainbow", "diploma", "gift", "honor", "miracle", "sunrise", "family", "happy", "laughter", "paradise", "vacation"
B: "abuse", "crash", "filth", "murder", "sickness", "accident", "death", "grief", "poison", "stink", "assault", "disaster", "hatred", "pollute", "tragedy", "divorce", "jail", "poverty", "ugly", "cancer", "kill", "rotten", "vomit", "agony", "prison"
Male vs Female / Career vs Family:
X: "brother", "father", "uncle", "grandfather", "son", "he", "his", "him", "man", "himself", "men", "husband", "boy", "uncle", "nephew", "boyfriend", "king", "actor"
Y: "sister", "mother", "aunt", "grandmother", "daughter", "she", "hers", "her", "woman", "herself", "women", "wife", "aunt", "niece", "girlfriend", "queen", "actress"
A: "executive", "management", "professional", "corporation", "salary", "office", "business", "career", "industry", "company", "promotion", "profession", "CEO", "manager", "coworker", "entrepreneur"
B: "home", "parents", "children", "family", "cousins", "marriage", "wedding", "relatives", "grandparents", "grandchildren", "nurture", "child", "toddler", "infant", "teenager"
Math vs Art / Male vs Female:
X: "math", "algebra", "geometry", "calculus", "equations", "computation", "numbers", "addition", "trigonometry", "arithmetic", "logic", "proofs", "multiplication", "mathematics"
Y: "poetry", "art", "Shakespeare", "dance", "literature", "novel", "symphony", "drama", "orchestra", "music", "ballet", "arts", "creative", "sculpture"
A: "brother", "father", "uncle", "grandfather", "son", "he", "his", "him", "man", "himself", "men", "husband", "boy", "uncle", "nephew", "boyfriend", "king", "actor"
B: "sister", "mother", "aunt", "grandmother", "daughter", "she", "hers", "her", "woman", "herself", "women", "wife", "aunt", "niece", "girlfriend", "queen", "actress"
Science vs Art / Male8 vs Female8:
X:"science", "technology", "physics", "chemistry", "Einstein", "NASA", "experiment", "astronomy", "biology", "aeronautics", "mechanics", "thermodynamics"
Y: "poetry", "art", "Shakespeare", "dance", "literature", "novel", "symphony", "drama", "orchestra", "music", "ballet", "arts", "creative", "sculpture"
A: "brother", "father", "uncle", "grandfather", "son", "he", "his", "him", "man", "himself", "men", "husband", "boy", "uncle", "nephew", "boyfriend"
B: "sister", "mother", "aunt", "grandmother", "daughter", "she", "hers", "her", "woman", "herself", "women", "wife", "aunt", "niece", "girlfriend"
<<</WEAT Word Sets>>>
<<</Title>>>
| {
"references": [
"We compare this method of bias mitigation with the no bias mitigation (\"Orig\"), geometric bias mitigation (\"Geo\"), the two pieces of our method alone (\"Prob\" and \"KNN\") and the composite method (\"KNN+Prob\"). We note that the composite method performs reasonably well according the the RIPA metric, and much better than traditional geometric bias mitigation according to the neighborhood metric, without significant performance loss according to the accepted benchmarks. To our knowledge this is the first bias mitigation method to perform reasonably both on both metrics."
],
"type": "extractive"
} |
1912.02481 | Please extract a concise answer without any additional explanation for the following question based on the given text.
Question: What turn out to be more important high volume or high quality data?
Context: <<<Title>>>
Massive vs. Curated Word Embeddings for Low-Resourced Languages. The Case of Yor\`ub\'a and Twi
<<<Abstract>>>
The success of several architectures to learn semantic representations from unannotated text and the availability of these kind of texts in online multilingual resources such as Wikipedia has facilitated the massive and automatic creation of resources for multiple languages. The evaluation of such resources is usually done for the high-resourced languages, where one has a smorgasbord of tasks and test sets to evaluate on. For low-resourced languages, the evaluation is more difficult and normally ignored, with the hope that the impressive capability of deep learning architectures to learn (multilingual) representations in the high-resourced setting holds in the low-resourced setting too. In this paper we focus on two African languages, Yor\`ub\'a and Twi, and compare the word embeddings obtained in this way, with word embeddings obtained from curated corpora and a language-dependent processing. We analyse the noise in the publicly available corpora, collect high quality and noisy data for the two languages and quantify the improvements that depend not only on the amount of data but on the quality too. We also use different architectures that learn word representations both from surface forms and characters to further exploit all the available information which showed to be important for these languages. For the evaluation, we manually translate the wordsim-353 word pairs dataset from English into Yor\`ub\'a and Twi. As output of the work, we provide corpora, embeddings and the test suits for both languages.
<<</Abstract>>>
<<<Introduction>>>
In recent years, word embeddings BIBREF0, BIBREF1, BIBREF2 have been proven to be very useful for training downstream natural language processing (NLP) tasks. Moreover, contextualized embeddings BIBREF3, BIBREF4 have been shown to further improve the performance of NLP tasks such as named entity recognition, question answering, or text classification when used as word features because they are able to resolve ambiguities of word representations when they appear in different contexts. Different deep learning architectures such as multilingual BERT BIBREF4, LASER BIBREF5 and XLM BIBREF6 have proved successful in the multilingual setting. All these architectures learn the semantic representations from unannotated text, making them cheap given the availability of texts in online multilingual resources such as Wikipedia. However, the evaluation of such resources is usually done for the high-resourced languages, where one has a smorgasbord of tasks and test sets to evaluate on. This is the best-case scenario, languages with tones of data for training that generate high-quality models.
For low-resourced languages, the evaluation is more difficult and therefore normally ignored simply because of the lack of resources. In these cases, training data is scarce, and the assumption that the capability of deep learning architectures to learn (multilingual) representations in the high-resourced setting holds in the low-resourced one does not need to be true. In this work, we focus on two African languages, Yorùbá and Twi, and carry out several experiments to verify this claim. Just by a simple inspection of the word embeddings trained on Wikipedia by fastText, we see a high number of non-Yorùbá or non-Twi words in the vocabularies. For Twi, the vocabulary has only 935 words, and for Yorùbá we estimate that 135 k out of the 150 k words belong to other languages such as English, French and Arabic.
In order to improve the semantic representations for these languages, we collect online texts and study the influence of the quality and quantity of the data in the final models. We also examine the most appropriate architecture depending on the characteristics of each language. Finally, we translate test sets and annotate corpora to evaluate the performance of both our models together with fastText and BERT pre-trained embeddings which could not be evaluated otherwise for Yorùbá and Twi. The evaluation is carried out in a word similarity and relatedness task using the wordsim-353 test set, and in a named entity recognition (NER) task where embeddings play a crucial role. Of course, the evaluation of the models in only two tasks is not exhaustive but it is an indication of the quality we can obtain for these two low-resourced languages as compared to others such as English where these evaluations are already available.
The rest of the paper is organized as follows. Related works are reviewed in Section SECREF2 The two languages under study are described in Section SECREF3. We introduce the corpora and test sets in Section SECREF4. The fifth section explores the different training architectures we consider, and the experiments that are carried out. Finally, discussion and concluding remarks are given in Section SECREF6
<<</Introduction>>>
<<<Related Work>>>
The large amount of freely available text in the internet for multiple languages is facilitating the massive and automatic creation of multilingual resources. The resource par excellence is Wikipedia, an online encyclopedia currently available in 307 languages. Other initiatives such as Common Crawl or the Jehovah’s Witnesses site are also repositories for multilingual data, usually assumed to be noisier than Wikipedia. Word and contextual embeddings have been pre-trained on these data, so that the resources are nowadays at hand for more than 100 languages. Some examples include fastText word embeddings BIBREF2, BIBREF7, MUSE embeddings BIBREF8, BERT multilingual embeddings BIBREF4 and LASER sentence embeddings BIBREF5. In all cases, embeddings are trained either simultaneously for multiple languages, joining high- and low-resource data, or following the same methodology.
On the other hand, different approaches try to specifically design architectures to learn embeddings in a low-resourced setting. ChaudharyEtAl:2018 follow a transfer learning approach that uses phonemes, lemmas and morphological tags to transfer the knowledge from related high-resource language into the low-resource one. jiangEtal:2018 apply Positive-Unlabeled Learning for word embedding calculations, assuming that unobserved pairs of words in a corpus also convey information, and this is specially important for small corpora.
In order to assess the quality of word embeddings, word similarity and relatedness tasks are usually used. wordsim-353 BIBREF9 is a collection of 353 pairs annotated with semantic similarity scores in a scale from 0 to 10. Even the problems detected in this dataset BIBREF10, it is widely used by the community. The test set was originally created for English, but the need for comparison with other languages has motivated several translations/adaptations. In hassanMihalcea:2009 the test was translated manually into Spanish, Romanian and Arabic and the scores were adapted to reflect similarities in the new language. The reported correlation between the English scores and the Spanish ones is 0.86. Later, JoubarneInkpen:2011 show indications that the measures of similarity highly correlate across languages. leviantReichart:2015 translated also wordsim-353 into German, Italian and Russian and used crowdsourcing to score the pairs. Finally, jiangEtal:2018 translated with Google Cloud the test set from English into Czech, Danish and Dutch. In our work, native speakers translate wordsim-353 into Yorùbá and Twi, and similarity scores are kept unless the discrepancy with English is big (see Section SECREF11 for details). A similar approach to our work is done for Gujarati in JoshiEtAl:2019.
<<</Related Work>>>
<<<Languages under Study>>>
<<<Yorùbá>>>
is a language in the West Africa with over 50 million speakers. It is spoken among other languages in Nigeria, republic of Togo, Benin Republic, Ghana and Sierra Leon. It is also a language of Òrìsà in Cuba, Brazil, and some Caribbean countries. It is one of the three major languages in Nigeria and it is regarded as the third most spoken native African language. There are different dialects of Yorùbá in Nigeria BIBREF11, BIBREF12, BIBREF13. However, in this paper our focus is the standard Yorùbá based upon a report from the 1974 Joint Consultative Committee on Education BIBREF14.
Standard Yorùbá has 25 letters without the Latin characters c, q, v, x and z. There are 18 consonants (b, d, f, g, gb, j[dz], k, l, m, n, p[kp], r, s, ṣ, t, w y[j]), 7 oral vowels (a, e, ẹ, i, o, ọ, u), five nasal vowels, (an, $ \underaccent{\dot{}}{e}$n, in, $ \underaccent{\dot{}}{o}$n, un) and syllabic nasals (m̀, ḿ, ǹ, ń). Yorùbá is a tone language which makes heavy use of lexical tones which are indicated by the use of diacritics. There are three tones in Yorùbá namely low, mid and high which are represented as grave ($\setminus $), macron ($-$) and acute ($/$) symbols respectively. These tones are applied on vowels and syllabic nasals. Mid tone is usually left unmarked on vowels and every initial or first vowel in a word cannot have a high tone. It is important to note that tone information is needed for correct pronunciation and to have the meaning of a word BIBREF15, BIBREF12, BIBREF14. For example, owó (money), ọw (broom), òwò (business), w (honour), ọw (hand), and w (group) are different words with different dots and diacritic combinations. According to Asahiah2014, Standard Yorùbá uses 4 diacritics, 3 are for marking tones while the fourth which is the dot below is used to indicate the open phonetic variants of letter "e" and "o" and the long variant of "s". Also, there are 19 single diacritic letters, 3 are marked with dots below (ẹ, ọ, ṣ) while the rest are either having the grave or acute accent. The four double diacritics are divided between the grave and the acute accent as well.
As noted in Asahiah2014, most of the Yorùbá texts found in websites or public domain repositories (i) either use the correct Yorùbá orthography or (ii) replace diacritized characters with un-diacritized ones.
This happens as a result of many factors, but most especially to the unavailability of appropriate input devices for the accurate application of the diacritical marks BIBREF11. This has led to research on restoration models for diacritics BIBREF16, but the problem is not well solved and we find that most Yorùbá text in the public domain today is not well diacritized. Wikipedia is not an exception.
<<</Yorùbá>>>
<<<Twi>>>
is an Akan language of the Central Tano Branch of the Niger Congo family of languages. It is the most widely spoken of the about 80 indigenous languages in Ghana BIBREF17. It has about 9 million native speakers and about a total of 17–18 million Ghanaians have it as either first or second language. There are two mutually intelligible dialects, Asante and Akuapem, and sub-dialectical variants which are mostly unknown to and unnoticed by non-native speakers. It is also mutually intelligible with Fante and to a large extent Bono, another of the Akan languages. It is one of, if not the, easiest to learn to speak of the indigenous Ghanaian languages. The same is however not true when it comes to reading and especially writing. This is due to a number of easily overlooked complexities in the structure of the language. First of all, similarly to Yorùbá, Twi is a tonal language but written without diacritics or accents. As a result, words which are pronounced differently and unambiguous in speech tend to be ambiguous in writing. Besides, most of such words fit interchangeably in the same context and some of them can have more than two meanings. A simple example is:
Me papa aba nti na me ne wo redi no yie no. S wo ara wo nim s me papa ba a, me suban fofor adi.
This sentence could be translated as
(i) I'm only treating you nicely because I'm in a good mood. You already know I'm a completely different person when I'm in a good mood.
(ii) I'm only treating you nicely because my dad is around. You already know I'm a completely different person when my dad comes around.
Another characteristic of Twi is the fact that a good number of stop words have the same written form as content words. For instance, “na” or “na” could be the words “and, then”, the phrase “and then” or the word “mother”. This kind of ambiguity has consequences in several natural language applications where stop words are removed from text.
Finally, we want to point out that words can also be written with or without prefixes. An example is this same na and na which happen to be the same word with an omissible prefix across its multiple senses. For some words, the prefix characters are mostly used when the word begins a sentence and omitted in the middle. This however depends on the author/speaker. For the word embeddings calculation, this implies that one would have different embeddings for the same word found in different contexts.
<<</Twi>>>
<<</Languages under Study>>>
<<<Data>>>
We collect clean and noisy corpora for Yorùbá and Twi in order to quantify the effect of noise on the quality of the embeddings, where noisy has a different meaning depending on the language as it will be explained in the next subsections.
<<<Training Corpora>>>
For Yorùbá, we use several corpora collected by the Niger-Volta Language Technologies Institute with texts from different sources, including the Lagos-NWU conversational speech corpus, fully-diacritized Yorùbá language websites and an online Bible. The largest source with clean data is the JW300 corpus. We also created our own small-sized corpus by web-crawling three Yorùbá language websites (Alàkwé, r Yorùbá and Èdè Yorùbá Rẹw in Table TABREF7), some Yoruba Tweets with full diacritics and also news corpora (BBC Yorùbá and VON Yorùbá) with poor diacritics which we use to introduce noise. By noisy corpus, we refer to texts with incorrect diacritics (e.g in BBC Yorùbá), removal of tonal symbols (e.g in VON Yorùbá) and removal of all diacritics/under-dots (e.g some articles in Yorùbá Wikipedia). Furthermore, we got two manually typed fully-diacritized Yorùbá literature (Ìrìnkèrindò nínú igbó elégbèje and Igbó Olódùmarè) both written by Daniel Orowole Olorunfemi Fagunwa a popular Yorùbá author. The number of tokens available from each source, the link to the original source and the quality of the data is summarised in Table TABREF7.
The gathering of clean data in Twi is more difficult. We use as the base text as it has been shown that the Bible is the most available resource for low and endangered languages BIBREF18. This is the cleanest of all the text we could obtain. In addition, we use the available (and small) Wikipedia dumps which are quite noisy, i.e. Wikipedia contains a good number of English words, spelling errors and Twi sentences formulated in a non-natural way (formulated as L2 speakers would speak Twi as compared to native speakers). Lastly, we added text crawled from jw and the JW300 Twi corpus. Notice that the Bible text, is mainly written in the Asante dialect whilst the last, Jehovah's Witnesses, was written mainly in the Akuapem dialect. The Wikipedia text is a mixture of the two dialects. This introduces a lot of noise into the embeddings as the spelling of most words differs especially at the end of the words due to the mixture of dialects. The JW300 Twi corpus also contains mixed dialects but is mainly Akuampem. In this case, the noise comes also from spelling errors and the uncommon addition of diacritics which are not standardised on certain vowels. Figures for Twi corpora are summarised in the bottom block of Table TABREF7.
<<</Training Corpora>>>
<<<Evaluation Test Sets>>>
<<<Yorùbá.>>>
One of the contribution of this work is the introduction of the wordsim-353 word pairs dataset for Yorùbá. All the 353 word pairs were translated from English to Yorùbá by 3 native speakers. The set is composed of 446 unique English words, 348 of which can be expressed as one-word translation in Yorùbá (e.g. book translates to ìwé). In 61 cases (most countries and locations but also other content words) translations are transliterations (e.g. Doctor is dókítà and cucumber kùkúmbà.). 98 words were translated by short phrases instead of single words. This mostly affects words from science and technology (e.g. keyboard translates to pátákó ìtwé —literally meaning typing board—, laboratory translates to ìyàrá ìṣèwádìí —research room—, and ecology translates to ìm nípa àyíká while psychology translates to ìm nípa dá). Finally, 6 terms have the same form in English and Yorùbá therefore they are retained like that in the dataset (e.g. Jazz, Rock and acronyms such as FBI or OPEC).
We also annotate the Global Voices Yorùbá corpus to test the performance of our trained Yorùbá BERT embeddings on the named entity recognition task. The corpus consists of 25 k tokens which we annotate with four named entity types: DATE, location (LOC), organization (ORG) and personal names (PER). Any other token that does not belong to the four named entities is tagged with "O". The dataset is further split into training (70%), development (10%) and test (20%) partitions. Table TABREF12 shows the number of named entities per type and partition.
<<</Yorùbá.>>>
<<</Evaluation Test Sets>>>
<<</Data>>>
<<<Semantic Representations>>>
In this section, we describe the architectures used for learning word embeddings for the Twi and Yorùbá languages. Also, we discuss the quality of the embeddings as measured by the correlation with human judgements on the translated wordSim-353 test sets and by the F1 score in a NER task.
<<<Word Embeddings Architectures>>>
Modeling sub-word units has recently become a popular way to address out-of-vocabulary word problem in NLP especially in word representation learning BIBREF19, BIBREF2, BIBREF4. A sub-word unit can be a character, character $n$-grams, or heuristically learned Byte Pair Encodings (BPE) which work very well in practice especially for morphologically rich languages. Here, we consider two word embedding models that make use of character-level information together with word information: Character Word Embedding (CWE) BIBREF20 and fastText BIBREF2. Both of them are extensions of the Word2Vec architectures BIBREF0 that model sub-word units, character embeddings in the case of CWE and character $n$-grams for fastText.
CWE was introduced in 2015 to model the embeddings of characters jointly with words in order to address the issues of character ambiguities and non-compositional words especially in the Chinese language. A word or character embedding is learned in CWE using either CBOW or skipgram architectures, and then the final word embedding is computed by adding the character embeddings to the word itself:
where $w_j$ is the word embedding of $x_j$, $N_j$ is the number of characters in $x_j$, and $c_k$ is the embedding of the $k$-th character $c_k$ in $x_j$.
Similarly, in 2017 fastText was introduced as an extension to skipgram in order to take into account morphology and improve the representation of rare words. In this case the embedding of a word also includes the embeddings of its character $n$-grams:
where $w_j$ is the word embedding of $x_j$, $G_j$ is the number of character $n$-grams in $x_j$ and $g_k$ is the embedding of the $k$-th $n$-gram.
cwe also proposed three alternatives to learn multiple embeddings per character and resolve ambiguities: (i) position-based character embeddings where each character has different embeddings depending on the position it appears in a word, i.e., beginning, middle or end (ii) cluster-based character embeddings where a character can have $K$ different cluster embeddings, and (iii) position-based cluster embeddings (CWE-LP) where for each position $K$ different embeddings are learned. We use the latter in our experiments with CWE but no positional embeddings are used with fastText.
Finally, we consider a contextualized embedding architecture, BERT BIBREF4. BERT is a masked language model based on the highly efficient and parallelizable Transformer architecture BIBREF21 known to produce very rich contextualized representations for downstream NLP tasks.
The architecture is trained by jointly conditioning on both left and right contexts in all the transformer layers using two unsupervised objectives: Masked LM and Next-sentence prediction. The representation of a word is therefore learned according to the context it is found in.
Training contextual embeddings needs of huge amounts of corpora which are not available for low-resourced languages such as Yorùbá and Twi. However, Google provided pre-trained multilingual embeddings for 102 languages including Yorùbá (but not Twi).
<<</Word Embeddings Architectures>>>
<<<Experiments>>>
<<<FastText Training and Evaluation>>>
As a first experiment, we compare the quality of fastText embeddings trained on (high-quality) curated data and (low-quality) massively extracted data for Twi and Yorùbá languages.
Facebook released pre-trained word embeddings using fastText for 294 languages trained on Wikipedia BIBREF2 (F1 in tables) and for 157 languages trained on Wikipedia and Common Crawl BIBREF7 (F2). For Yorùbá, both versions are available but only embeddings trained on Wikipedia are available for Twi. We consider these embeddings the result of training on what we call massively-extracted corpora. Notice that training settings for both embeddings are not exactly the same, and differences in performance might come both from corpus size/quality but also from the background model. The 294-languages version is trained using skipgram, in dimension 300, with character $n$-grams of length 5, a window of size 5 and 5 negatives. The 157-languages version is trained using CBOW with position-weights, in dimension 300, with character $n$-grams of length 5, a window of size 5 and 10 negatives.
We want to compare the performance of these embeddings with the equivalent models that can be obtained by training on the different sources verified by native speakers of Twi and Yorùbá; what we call curated corpora and has been described in Section SECREF4 For the comparison, we define 3 datasets according to the quality and quantity of textual data used for training: (i) Curated Small Dataset (clean), C1, about 1.6 million tokens for Yorùbá and over 735 k tokens for Twi. The clean text for Twi is the Bible and for Yoruba all texts marked under the C1 column in Table TABREF7. (ii) In Curated Small Dataset (clean + noisy), C2, we add noise to the clean corpus (Wikipedia articles for Twi, and BBC Yorùbá news articles for Yorùbá). This increases the number of training tokens for Twi to 742 k tokens and Yorùbá to about 2 million tokens. (iii) Curated Large Dataset, C3 consists of all available texts we are able to crawl and source out for, either clean or noisy. The addition of JW300 BIBREF22 texts increases the vocabulary to more than 10 k tokens in both languages.
We train our fastText systems using a skipgram model with an embedding size of 300 dimensions, context window size of 5, 10 negatives and $n$-grams ranging from 3 to 6 characters similarly to the pre-trained models for both languages. Best results are obtained with minimum word count of 3.
Table TABREF15 shows the Spearman correlation between human judgements and cosine similarity scores on the wordSim-353 test set. Notice that pre-trained embeddings on Wikipedia show a very low correlation with humans on the similarity task for both languages ($\rho $=$0.14$) and their performance is even lower when Common Crawl is also considered ($\rho $=$0.07$ for Yorùbá). An important reason for the low performance is the limited vocabulary. The pre-trained Twi model has only 935 tokens. For Yorùbá, things are apparently better with more than 150 k tokens when both Wikipedia and Common Crawl are used but correlation is even lower. An inspection of the pre-trained embeddings indicates that over 135 k words belong to other languages mostly English, French and Arabic.
If we focus only on Wikipedia, we see that many texts are without diacritics in Yorùbá and often make use of mixed dialects and English sentences in Twi.
The Spearman $\rho $ correlation for fastText models on the curated small dataset (clean), C1, improves the baselines by a large margin ($\rho =0.354$ for Twi and 0.322 for Yorùbá) even with a small dataset. The improvement could be justified just by the larger vocabulary in Twi, but in the case of Yorùbá the enhancement is there with almost half of the vocabulary size. We found out that adding some noisy texts (C2 dataset) slightly improves the correlation for Twi language but not for the Yorùbá language. The Twi language benefits from Wikipedia articles because its inclusion doubles the vocabulary and reduces the bias of the model towards religious texts. However, for Yorùbá, noisy texts often ignore diacritics or tonal marks which increases the vocabulary size at the cost of an increment in the ambiguity too. As a result, the correlation is slightly hurt. One would expect that training with more data would improve the quality of the embeddings, but we found out with the results obtained with the C3 dataset, that only high-quality data helps. The addition of JW300 boosts the vocabulary in both cases, but whereas for Twi the corpus mixes dialects and is noisy, for Yorùbá it is very clean and with full diacritics. Consequently, the best embeddings for Yorùbá are obtained when training with the C3 dataset, whereas for Twi, C2 is the best option. In both cases, the curated embeddings improve the correlation with human judgements on the similarity task a $\Delta \rho =+0.25$ or, equivalently, by an increment on $\rho $ of 170% (Twi) and 180% (Yorùbá).
<<</FastText Training and Evaluation>>>
<<<CWE Training and Evaluation>>>
The huge ambiguity in the written Twi language motivates the exploration of different approaches to word embedding estimations. In this work, we compare the standard fastText methodology to include sub-word information with the character-enhanced approach with position-based clustered embeddings (CWE-LP as introduced in Section SECREF17). With the latter, we expect to specifically address the ambiguity present in a language that does not translate the different oral tones on vowels into the written language.
The character-enhanced word embeddings are trained using a skipgram architecture with cluster-based embeddings and an embedding size of 300 dimensions, context window-size of 5, and 5 negative samples. In this case, the best performance is obtained with a minimum word count of 1, and that increases the effective vocabulary that is used for training the embeddings with respect to the fastText experiments reported in Table TABREF15.
We repeat the same experiments as with fastText and summarise them in Table TABREF16. If we compare the relative numbers for the three datasets (C1, C2 and C3) we observe the same trends as before: the performance of the embeddings in the similarity task improves with the vocabulary size when the training data can be considered clean, but the performance diminishes when the data is noisy.
According to the results, CWE is specially beneficial for Twi but not always for Yorùbá. Clean Yorùbá text, does not have the ambiguity issues at character-level, therefore the $n$-gram approximation works better when enough clean data is used ($\rho ^{C3}_{CWE}=0.354$ vs. $\rho ^{C3}_{fastText}=0.391$) but it does not when too much noisy data (no diacritics, therefore character-level information would be needed) is used ($\rho ^{C2}_{CWE}=0.345$ vs. $\rho ^{C2}_{fastText}=0.302$). For Twi, the character-level information reinforces the benefits of clean data and the best correlation with human judgements is reached with CWE embeddings ($\rho ^{C2}_{CWE}=0.437$ vs. $\rho ^{C2}_{fastText}=0.388$).
<<</CWE Training and Evaluation>>>
<<<BERT Evaluation on NER Task>>>
In order to go beyond the similarity task using static word vectors, we also investigate the quality of the multilingual BERT embeddings by fine-tuning a named entity recognition task on the Yorùbá Global Voices corpus.
One of the major advantages of pre-trained BERT embeddings is that fine-tuning of the model on downstream NLP tasks is typically computationally inexpensive, often with few number of epochs. However, the data the embeddings are trained on has the same limitations as that used in massive word embeddings. Fine-tuning involves replacing the last layer of BERT used optimizing the masked LM with a task-dependent linear classifier or any other deep learning architecture, and training all the model parameters end-to-end. For the NER task, we obtain the token-level representation from BERT and train a linear classifier for sequence tagging.
Similar to our observations with non-contextualized embeddings, we find out that fine-tuning the pre-trained multilingual-uncased BERT for 4 epochs on the NER task gives an F1 score of 0. If we do the same experiment in English, F1 is 58.1 after 4 epochs.
That shows how pre-trained embeddings by themselves do not perform well in downstream tasks on low-resource languages. To address this problem for Yorùbá, we fine-tune BERT representations on the Yorùbá corpus in two ways: (i) using the multilingual vocabulary, and (ii) using only Yorùbá vocabulary. In both cases diacritics are ignored to be consistent with the base model training.
As expected, the fine-tuning of the pre-trained BERT on the Yorùbá corpus in the two configurations generates better representations than the base model. These models are able to achieve a better performance on the NER task with an average F1 score of over 47% (see Table TABREF26 for the comparative). The fine-tuned BERT model with only Yorùbá vocabulary further increases by more than 4% in F1 score obtained with the tuning that uses the multilingual vocabulary. Although we do not have enough data to train BERT from scratch, we observe that fine-tuning BERT on a limited amount of monolingual data of a low-resource language helps to improve the quality of the embeddings. The same observation holds true for high-resource languages like German and French BIBREF23.
<<</BERT Evaluation on NER Task>>>
<<</Experiments>>>
<<</Semantic Representations>>>
<<<Summary and Discussion>>>
In this paper, we present curated word and contextual embeddings for Yorùbá and Twi. For this purpose, we gather and select corpora and study the most appropriate techniques for the languages. We also create test sets for the evaluation of the word embeddings within a word similarity task (wordsim353) and the contextual embeddings within a NER task. Corpora, embeddings and test sets are available in github.
In our analysis, we show how massively generated embeddings perform poorly for low-resourced languages as compared to the performance for high-resourced ones. This is due both to the quantity but also the quality of the data used. While the Pearson $\rho $ correlation for English obtained with fastText embeddings trained on Wikipedia (WP) and Common Crawl (CC) are $\rho _{WP}$=$0.67$ and $\rho _{WP+CC}$=$0.78$, the equivalent ones for Yorùbá are $\rho _{WP}$=$0.14$ and $\rho _{WP+CC}$=$0.07$. For Twi, only embeddings with Wikipedia are available ($\rho _{WP}$=$0.14$). By carefully gathering high-quality data and optimising the models to the characteristics of each language, we deliver embeddings with correlations of $\rho $=$0.39$ (Yorùbá) and $\rho $=$0.44$ (Twi) on the same test set, still far from the high-resourced models, but representing an improvement over $170\%$ on the task.
In a low-resourced setting, the data quality, processing and model selection is more critical than in a high-resourced scenario. We show how the characteristics of a language (such as diacritization in our case) should be taken into account in order to choose the relevant data and model to use. As an example, Twi word embeddings are significantly better when training on 742 k selected tokens than on 16 million noisy tokens, and when using a model that takes into account single character information (CWE-LP) instead of $n$-gram information (fastText).
Finally, we want to note that, even within a corpus, the quality of the data might depend on the language. Wikipedia is usually used as a high-quality freely available multilingual corpus as compared to noisier data such as Common Crawl. However, for the two languages under study, Wikipedia resulted to have too much noise: interference from other languages, text clearly written by non-native speakers, lack of diacritics and mixture of dialects. The JW300 corpus on the other hand, has been rated as high-quality by our native Yorùbá speakers, but as noisy by our native Twi speakers. In both cases, experiments confirm the conclusions.
<<</Summary and Discussion>>>
<<<Acknowledgements>>>
The authors thank Dr. Clement Odoje of the Department of Linguistics and African Languages, University of Ibadan, Nigeria and Olóyè Gbémisóyè Àrdèó for helping us with the Yorùbá translation of the WordSim-353 word pairs and Dr. Felix Y. Adu-Gyamfi and Ps. Isaac Sarfo for helping with the Twi translation. We also thank the members of the Niger-Volta Language Technologies Institute for providing us with clean Yorùbá corpus
The project on which this paper is based was partially funded by the German Federal Ministry of Education and Research under the funding code 01IW17001 (Deeplee). Responsibility for the content of this publication is with the authors.
<<</Acknowledgements>>>
<<</Title>>>
| {
"references": [
"only high-quality data helps",
"high-quality"
],
"type": "extractive"
} |
1912.02481 | Please extract a concise answer without any additional explanation for the following question based on the given text.
Question: What two architectures are used?
Context: <<<Title>>>
Massive vs. Curated Word Embeddings for Low-Resourced Languages. The Case of Yor\`ub\'a and Twi
<<<Abstract>>>
The success of several architectures to learn semantic representations from unannotated text and the availability of these kind of texts in online multilingual resources such as Wikipedia has facilitated the massive and automatic creation of resources for multiple languages. The evaluation of such resources is usually done for the high-resourced languages, where one has a smorgasbord of tasks and test sets to evaluate on. For low-resourced languages, the evaluation is more difficult and normally ignored, with the hope that the impressive capability of deep learning architectures to learn (multilingual) representations in the high-resourced setting holds in the low-resourced setting too. In this paper we focus on two African languages, Yor\`ub\'a and Twi, and compare the word embeddings obtained in this way, with word embeddings obtained from curated corpora and a language-dependent processing. We analyse the noise in the publicly available corpora, collect high quality and noisy data for the two languages and quantify the improvements that depend not only on the amount of data but on the quality too. We also use different architectures that learn word representations both from surface forms and characters to further exploit all the available information which showed to be important for these languages. For the evaluation, we manually translate the wordsim-353 word pairs dataset from English into Yor\`ub\'a and Twi. As output of the work, we provide corpora, embeddings and the test suits for both languages.
<<</Abstract>>>
<<<Introduction>>>
In recent years, word embeddings BIBREF0, BIBREF1, BIBREF2 have been proven to be very useful for training downstream natural language processing (NLP) tasks. Moreover, contextualized embeddings BIBREF3, BIBREF4 have been shown to further improve the performance of NLP tasks such as named entity recognition, question answering, or text classification when used as word features because they are able to resolve ambiguities of word representations when they appear in different contexts. Different deep learning architectures such as multilingual BERT BIBREF4, LASER BIBREF5 and XLM BIBREF6 have proved successful in the multilingual setting. All these architectures learn the semantic representations from unannotated text, making them cheap given the availability of texts in online multilingual resources such as Wikipedia. However, the evaluation of such resources is usually done for the high-resourced languages, where one has a smorgasbord of tasks and test sets to evaluate on. This is the best-case scenario, languages with tones of data for training that generate high-quality models.
For low-resourced languages, the evaluation is more difficult and therefore normally ignored simply because of the lack of resources. In these cases, training data is scarce, and the assumption that the capability of deep learning architectures to learn (multilingual) representations in the high-resourced setting holds in the low-resourced one does not need to be true. In this work, we focus on two African languages, Yorùbá and Twi, and carry out several experiments to verify this claim. Just by a simple inspection of the word embeddings trained on Wikipedia by fastText, we see a high number of non-Yorùbá or non-Twi words in the vocabularies. For Twi, the vocabulary has only 935 words, and for Yorùbá we estimate that 135 k out of the 150 k words belong to other languages such as English, French and Arabic.
In order to improve the semantic representations for these languages, we collect online texts and study the influence of the quality and quantity of the data in the final models. We also examine the most appropriate architecture depending on the characteristics of each language. Finally, we translate test sets and annotate corpora to evaluate the performance of both our models together with fastText and BERT pre-trained embeddings which could not be evaluated otherwise for Yorùbá and Twi. The evaluation is carried out in a word similarity and relatedness task using the wordsim-353 test set, and in a named entity recognition (NER) task where embeddings play a crucial role. Of course, the evaluation of the models in only two tasks is not exhaustive but it is an indication of the quality we can obtain for these two low-resourced languages as compared to others such as English where these evaluations are already available.
The rest of the paper is organized as follows. Related works are reviewed in Section SECREF2 The two languages under study are described in Section SECREF3. We introduce the corpora and test sets in Section SECREF4. The fifth section explores the different training architectures we consider, and the experiments that are carried out. Finally, discussion and concluding remarks are given in Section SECREF6
<<</Introduction>>>
<<<Related Work>>>
The large amount of freely available text in the internet for multiple languages is facilitating the massive and automatic creation of multilingual resources. The resource par excellence is Wikipedia, an online encyclopedia currently available in 307 languages. Other initiatives such as Common Crawl or the Jehovah’s Witnesses site are also repositories for multilingual data, usually assumed to be noisier than Wikipedia. Word and contextual embeddings have been pre-trained on these data, so that the resources are nowadays at hand for more than 100 languages. Some examples include fastText word embeddings BIBREF2, BIBREF7, MUSE embeddings BIBREF8, BERT multilingual embeddings BIBREF4 and LASER sentence embeddings BIBREF5. In all cases, embeddings are trained either simultaneously for multiple languages, joining high- and low-resource data, or following the same methodology.
On the other hand, different approaches try to specifically design architectures to learn embeddings in a low-resourced setting. ChaudharyEtAl:2018 follow a transfer learning approach that uses phonemes, lemmas and morphological tags to transfer the knowledge from related high-resource language into the low-resource one. jiangEtal:2018 apply Positive-Unlabeled Learning for word embedding calculations, assuming that unobserved pairs of words in a corpus also convey information, and this is specially important for small corpora.
In order to assess the quality of word embeddings, word similarity and relatedness tasks are usually used. wordsim-353 BIBREF9 is a collection of 353 pairs annotated with semantic similarity scores in a scale from 0 to 10. Even the problems detected in this dataset BIBREF10, it is widely used by the community. The test set was originally created for English, but the need for comparison with other languages has motivated several translations/adaptations. In hassanMihalcea:2009 the test was translated manually into Spanish, Romanian and Arabic and the scores were adapted to reflect similarities in the new language. The reported correlation between the English scores and the Spanish ones is 0.86. Later, JoubarneInkpen:2011 show indications that the measures of similarity highly correlate across languages. leviantReichart:2015 translated also wordsim-353 into German, Italian and Russian and used crowdsourcing to score the pairs. Finally, jiangEtal:2018 translated with Google Cloud the test set from English into Czech, Danish and Dutch. In our work, native speakers translate wordsim-353 into Yorùbá and Twi, and similarity scores are kept unless the discrepancy with English is big (see Section SECREF11 for details). A similar approach to our work is done for Gujarati in JoshiEtAl:2019.
<<</Related Work>>>
<<<Languages under Study>>>
<<<Yorùbá>>>
is a language in the West Africa with over 50 million speakers. It is spoken among other languages in Nigeria, republic of Togo, Benin Republic, Ghana and Sierra Leon. It is also a language of Òrìsà in Cuba, Brazil, and some Caribbean countries. It is one of the three major languages in Nigeria and it is regarded as the third most spoken native African language. There are different dialects of Yorùbá in Nigeria BIBREF11, BIBREF12, BIBREF13. However, in this paper our focus is the standard Yorùbá based upon a report from the 1974 Joint Consultative Committee on Education BIBREF14.
Standard Yorùbá has 25 letters without the Latin characters c, q, v, x and z. There are 18 consonants (b, d, f, g, gb, j[dz], k, l, m, n, p[kp], r, s, ṣ, t, w y[j]), 7 oral vowels (a, e, ẹ, i, o, ọ, u), five nasal vowels, (an, $ \underaccent{\dot{}}{e}$n, in, $ \underaccent{\dot{}}{o}$n, un) and syllabic nasals (m̀, ḿ, ǹ, ń). Yorùbá is a tone language which makes heavy use of lexical tones which are indicated by the use of diacritics. There are three tones in Yorùbá namely low, mid and high which are represented as grave ($\setminus $), macron ($-$) and acute ($/$) symbols respectively. These tones are applied on vowels and syllabic nasals. Mid tone is usually left unmarked on vowels and every initial or first vowel in a word cannot have a high tone. It is important to note that tone information is needed for correct pronunciation and to have the meaning of a word BIBREF15, BIBREF12, BIBREF14. For example, owó (money), ọw (broom), òwò (business), w (honour), ọw (hand), and w (group) are different words with different dots and diacritic combinations. According to Asahiah2014, Standard Yorùbá uses 4 diacritics, 3 are for marking tones while the fourth which is the dot below is used to indicate the open phonetic variants of letter "e" and "o" and the long variant of "s". Also, there are 19 single diacritic letters, 3 are marked with dots below (ẹ, ọ, ṣ) while the rest are either having the grave or acute accent. The four double diacritics are divided between the grave and the acute accent as well.
As noted in Asahiah2014, most of the Yorùbá texts found in websites or public domain repositories (i) either use the correct Yorùbá orthography or (ii) replace diacritized characters with un-diacritized ones.
This happens as a result of many factors, but most especially to the unavailability of appropriate input devices for the accurate application of the diacritical marks BIBREF11. This has led to research on restoration models for diacritics BIBREF16, but the problem is not well solved and we find that most Yorùbá text in the public domain today is not well diacritized. Wikipedia is not an exception.
<<</Yorùbá>>>
<<<Twi>>>
is an Akan language of the Central Tano Branch of the Niger Congo family of languages. It is the most widely spoken of the about 80 indigenous languages in Ghana BIBREF17. It has about 9 million native speakers and about a total of 17–18 million Ghanaians have it as either first or second language. There are two mutually intelligible dialects, Asante and Akuapem, and sub-dialectical variants which are mostly unknown to and unnoticed by non-native speakers. It is also mutually intelligible with Fante and to a large extent Bono, another of the Akan languages. It is one of, if not the, easiest to learn to speak of the indigenous Ghanaian languages. The same is however not true when it comes to reading and especially writing. This is due to a number of easily overlooked complexities in the structure of the language. First of all, similarly to Yorùbá, Twi is a tonal language but written without diacritics or accents. As a result, words which are pronounced differently and unambiguous in speech tend to be ambiguous in writing. Besides, most of such words fit interchangeably in the same context and some of them can have more than two meanings. A simple example is:
Me papa aba nti na me ne wo redi no yie no. S wo ara wo nim s me papa ba a, me suban fofor adi.
This sentence could be translated as
(i) I'm only treating you nicely because I'm in a good mood. You already know I'm a completely different person when I'm in a good mood.
(ii) I'm only treating you nicely because my dad is around. You already know I'm a completely different person when my dad comes around.
Another characteristic of Twi is the fact that a good number of stop words have the same written form as content words. For instance, “na” or “na” could be the words “and, then”, the phrase “and then” or the word “mother”. This kind of ambiguity has consequences in several natural language applications where stop words are removed from text.
Finally, we want to point out that words can also be written with or without prefixes. An example is this same na and na which happen to be the same word with an omissible prefix across its multiple senses. For some words, the prefix characters are mostly used when the word begins a sentence and omitted in the middle. This however depends on the author/speaker. For the word embeddings calculation, this implies that one would have different embeddings for the same word found in different contexts.
<<</Twi>>>
<<</Languages under Study>>>
<<<Data>>>
We collect clean and noisy corpora for Yorùbá and Twi in order to quantify the effect of noise on the quality of the embeddings, where noisy has a different meaning depending on the language as it will be explained in the next subsections.
<<<Training Corpora>>>
For Yorùbá, we use several corpora collected by the Niger-Volta Language Technologies Institute with texts from different sources, including the Lagos-NWU conversational speech corpus, fully-diacritized Yorùbá language websites and an online Bible. The largest source with clean data is the JW300 corpus. We also created our own small-sized corpus by web-crawling three Yorùbá language websites (Alàkwé, r Yorùbá and Èdè Yorùbá Rẹw in Table TABREF7), some Yoruba Tweets with full diacritics and also news corpora (BBC Yorùbá and VON Yorùbá) with poor diacritics which we use to introduce noise. By noisy corpus, we refer to texts with incorrect diacritics (e.g in BBC Yorùbá), removal of tonal symbols (e.g in VON Yorùbá) and removal of all diacritics/under-dots (e.g some articles in Yorùbá Wikipedia). Furthermore, we got two manually typed fully-diacritized Yorùbá literature (Ìrìnkèrindò nínú igbó elégbèje and Igbó Olódùmarè) both written by Daniel Orowole Olorunfemi Fagunwa a popular Yorùbá author. The number of tokens available from each source, the link to the original source and the quality of the data is summarised in Table TABREF7.
The gathering of clean data in Twi is more difficult. We use as the base text as it has been shown that the Bible is the most available resource for low and endangered languages BIBREF18. This is the cleanest of all the text we could obtain. In addition, we use the available (and small) Wikipedia dumps which are quite noisy, i.e. Wikipedia contains a good number of English words, spelling errors and Twi sentences formulated in a non-natural way (formulated as L2 speakers would speak Twi as compared to native speakers). Lastly, we added text crawled from jw and the JW300 Twi corpus. Notice that the Bible text, is mainly written in the Asante dialect whilst the last, Jehovah's Witnesses, was written mainly in the Akuapem dialect. The Wikipedia text is a mixture of the two dialects. This introduces a lot of noise into the embeddings as the spelling of most words differs especially at the end of the words due to the mixture of dialects. The JW300 Twi corpus also contains mixed dialects but is mainly Akuampem. In this case, the noise comes also from spelling errors and the uncommon addition of diacritics which are not standardised on certain vowels. Figures for Twi corpora are summarised in the bottom block of Table TABREF7.
<<</Training Corpora>>>
<<<Evaluation Test Sets>>>
<<<Yorùbá.>>>
One of the contribution of this work is the introduction of the wordsim-353 word pairs dataset for Yorùbá. All the 353 word pairs were translated from English to Yorùbá by 3 native speakers. The set is composed of 446 unique English words, 348 of which can be expressed as one-word translation in Yorùbá (e.g. book translates to ìwé). In 61 cases (most countries and locations but also other content words) translations are transliterations (e.g. Doctor is dókítà and cucumber kùkúmbà.). 98 words were translated by short phrases instead of single words. This mostly affects words from science and technology (e.g. keyboard translates to pátákó ìtwé —literally meaning typing board—, laboratory translates to ìyàrá ìṣèwádìí —research room—, and ecology translates to ìm nípa àyíká while psychology translates to ìm nípa dá). Finally, 6 terms have the same form in English and Yorùbá therefore they are retained like that in the dataset (e.g. Jazz, Rock and acronyms such as FBI or OPEC).
We also annotate the Global Voices Yorùbá corpus to test the performance of our trained Yorùbá BERT embeddings on the named entity recognition task. The corpus consists of 25 k tokens which we annotate with four named entity types: DATE, location (LOC), organization (ORG) and personal names (PER). Any other token that does not belong to the four named entities is tagged with "O". The dataset is further split into training (70%), development (10%) and test (20%) partitions. Table TABREF12 shows the number of named entities per type and partition.
<<</Yorùbá.>>>
<<</Evaluation Test Sets>>>
<<</Data>>>
<<<Semantic Representations>>>
In this section, we describe the architectures used for learning word embeddings for the Twi and Yorùbá languages. Also, we discuss the quality of the embeddings as measured by the correlation with human judgements on the translated wordSim-353 test sets and by the F1 score in a NER task.
<<<Word Embeddings Architectures>>>
Modeling sub-word units has recently become a popular way to address out-of-vocabulary word problem in NLP especially in word representation learning BIBREF19, BIBREF2, BIBREF4. A sub-word unit can be a character, character $n$-grams, or heuristically learned Byte Pair Encodings (BPE) which work very well in practice especially for morphologically rich languages. Here, we consider two word embedding models that make use of character-level information together with word information: Character Word Embedding (CWE) BIBREF20 and fastText BIBREF2. Both of them are extensions of the Word2Vec architectures BIBREF0 that model sub-word units, character embeddings in the case of CWE and character $n$-grams for fastText.
CWE was introduced in 2015 to model the embeddings of characters jointly with words in order to address the issues of character ambiguities and non-compositional words especially in the Chinese language. A word or character embedding is learned in CWE using either CBOW or skipgram architectures, and then the final word embedding is computed by adding the character embeddings to the word itself:
where $w_j$ is the word embedding of $x_j$, $N_j$ is the number of characters in $x_j$, and $c_k$ is the embedding of the $k$-th character $c_k$ in $x_j$.
Similarly, in 2017 fastText was introduced as an extension to skipgram in order to take into account morphology and improve the representation of rare words. In this case the embedding of a word also includes the embeddings of its character $n$-grams:
where $w_j$ is the word embedding of $x_j$, $G_j$ is the number of character $n$-grams in $x_j$ and $g_k$ is the embedding of the $k$-th $n$-gram.
cwe also proposed three alternatives to learn multiple embeddings per character and resolve ambiguities: (i) position-based character embeddings where each character has different embeddings depending on the position it appears in a word, i.e., beginning, middle or end (ii) cluster-based character embeddings where a character can have $K$ different cluster embeddings, and (iii) position-based cluster embeddings (CWE-LP) where for each position $K$ different embeddings are learned. We use the latter in our experiments with CWE but no positional embeddings are used with fastText.
Finally, we consider a contextualized embedding architecture, BERT BIBREF4. BERT is a masked language model based on the highly efficient and parallelizable Transformer architecture BIBREF21 known to produce very rich contextualized representations for downstream NLP tasks.
The architecture is trained by jointly conditioning on both left and right contexts in all the transformer layers using two unsupervised objectives: Masked LM and Next-sentence prediction. The representation of a word is therefore learned according to the context it is found in.
Training contextual embeddings needs of huge amounts of corpora which are not available for low-resourced languages such as Yorùbá and Twi. However, Google provided pre-trained multilingual embeddings for 102 languages including Yorùbá (but not Twi).
<<</Word Embeddings Architectures>>>
<<<Experiments>>>
<<<FastText Training and Evaluation>>>
As a first experiment, we compare the quality of fastText embeddings trained on (high-quality) curated data and (low-quality) massively extracted data for Twi and Yorùbá languages.
Facebook released pre-trained word embeddings using fastText for 294 languages trained on Wikipedia BIBREF2 (F1 in tables) and for 157 languages trained on Wikipedia and Common Crawl BIBREF7 (F2). For Yorùbá, both versions are available but only embeddings trained on Wikipedia are available for Twi. We consider these embeddings the result of training on what we call massively-extracted corpora. Notice that training settings for both embeddings are not exactly the same, and differences in performance might come both from corpus size/quality but also from the background model. The 294-languages version is trained using skipgram, in dimension 300, with character $n$-grams of length 5, a window of size 5 and 5 negatives. The 157-languages version is trained using CBOW with position-weights, in dimension 300, with character $n$-grams of length 5, a window of size 5 and 10 negatives.
We want to compare the performance of these embeddings with the equivalent models that can be obtained by training on the different sources verified by native speakers of Twi and Yorùbá; what we call curated corpora and has been described in Section SECREF4 For the comparison, we define 3 datasets according to the quality and quantity of textual data used for training: (i) Curated Small Dataset (clean), C1, about 1.6 million tokens for Yorùbá and over 735 k tokens for Twi. The clean text for Twi is the Bible and for Yoruba all texts marked under the C1 column in Table TABREF7. (ii) In Curated Small Dataset (clean + noisy), C2, we add noise to the clean corpus (Wikipedia articles for Twi, and BBC Yorùbá news articles for Yorùbá). This increases the number of training tokens for Twi to 742 k tokens and Yorùbá to about 2 million tokens. (iii) Curated Large Dataset, C3 consists of all available texts we are able to crawl and source out for, either clean or noisy. The addition of JW300 BIBREF22 texts increases the vocabulary to more than 10 k tokens in both languages.
We train our fastText systems using a skipgram model with an embedding size of 300 dimensions, context window size of 5, 10 negatives and $n$-grams ranging from 3 to 6 characters similarly to the pre-trained models for both languages. Best results are obtained with minimum word count of 3.
Table TABREF15 shows the Spearman correlation between human judgements and cosine similarity scores on the wordSim-353 test set. Notice that pre-trained embeddings on Wikipedia show a very low correlation with humans on the similarity task for both languages ($\rho $=$0.14$) and their performance is even lower when Common Crawl is also considered ($\rho $=$0.07$ for Yorùbá). An important reason for the low performance is the limited vocabulary. The pre-trained Twi model has only 935 tokens. For Yorùbá, things are apparently better with more than 150 k tokens when both Wikipedia and Common Crawl are used but correlation is even lower. An inspection of the pre-trained embeddings indicates that over 135 k words belong to other languages mostly English, French and Arabic.
If we focus only on Wikipedia, we see that many texts are without diacritics in Yorùbá and often make use of mixed dialects and English sentences in Twi.
The Spearman $\rho $ correlation for fastText models on the curated small dataset (clean), C1, improves the baselines by a large margin ($\rho =0.354$ for Twi and 0.322 for Yorùbá) even with a small dataset. The improvement could be justified just by the larger vocabulary in Twi, but in the case of Yorùbá the enhancement is there with almost half of the vocabulary size. We found out that adding some noisy texts (C2 dataset) slightly improves the correlation for Twi language but not for the Yorùbá language. The Twi language benefits from Wikipedia articles because its inclusion doubles the vocabulary and reduces the bias of the model towards religious texts. However, for Yorùbá, noisy texts often ignore diacritics or tonal marks which increases the vocabulary size at the cost of an increment in the ambiguity too. As a result, the correlation is slightly hurt. One would expect that training with more data would improve the quality of the embeddings, but we found out with the results obtained with the C3 dataset, that only high-quality data helps. The addition of JW300 boosts the vocabulary in both cases, but whereas for Twi the corpus mixes dialects and is noisy, for Yorùbá it is very clean and with full diacritics. Consequently, the best embeddings for Yorùbá are obtained when training with the C3 dataset, whereas for Twi, C2 is the best option. In both cases, the curated embeddings improve the correlation with human judgements on the similarity task a $\Delta \rho =+0.25$ or, equivalently, by an increment on $\rho $ of 170% (Twi) and 180% (Yorùbá).
<<</FastText Training and Evaluation>>>
<<<CWE Training and Evaluation>>>
The huge ambiguity in the written Twi language motivates the exploration of different approaches to word embedding estimations. In this work, we compare the standard fastText methodology to include sub-word information with the character-enhanced approach with position-based clustered embeddings (CWE-LP as introduced in Section SECREF17). With the latter, we expect to specifically address the ambiguity present in a language that does not translate the different oral tones on vowels into the written language.
The character-enhanced word embeddings are trained using a skipgram architecture with cluster-based embeddings and an embedding size of 300 dimensions, context window-size of 5, and 5 negative samples. In this case, the best performance is obtained with a minimum word count of 1, and that increases the effective vocabulary that is used for training the embeddings with respect to the fastText experiments reported in Table TABREF15.
We repeat the same experiments as with fastText and summarise them in Table TABREF16. If we compare the relative numbers for the three datasets (C1, C2 and C3) we observe the same trends as before: the performance of the embeddings in the similarity task improves with the vocabulary size when the training data can be considered clean, but the performance diminishes when the data is noisy.
According to the results, CWE is specially beneficial for Twi but not always for Yorùbá. Clean Yorùbá text, does not have the ambiguity issues at character-level, therefore the $n$-gram approximation works better when enough clean data is used ($\rho ^{C3}_{CWE}=0.354$ vs. $\rho ^{C3}_{fastText}=0.391$) but it does not when too much noisy data (no diacritics, therefore character-level information would be needed) is used ($\rho ^{C2}_{CWE}=0.345$ vs. $\rho ^{C2}_{fastText}=0.302$). For Twi, the character-level information reinforces the benefits of clean data and the best correlation with human judgements is reached with CWE embeddings ($\rho ^{C2}_{CWE}=0.437$ vs. $\rho ^{C2}_{fastText}=0.388$).
<<</CWE Training and Evaluation>>>
<<<BERT Evaluation on NER Task>>>
In order to go beyond the similarity task using static word vectors, we also investigate the quality of the multilingual BERT embeddings by fine-tuning a named entity recognition task on the Yorùbá Global Voices corpus.
One of the major advantages of pre-trained BERT embeddings is that fine-tuning of the model on downstream NLP tasks is typically computationally inexpensive, often with few number of epochs. However, the data the embeddings are trained on has the same limitations as that used in massive word embeddings. Fine-tuning involves replacing the last layer of BERT used optimizing the masked LM with a task-dependent linear classifier or any other deep learning architecture, and training all the model parameters end-to-end. For the NER task, we obtain the token-level representation from BERT and train a linear classifier for sequence tagging.
Similar to our observations with non-contextualized embeddings, we find out that fine-tuning the pre-trained multilingual-uncased BERT for 4 epochs on the NER task gives an F1 score of 0. If we do the same experiment in English, F1 is 58.1 after 4 epochs.
That shows how pre-trained embeddings by themselves do not perform well in downstream tasks on low-resource languages. To address this problem for Yorùbá, we fine-tune BERT representations on the Yorùbá corpus in two ways: (i) using the multilingual vocabulary, and (ii) using only Yorùbá vocabulary. In both cases diacritics are ignored to be consistent with the base model training.
As expected, the fine-tuning of the pre-trained BERT on the Yorùbá corpus in the two configurations generates better representations than the base model. These models are able to achieve a better performance on the NER task with an average F1 score of over 47% (see Table TABREF26 for the comparative). The fine-tuned BERT model with only Yorùbá vocabulary further increases by more than 4% in F1 score obtained with the tuning that uses the multilingual vocabulary. Although we do not have enough data to train BERT from scratch, we observe that fine-tuning BERT on a limited amount of monolingual data of a low-resource language helps to improve the quality of the embeddings. The same observation holds true for high-resource languages like German and French BIBREF23.
<<</BERT Evaluation on NER Task>>>
<<</Experiments>>>
<<</Semantic Representations>>>
<<<Summary and Discussion>>>
In this paper, we present curated word and contextual embeddings for Yorùbá and Twi. For this purpose, we gather and select corpora and study the most appropriate techniques for the languages. We also create test sets for the evaluation of the word embeddings within a word similarity task (wordsim353) and the contextual embeddings within a NER task. Corpora, embeddings and test sets are available in github.
In our analysis, we show how massively generated embeddings perform poorly for low-resourced languages as compared to the performance for high-resourced ones. This is due both to the quantity but also the quality of the data used. While the Pearson $\rho $ correlation for English obtained with fastText embeddings trained on Wikipedia (WP) and Common Crawl (CC) are $\rho _{WP}$=$0.67$ and $\rho _{WP+CC}$=$0.78$, the equivalent ones for Yorùbá are $\rho _{WP}$=$0.14$ and $\rho _{WP+CC}$=$0.07$. For Twi, only embeddings with Wikipedia are available ($\rho _{WP}$=$0.14$). By carefully gathering high-quality data and optimising the models to the characteristics of each language, we deliver embeddings with correlations of $\rho $=$0.39$ (Yorùbá) and $\rho $=$0.44$ (Twi) on the same test set, still far from the high-resourced models, but representing an improvement over $170\%$ on the task.
In a low-resourced setting, the data quality, processing and model selection is more critical than in a high-resourced scenario. We show how the characteristics of a language (such as diacritization in our case) should be taken into account in order to choose the relevant data and model to use. As an example, Twi word embeddings are significantly better when training on 742 k selected tokens than on 16 million noisy tokens, and when using a model that takes into account single character information (CWE-LP) instead of $n$-gram information (fastText).
Finally, we want to note that, even within a corpus, the quality of the data might depend on the language. Wikipedia is usually used as a high-quality freely available multilingual corpus as compared to noisier data such as Common Crawl. However, for the two languages under study, Wikipedia resulted to have too much noise: interference from other languages, text clearly written by non-native speakers, lack of diacritics and mixture of dialects. The JW300 corpus on the other hand, has been rated as high-quality by our native Yorùbá speakers, but as noisy by our native Twi speakers. In both cases, experiments confirm the conclusions.
<<</Summary and Discussion>>>
<<<Acknowledgements>>>
The authors thank Dr. Clement Odoje of the Department of Linguistics and African Languages, University of Ibadan, Nigeria and Olóyè Gbémisóyè Àrdèó for helping us with the Yorùbá translation of the WordSim-353 word pairs and Dr. Felix Y. Adu-Gyamfi and Ps. Isaac Sarfo for helping with the Twi translation. We also thank the members of the Niger-Volta Language Technologies Institute for providing us with clean Yorùbá corpus
The project on which this paper is based was partially funded by the German Federal Ministry of Education and Research under the funding code 01IW17001 (Deeplee). Responsibility for the content of this publication is with the authors.
<<</Acknowledgements>>>
<<</Title>>>
| {
"references": [
"fastText,CWE-LP"
],
"type": "extractive"
} |
2002.02224 | Please extract a concise answer without any additional explanation for the following question based on the given text.
Question: How is quality of the citation measured?
Context: <<<Title>>>
Citation Data of Czech Apex Courts
<<<Abstract>>>
In this paper, we introduce the citation data of the Czech apex courts (Supreme Court, Supreme Administrative Court and Constitutional Court). This dataset was automatically extracted from the corpus of texts of Czech court decisions - CzCDC 1.0. We obtained the citation data by building the natural language processing pipeline for extraction of the court decision identifiers. The pipeline included the (i) document segmentation model and the (ii) reference recognition model. Furthermore, the dataset was manually processed to achieve high-quality citation data as a base for subsequent qualitative and quantitative analyses. The dataset will be made available to the general public.
<<</Abstract>>>
<<<Introduction>>>
Analysis of the way court decisions refer to each other provides us with important insights into the decision-making process at courts. This is true both for the common law courts and for their counterparts in the countries belonging to the continental legal system. Citation data can be used for both qualitative and quantitative studies, casting light in the behavior of specific judges through document analysis or allowing complex studies into changing the nature of courts in transforming countries.
That being said, it is still difficult to create sufficiently large citation datasets to allow a complex research. In the case of the Czech Republic, it was difficult to obtain a relevant dataset of the court decisions of the apex courts (Supreme Court, Supreme Administrative Court and Constitutional Court). Due to its size, it is nearly impossible to extract the references manually. One has to reach out for an automation of such task. However, study of court decisions displayed many different ways that courts use to cite even decisions of their own, not to mention the decisions of other courts.The great diversity in citations led us to the use of means of the natural language processing for the recognition and the extraction of the citation data from court decisions of the Czech apex courts.
In this paper, we describe the tool ultimately used for the extraction of the references from the court decisions, together with a subsequent way of manual processing of the raw data to achieve a higher-quality dataset. Section SECREF2 maps the related work in the area of legal citation analysis (SectionSECREF1), reference recognition (Section SECREF2), text segmentation (Section SECREF4), and data availability (Section SECREF3). Section SECREF3 describes the method we used for the citation extraction, listing the individual models and the way we have combined these models into the NLP pipeline. Section SECREF4 presents results in the terms of evaluation of the performance of our pipeline, the statistics of the raw data, further manual processing and statistics of the final citation dataset. Section SECREF5 discusses limitations of our work and outlines the possible future development. Section SECREF6 concludes this paper.
<<</Introduction>>>
<<<Related work>>>
<<<Legal Citation Analysis>>>
The legal citation analysis is an emerging phenomenon in the field of the legal theory and the legal empirical research.The legal citation analysis employs tools provided by the field of network analysis.
In spite of the long-term use of the citations in the legal domain (eg. the use of Shepard's Citations since 1873), interest in the network citation analysis increased significantly when Fowler et al. published the two pivotal works on the case law citations by the Supreme Court of the United States BIBREF0, BIBREF1. Authors used the citation data and network analysis to test the hypotheses about the function of stare decisis the doctrine and other issues of legal precedents. In the continental legal system, this work was followed by Winkels and de Ruyter BIBREF2. Authors adopted similar approach to Fowler to the court decisions of the Dutch Supreme Court. Similar methods were later used by Derlén and Lindholm BIBREF3, BIBREF4 and Panagis and Šadl BIBREF5 for the citation data of the Court of Justice of the European Union, and by Olsen and Küçüksu for the citation data of the European Court of Human Rights BIBREF6.
Additionally, a minor part in research in the legal network analysis resulted in the past in practical tools designed to help lawyers conduct the case law research. Kuppevelt and van Dijck built prototypes employing these techniques in the Netherlands BIBREF7. Görög a Weisz introduced the new legal information retrieval system, Justeus, based on a large database of the legal sources and partly on the network analysis methods. BIBREF8
<<</Legal Citation Analysis>>>
<<<Reference Recognition>>>
The area of reference recognition already contains a large amount of work. It is concerned with recognizing text spans in documents that are referring to other documents. As such, it is a classical topic within the AI & Law literature.
The extraction of references from the Italian legislation based on regular expressions was reported by Palmirani et al. BIBREF9. The main goal was to bring references under a set of common standards to ensure the interoperability between different legal information systems.
De Maat et al. BIBREF10 focused on an automated detection of references to legal acts in Dutch language. Their approach consisted of a grammar covering increasingly complex citation patterns.
Opijnen BIBREF11 aimed for a reference recognition and a reference standardization using regular expressions accounting for multiple the variant of the same reference and multiple vendor-specific identifiers.
The language specific work by Kríž et al. BIBREF12 focused on the detecting and classification references to other court decisions and legal acts. Authors used a statistical recognition (HMM and Perceptron algorithms) and reported F1-measure over 90% averaged over all entities. It is the state-of-art in the automatic recognition of references in the Czech court decisions. Unfortunately, it allows only for the detection of docket numbers and it is unable to recognize court-specific or vendor-specific identifiers in the court decisions.
Other language specific-work includes our previous reference recognition model presented in BIBREF13. Prediction model is based on conditional random fields and it allows recognition of different constituents which then establish both explicit and implicit case-law and doctrinal references. Parts of this model were used in the pipeline described further within this paper in Section SECREF3.
<<</Reference Recognition>>>
<<<Data Availability>>>
Large scale quantitative and qualitative studies are often hindered by the unavailability of court data. Access to court decisions is often hindered by different obstacles. In some countries, court decisions are not available at all, while in some other they are accessible only through legal information systems, often proprietary. This effectively restricts the access to court decisions in terms of the bulk data. This issue was already approached by many researchers either through making available selected data for computational linguistics studies or by making available datasets of digitized data for various purposes. Non-exhaustive list of publicly available corpora includes British Law Report Corpus BIBREF14, The Corpus of US Supreme Court Opinions BIBREF15,the HOLJ corpus BIBREF16, the Corpus of Historical English Law Reports, Corpus de Sentencias Penales BIBREF17, Juristisches Referenzkorpus BIBREF18 and many others.
Language specific work in this area is presented by the publicly available Czech Court Decisions Corpus (CzCDC 1.0) BIBREF19. This corpus contains majority of court decisions of the Czech Supreme Court, the Supreme Administrative Court and the Constitutional Court, hence allowing a large-scale extraction of references to yield representative results. The CzCDC 1.0 was used as a dataset for extraction of the references as is described further within this paper in Section SECREF3. Unfortunately, despite containing 237 723 court decisions issued between 1st January 1993 and 30th September 2018, it is not complete. This fact is reflected in the analysis of the results.
<<</Data Availability>>>
<<<Document Segmentation>>>
A large volume of legal information is available in unstructured form, which makes processing these data a challenging task – both for human lawyers and for computers. Schweighofer BIBREF20 called for generic tools allowing a document segmentation to ease the processing of unstructured data by giving them some structure.
Topic-based segmentation often focuses on the identifying specific sentences that present borderlines of different textual segments.
The automatic segmentation is not an individual goal – it always serves as a prerequisite for further tasks requiring structured data. Segmentation is required for the text summarization BIBREF21, BIBREF22, keyword extraction BIBREF23, textual information retrieval BIBREF24, and other applications requiring input in the form of structured data.
Major part of research is focused on semantic similarity methods.The computing similarity between the parts of text presumes that a decrease of similarity means a topical border of two text segments. This approach was introduced by Hearst BIBREF22 and was used by Choi BIBREF25 and Heinonen BIBREF26 as well.
Another approach takes word frequencies and presumes a border according to different key words extracted. Reynar BIBREF27 authored graphical method based on statistics called dotplotting. Similar techniques were used by Ye BIBREF28 or Saravanan BIBREF29. Bommarito et al. BIBREF30 introduced a Python library combining different features including pre-trained models to the use for automatic legal text segmentation. Li BIBREF31 included neural network into his method to segment Chinese legal texts.
Šavelka and Ashley BIBREF32 similarly introduced the machine learning based approach for the segmentation of US court decisions texts into seven different parts. Authors reached high success rates in recognizing especially the Introduction and Analysis parts of the decisions.
Language specific work includes the model presented by Harašta et al. BIBREF33. This work focuses on segmentation of the Czech court decisions into pre-defined topical segments. Parts of this segmentation model were used in the pipeline described further within this paper in Section SECREF3.
<<</Document Segmentation>>>
<<</Related work>>>
<<<Methodology>>>
In this paper, we present and describe the citation dataset of the Czech top-tier courts. To obtain this dataset, we have processed the court decisions contained in CzCDC 1.0 dataset by the NLP pipeline consisting of the segmentation model introduced in BIBREF33, and parts of the reference recognition model presented in BIBREF13. The process is described in this section.
<<<Dataset and models>>>
<<<CzCDC 1.0 dataset>>>
Novotná and Harašta BIBREF19 prepared a dataset of the court decisions of the Czech Supreme Court, the Supreme Administrative Court and the Constitutional Court. The dataset contains 237,723 decisions published between 1st January 1993 and the 30th September 2018. These decisions are organised into three sub-corpora. The sub-corpus of the Supreme Court contains 111,977 decisions, the sub-corpus of the Supreme Administrative Court contains 52,660 decisions and the sub-corpus of the Constitutional Court contains 73,086 decisions. Authors in BIBREF19 assessed that the CzCDC currently contains approximately 91% of all decisions of the Supreme Court, 99,5% of all decisions of the Constitutional Court, and 99,9% of all decisions of the Supreme Administrative Court. As such, it presents the best currently available dataset of the Czech top-tier court decisions.
<<</CzCDC 1.0 dataset>>>
<<<Reference recognition model>>>
Harašta and Šavelka BIBREF13 introduced a reference recognition model trained specifically for the Czech top-tier courts. Moreover, authors made their training data available in the BIBREF34. Given the lack of a single citation standard, references in this work consist of smaller units, because these were identified as more uniform and therefore better suited for the automatic detection. The model was trained using conditional random fields, which is a random field model that is globally conditioned on an observation sequence O. The states of the model correspond to event labels E. Authors used a first-order conditional random fields. Model was trained for each type of the smaller unit independently.
<<</Reference recognition model>>>
<<<Text segmentation model>>>
Harašta et al. BIBREF33, authors introduced the model for the automatic segmentation of the Czech court decisions into pre-defined multi-paragraph parts. These segments include the Header (introduction of given case), History (procedural history prior the apex court proceeding), Submission/Rejoinder (petition of plaintiff and response of defendant), Argumentation (argumentation of the court hearing the case), Footer (legally required information, such as information about further proceedings), Dissent and Footnotes. The model for automatic segmentation of the text was trained using conditional random fields. The model was trained for each type independently.
<<</Text segmentation model>>>
<<</Dataset and models>>>
<<<Pipeline>>>
In order to obtain the citation data of the Czech apex courts, it was necessary to recognize and extract the references from the CzCDC 1.0. Given that training data for both the reference recognition model BIBREF13, BIBREF34 and the text segmentation model BIBREF33 are publicly available, we were able to conduct extensive error analysis and put together a pipeline to arguably achieve the maximum efficiency in the task. The pipeline described in this part is graphically represented in Figure FIGREF10.
As the first step, every document in the CzCDC 1.0 was segmented using the text segmentation model. This allowed us to treat different parts of processed court documents differently in the further text processing. Specifically, it allowed us to subject only the specific part of a court decision, in this case the court argumentation, to further the reference recognition and extraction. A textual segment recognised as the court argumentation is then processed further.
As the second step, parts recognised by the text segmentation model as a court argumentation was processed using the reference recognition model. After carefully studying the evaluation of the model's performance in BIBREF13, we have decided to use only part of the said model. Specifically, we have employed the recognition of the court identifiers, as we consider the rest of the smaller units introduced by Harašta and Šavelka of a lesser value for our task. Also, deploying only the recognition of the court identifiers allowed us to avoid the problematic parsing of smaller textual units into the references. The text spans recognised as identifiers of court decisions are then processed further.
At this point, it is necessary to evaluate the performance of the above mentioned part of the pipeline before proceeding further. The evaluation of the performance is summarised in Table TABREF11. It shows that organising the two models into the pipeline boosted the performance of the reference recognition model, leading to a higher F1 measure in the initial recognition of the text spans and their classification.
Further processing included:
control and repair of incompletely identified court identifiers (manual);
identification and sorting of identifiers as belonging to Supreme Court, Supreme Administrative Court or Constitutional Court (rule-based, manual);
standardisation of different types of court identifiers (rule-based, manual);
parsing of identifiers with court decisions available in CzCDC 1.0.
<<</Pipeline>>>
<<</Methodology>>>
<<<Results>>>
Overall, through the process described in Section SECREF3, we have retrieved three datasets of extracted references - one dataset per each of the apex courts. These datasets consist of the individual pairs containing the identification of the decision from which the reference was retrieved, and the identification of the referred documents. As we only extracted references to other judicial decisions, we obtained 471,319 references from Supreme Court decisions, 167,237 references from Supreme Administrative Court decisions and 264,463 references from Constitutional Court Decisions. These are numbers of text spans identified as references prior the further processing described in Section SECREF3.
These references include all identifiers extracted from the court decisions contained in the CzCDC 1.0. Therefore, this number includes all other court decisions, including lower courts, the Court of Justice of the European Union, the European Court of Human Rights, decisions of other public authorities etc. Therefore, it was necessary to classify these into references referring to decisions of the Supreme Court, Supreme Administrative Court, Constitutional Court and others. These groups then underwent a standardisation - or more precisely a resolution - of different court identifiers used by the Czech courts. Numbers of the references resulting from this step are shown in Table TABREF16.
Following this step, we linked court identifiers with court decisions contained in the CzCDC 1.0. Given that, the CzCDC 1.0 does not contain all the decisions of the respective courts, we were not able to parse all the references. Numbers of the references resulting from this step are shown in Table TABREF17.
<<</Results>>>
<<<Discussion>>>
This paper introduced the first dataset of citation data of the three Czech apex courts. Understandably, there are some pitfalls and limitations to our approach.
As we admitted in the evaluation in Section SECREF9, the models we included in our NLP pipelines are far from perfect. Overall, we were able to achieve a reasonable recall and precision rate, which was further enhanced by several round of manual processing of the resulting data. However, it is safe to say that we did not manage to extract all the references. Similarly, because the CzCDC 1.0 dataset we used does not contain all the decisions of the respective courts, we were not able to parse all court identifiers to the documents these refer to. Therefore, the future work in this area may include further development of the resources we used. The CzCDC 1.0 would benefit from the inclusion of more documents of the Supreme Court, the reference recognition model would benefit from more refined training methods etc.
That being said, the presented dataset is currently the only available resource of its kind focusing on the Czech court decisions that is freely available to research teams. This significantly reduces the costs necessary to conduct these types of studies involving network analysis, and the similar techniques requiring a large amount of citation data.
<<</Discussion>>>
<<<Conclusion>>>
In this paper, we have described the process of the creation of the first dataset of citation data of the three Czech apex courts. The dataset is publicly available for download at https://github.com/czech-case-law-relevance/czech-court-citations-dataset.
<<</Conclusion>>>
<<</Title>>>
| {
"references": [
"it is necessary to evaluate the performance of the above mentioned part of the pipeline before proceeding further. The evaluation of the performance is summarised in Table TABREF11. It shows that organising the two models into the pipeline boosted the performance of the reference recognition model, leading to a higher F1 measure in the initial recognition of the text spans and their classification."
],
"type": "extractive"
} |
2003.06651 | Please answer the following question with yes or no based on the given text. You only need to output 'Yes' or 'No' without any additional explanation.
Question: Is the method described in this work a clustering-based method?
Context: <<<Title>>>
Word Sense Disambiguation for 158 Languages using Word Embeddings Only
<<<Abstract>>>
Disambiguation of word senses in context is easy for humans, but is a major challenge for automatic approaches. Sophisticated supervised and knowledge-based models were developed to solve this task. However, (i) the inherent Zipfian distribution of supervised training instances for a given word and/or (ii) the quality of linguistic knowledge representations motivate the development of completely unsupervised and knowledge-free approaches to word sense disambiguation (WSD). They are particularly useful for under-resourced languages which do not have any resources for building either supervised and/or knowledge-based models. In this paper, we present a method that takes as input a standard pre-trained word embedding model and induces a fully-fledged word sense inventory, which can be used for disambiguation in context. We use this method to induce a collection of sense inventories for 158 languages on the basis of the original pre-trained fastText word embeddings by Grave et al. (2018), enabling WSD in these languages. Models and system are available online.
<<</Abstract>>>
<<<>>>
1.1em
<<</>>>
<<<Introduction>>>
There are many polysemous words in virtually any language. If not treated as such, they can hamper the performance of all semantic NLP tasks BIBREF0. Therefore, the task of resolving the polysemy and choosing the most appropriate meaning of a word in context has been an important NLP task for a long time. It is usually referred to as Word Sense Disambiguation (WSD) and aims at assigning meaning to a word in context.
The majority of approaches to WSD are based on the use of knowledge bases, taxonomies, and other external manually built resources BIBREF1, BIBREF2. However, different senses of a polysemous word occur in very diverse contexts and can potentially be discriminated with their help. The fact that semantically related words occur in similar contexts, and diverse words do not share common contexts, is known as distributional hypothesis and underlies the technique of constructing word embeddings from unlabelled texts. The same intuition can be used to discriminate between different senses of individual words. There exist methods of training word embeddings that can detect polysemous words and assign them different vectors depending on their contexts BIBREF3, BIBREF4. Unfortunately, many wide-spread word embedding models, such as GloVe BIBREF5, word2vec BIBREF6, fastText BIBREF7, do not handle polysemous words. Words in these models are represented with single vectors, which were constructed from diverse sets of contexts corresponding to different senses. In such cases, their disambiguation needs knowledge-rich approaches.
We tackle this problem by suggesting a method of post-hoc unsupervised WSD. It does not require any external knowledge and can separate different senses of a polysemous word using only the information encoded in pre-trained word embeddings. We construct a semantic similarity graph for words and partition it into densely connected subgraphs. This partition allows for separating different senses of polysemous words. Thus, the only language resource we need is a large unlabelled text corpus used to train embeddings. This makes our method applicable to under-resourced languages. Moreover, while other methods of unsupervised WSD need to train embeddings from scratch, we perform retrofitting of sense vectors based on existing word embeddings.
We create a massively multilingual application for on-the-fly word sense disambiguation. When receiving a text, the system identifies its language and performs disambiguation of all the polysemous words in it based on pre-extracted word sense inventories. The system works for 158 languages, for which pre-trained fastText embeddings available BIBREF8. The created inventories are based on these embeddings. To the best of our knowledge, our system is the only WSD system for the majority of the presented languages. Although it does not match the state of the art for resource-rich languages, it is fully unsupervised and can be used for virtually any language.
The contributions of our work are the following:
[noitemsep]
We release word sense inventories associated with fastText embeddings for 158 languages.
We release a system that allows on-the-fly word sense disambiguation for 158 languages.
We present egvi (Ego-Graph Vector Induction), a new algorithm of unsupervised word sense induction, which creates sense inventories based on pre-trained word vectors.
<<</Introduction>>>
<<<Related Work>>>
There are two main scenarios for WSD: the supervised approach that leverages training corpora explicitly labelled for word sense, and the knowledge-based approach that derives sense representation from lexical resources, such as WordNet BIBREF9. In the supervised case WSD can be treated as a classification problem. Knowledge-based approaches construct sense embeddings, i.e. embeddings that separate various word senses.
SupWSD BIBREF10 is a state-of-the-art system for supervised WSD. It makes use of linear classifiers and a number of features such as POS tags, surrounding words, local collocations, word embeddings, and syntactic relations. GlossBERT model BIBREF11, which is another implementation of supervised WSD, achieves a significant improvement by leveraging gloss information. This model benefits from sentence-pair classification approach, introduced by Devlin:19 in their BERT contextualized embedding model. The input to the model consists of a context (a sentence which contains an ambiguous word) and a gloss (sense definition) from WordNet. The context-gloss pair is concatenated through a special token ([SEP]) and classified as positive or negative.
On the other hand, sense embeddings are an alternative to traditional word vector models such as word2vec, fastText or GloVe, which represent monosemous words well but fail for ambiguous words. Sense embeddings represent individual senses of polysemous words as separate vectors. They can be linked to an explicit inventory BIBREF12 or induce a sense inventory from unlabelled data BIBREF13. LSTMEmbed BIBREF13 aims at learning sense embeddings linked to BabelNet BIBREF14, at the same time handling word ordering, and using pre-trained embeddings as an objective. Although it was tested only on English, the approach can be easily adapted to other languages present in BabelNet. However, manually labelled datasets as well as knowledge bases exist only for a small number of well-resourced languages. Thus, to disambiguate polysemous words in other languages one has to resort to fully unsupervised techniques.
The task of Word Sense Induction (WSI) can be seen as an unsupervised version of WSD. WSI aims at clustering word senses and does not require to map each cluster to a predefined sense. Instead of that, word sense inventories are induced automatically from the clusters, treating each cluster as a single sense of a word. WSI approaches fall into three main groups: context clustering, word ego-network clustering and synonyms (or substitute) clustering.
Context clustering approaches consist in creating vectors which characterise words' contexts and clustering these vectors. Here, the definition of context may vary from window-based context to latent topic-alike context. Afterwards, the resulting clusters are either used as senses directly BIBREF15, or employed further to learn sense embeddings via Chinese Restaurant Process algorithm BIBREF16, AdaGram, a Bayesian extension of the Skip-Gram model BIBREF17, AutoSense, an extension of the LDA topic model BIBREF18, and other techniques.
Word ego-network clustering is applied to semantic graphs. The nodes of a semantic graph are words, and edges between them denote semantic relatedness which is usually evaluated with cosine similarity of the corresponding embeddings BIBREF19 or by PMI-like measures BIBREF20. Word senses are induced via graph clustering algorithms, such as Chinese Whispers BIBREF21 or MaxMax BIBREF22. The technique suggested in our work belongs to this class of methods and is an extension of the method presented by Pelevina:16.
Synonyms and substitute clustering approaches create vectors which represent synonyms or substitutes of polysemous words. Such vectors are created using synonymy dictionaries BIBREF23 or context-dependent substitutes obtained from a language model BIBREF24. Analogously to previously described techniques, word senses are induced by clustering these vectors.
<<</Related Work>>>
<<<Algorithm for Word Sense Induction>>>
The majority of word vector models do not discriminate between multiple senses of individual words. However, a polysemous word can be identified via manual analysis of its nearest neighbours—they reflect different senses of the word. Table TABREF7 shows manually sense-labelled most similar terms to the word Ruby according to the pre-trained fastText model BIBREF8. As it was suggested early by Widdows:02, the distributional properties of a word can be used to construct a graph of words that are semantically related to it, and if a word is polysemous, such graph can easily be partitioned into a number of densely connected subgraphs corresponding to different senses of this word. Our algorithm is based on the same principle.
<<<SenseGram: A Baseline Graph-based Word Sense Induction Algorithm>>>
SenseGram is the method proposed by Pelevina:16 that separates nearest neighbours to induce word senses and constructs sense embeddings for each sense. It starts by constructing an ego-graph (semantic graph centred at a particular word) of the word and its nearest neighbours. The edges between the words denote their semantic relatedness, e.g. the two nodes are joined with an edge if cosine similarity of the corresponding embeddings is higher than a pre-defined threshold. The resulting graph can be clustered into subgraphs which correspond to senses of the word.
The sense vectors are then constructed by averaging embeddings of words in each resulting cluster. In order to use these sense vectors for word sense disambiguation in text, the authors compute the probabilities of sense vectors of a word given its context or the similarity of the sense vectors to the context.
<<</SenseGram: A Baseline Graph-based Word Sense Induction Algorithm>>>
<<<egvi (Ego-Graph Vector Induction): A Novel Word Sense Induction Algorithm>>>
<<<Induction of Sense Inventories>>>
One of the downsides of the described above algorithm is noise in the generated graph, namely, unrelated words and wrong connections. They hamper the separation of the graph. Another weak point is the imbalance in the nearest neighbour list, when a large part of it is attributed to the most frequent sense, not sufficiently representing the other senses. This can lead to construction of incorrect sense vectors.
We suggest a more advanced procedure of graph construction that uses the interpretability of vector addition and subtraction operations in word embedding space BIBREF6 while the previous algorithm only relies on the list of nearest neighbours in word embedding space. The key innovation of our algorithm is the use of vector subtraction to find pairs of most dissimilar graph nodes and construct the graph only from the nodes included in such “anti-edges”. Thus, our algorithm is based on graph-based word sense induction, but it also relies on vector-based operations between word embeddings to perform filtering of graph nodes. Analogously to the work of Pelevina:16, we construct a semantic relatedness graph from a list of nearest neighbours, but we filter this list using the following procedure:
Extract a list $\mathcal {N}$ = {$w_{1}$, $w_{2}$, ..., $w_{N}$} of $N$ nearest neighbours for the target (ego) word vector $w$.
Compute a list $\Delta $ = {$\delta _{1}$, $\delta _{2}$, ..., $\delta _{N}$} for each $w_{i}$ in $\mathcal {N}$, where $\delta _{i}~=~w-w_{i}$. The vectors in $\delta $ contain the components of sense of $w$ which are not related to the corresponding nearest neighbours from $\mathcal {N}$.
Compute a list $\overline{\mathcal {N}}$ = {$\overline{w_{1}}$, $\overline{w_{2}}$, ..., $\overline{w_{N}}$}, such that $\overline{w_{i}}$ is in the top nearest neighbours of $\delta _{i}$ in the embedding space. In other words, $\overline{w_{i}}$ is a word which is the most similar to the target (ego) word $w$ and least similar to its neighbour $w_{i}$. We refer to $\overline{w_{i}}$ as an anti-pair of $w_{i}$. The set of $N$ nearest neighbours and their anti-pairs form a set of anti-edges i.e. pairs of most dissimilar nodes – those which should not be connected: $\overline{E} = \lbrace (w_{1},\overline{w_{1}}), (w_{2},\overline{w_{2}}), ..., (w_{N},\overline{w_{N}})\rbrace $.
To clarify this, consider the target (ego) word $w = \textit {python}$, its top similar term $w_1 = \textit {Java}$ and the resulting anti-pair $\overline{w_i} = \textit {snake}$ which is the top related term of $\delta _1 = w - w_1$. Together they form an anti-edge $(w_i,\overline{w_i})=(\textit {Java}, \textit {snake})$ composed of a pair of semantically dissimilar terms.
Construct $V$, the set of vertices of our semantic graph $G=(V,E)$ from the list of anti-edges $\overline{E}$, with the following recurrent procedure: $V = V \cup \lbrace w_{i}, \overline{w_{i}}: w_{i} \in \mathcal {N}, \overline{w_{i}} \in \mathcal {N}\rbrace $, i.e. we add a word from the list of nearest neighbours and its anti-pair only if both of them are nearest neighbours of the original word $w$. We do not add $w$'s nearest neighbours if their anti-pairs do not belong to $\mathcal {N}$. Thus, we add only words which can help discriminating between different senses of $w$.
Construct the set of edges $E$ as follows. For each $w_{i}~\in ~\mathcal {N}$ we extract a set of its $K$ nearest neighbours $\mathcal {N}^{\prime }_{i} = \lbrace u_{1}, u_{2}, ..., u_{K}\rbrace $ and define $E = \lbrace (w_{i}, u_{j}): w_{i}~\in ~V, u_j~\in ~V, u_{j}~\in ~\mathcal {N}^{\prime }_{i}, u_{j}~\ne ~\overline{w_{i}}\rbrace $. In other words, we remove edges between a word $w_{i}$ and its nearest neighbour $u_j$ if $u_j$ is also its anti-pair. According to our hypothesis, $w_{i}$ and $\overline{w_{i}}$ belong to different senses of $w$, so they should not be connected (i.e. we never add anti-edges into $E$). Therefore, we consider any connection between them as noise and remove it.
Note that $N$ (the number of nearest neighbours for the target word $w$) and $K$ (the number of nearest neighbours of $w_{ci}$) do not have to match. The difference between these parameters is the following. $N$ defines how many words will be considered for the construction of ego-graph. On the other hand, $K$ defines the degree of relatedness between words in the ego-graph — if $K = 50$, then we will connect vertices $w$ and $u$ with an edge only if $u$ is in the list of 50 nearest neighbours of $w$. Increasing $K$ increases the graph connectivity and leads to lower granularity of senses.
According to our hypothesis, nearest neighbours of $w$ are grouped into clusters in the vector space, and each of the clusters corresponds to a sense of $w$. The described vertices selection procedure allows picking the most representative members of these clusters which are better at discriminating between the clusters. In addition to that, it helps dealing with the cases when one of the clusters is over-represented in the nearest neighbour list. In this case, many elements of such a cluster are not added to $V$ because their anti-pairs fall outside the nearest neighbour list. This also improves the quality of clustering.
After the graph construction, the clustering is performed using the Chinese Whispers algorithm BIBREF21. This is a bottom-up clustering procedure that does not require to pre-define the number of clusters, so it can correctly process polysemous words with varying numbers of senses as well as unambiguous words.
Figure FIGREF17 shows an example of the resulting pruned graph of for the word Ruby for $N = 50$ nearest neighbours in terms of the fastText cosine similarity. In contrast to the baseline method by BIBREF19 where all 50 terms are clustered, in the method presented in this section we sparsify the graph by removing 13 nodes which were not in the set of the “anti-edges” i.e. pairs of most dissimilar terms out of these 50 neighbours. Examples of anti-edges i.e. pairs of most dissimilar terms for this graph include: (Haskell, Sapphire), (Garnet, Rails), (Opal, Rubyist), (Hazel, RubyOnRails), and (Coffeescript, Opal).
<<</Induction of Sense Inventories>>>
<<<Labelling of Induced Senses>>>
We label each word cluster representing a sense to make them and the WSD results interpretable by humans. Prior systems used hypernyms to label the clusters BIBREF25, BIBREF26, e.g. “animal” in the “python (animal)”. However, neither hypernyms nor rules for their automatic extraction are available for all 158 languages. Therefore, we use a simpler method to select a keyword which would help to interpret each cluster. For each graph node $v \in V$ we count the number of anti-edges it belongs to: $count(v) = | \lbrace (w_i,\overline{w_i}) : (w_i,\overline{w_i}) \in \overline{E} \wedge (v = w_i \vee v = \overline{w_i}) \rbrace |$. A graph clustering yields a partition of $V$ into $n$ clusters: $V~=~\lbrace V_1, V_2, ..., V_n\rbrace $. For each cluster $V_i$ we define a keyword $w^{key}_i$ as the word with the largest number of anti-edges $count(\cdot )$ among words in this cluster.
<<</Labelling of Induced Senses>>>
<<<Word Sense Disambiguation>>>
We use keywords defined above to obtain vector representations of senses. In particular, we simply use word embedding of the keyword $w^{key}_i$ as a sense representation $\mathbf {s}_i$ of the target word $w$ to avoid explicit computation of sense embeddings like in BIBREF19. Given a sentence $\lbrace w_1, w_2, ..., w_{j}, w, w_{j+1}, ..., w_n\rbrace $ represented as a matrix of word vectors, we define the context of the target word $w$ as $\textbf {c}_w = \dfrac{\sum _{j=1}^{n} w_j}{n}$. Then, we define the most appropriate sense $\hat{s}$ as the sense with the highest cosine similarity to the embedding of the word's context:
<<</Word Sense Disambiguation>>>
<<</egvi (Ego-Graph Vector Induction): A Novel Word Sense Induction Algorithm>>>
<<</Algorithm for Word Sense Induction>>>
<<<System Design>>>
We release a system for on-the-fly WSD for 158 languages. Given textual input, it identifies polysemous words and retrieves senses that are the most appropriate in the context.
<<<Construction of Sense Inventories>>>
To build word sense inventories (sense vectors) for 158 languages, we utilised GPU-accelerated routines for search of similar vectors implemented in Faiss library BIBREF27. The search of nearest neighbours takes substantial time, therefore, acceleration with GPUs helps to significantly reduce the word sense construction time. To further speed up the process, we keep all intermediate results in memory, which results in substantial RAM consumption of up to 200 Gb.
The construction of word senses for all of the 158 languages takes a lot of computational resources and imposes high requirements to the hardware. For calculations, we use in parallel 10–20 nodes of the Zhores cluster BIBREF28 empowered with Nvidia Tesla V100 graphic cards. For each of the languages, we construct inventories based on 50, 100, and 200 neighbours for 100,000 most frequent words. The vocabulary was limited in order to make the computation time feasible. The construction of inventories for one language takes up to 10 hours, with $6.5$ hours on average. Building the inventories for all languages took more than 1,000 hours of GPU-accelerated computations. We release the constructed sense inventories for all the available languages. They contain all the necessary information for using them in the proposed WSD system or in other downstream tasks.
<<</Construction of Sense Inventories>>>
<<<Word Sense Disambiguation System>>>
The first text pre-processing step is language identification, for which we use the fastText language identification models by Bojanowski:17. Then the input is tokenised. For languages which use Latin, Cyrillic, Hebrew, or Greek scripts, we employ the Europarl tokeniser. For Chinese, we use the Stanford Word Segmenter BIBREF29. For Japanese, we use Mecab BIBREF30. We tokenise Vietnamese with UETsegmenter BIBREF31. All other languages are processed with the ICU tokeniser, as implemented in the PyICU project. After the tokenisation, the system analyses all the input words with pre-extracted sense inventories and defines the most appropriate sense for polysemous words.
Figure FIGREF19 shows the interface of the system. It has a textual input form. The automatically identified language of text is shown above. A click on any of the words displays a prompt (shown in black) with the most appropriate sense of a word in the specified context and the confidence score. In the given example, the word Jaguar is correctly identified as a car brand. This system is based on the system by Ustalov:18, extending it with a back-end for multiple languages, language detection, and sense browsing capabilities.
<<</Word Sense Disambiguation System>>>
<<</System Design>>>
<<<Evaluation>>>
We first evaluate our converted embedding models on multi-language lexical similarity and relatedness tasks, as a sanity check, to make sure the word sense induction process did not hurt the general performance of the embeddings. Then, we test the sense embeddings on WSD task.
<<<Lexical Similarity and Relatedness>>>
<<<Experimental Setup>>>
We use the SemR-11 datasets BIBREF32, which contain word pairs with manually assigned similarity scores from 0 (words are not related) to 10 (words are fully interchangeable) for 12 languages: English (en), Arabic (ar), German (de), Spanish (es), Farsi (fa), French (fr), Italian (it), Dutch (nl), Portuguese (pt), Russian (ru), Swedish (sv), Chinese (zh). The task is to assign relatedness scores to these pairs so that the ranking of the pairs by this score is close to the ranking defined by the oracle score. The performance is measured with Pearson correlation of the rankings. Since one word can have several different senses in our setup, we follow Remus:18 and define the relatedness score for a pair of words as the maximum cosine similarity between any of their sense vectors.
We extract the sense inventories from fastText embedding vectors. We set $N=K$ for all our experiments, i.e. the number of vertices in the graph and the maximum number of vertices' nearest neighbours match. We conduct experiments with $N=K$ set to 50, 100, and 200. For each cluster $V_i$ we create a sense vector $s_i$ by averaging vectors that belong to this cluster. We rely on the methodology of BIBREF33 shifting the generated sense vector to the direction of the original word vector: $s_i~=~\lambda ~w + (1-\lambda )~\dfrac{1}{n}~\sum _{u~\in ~V_i} cos(w, u)\cdot u, $ where, $\lambda \in [0, 1]$, $w$ is the embedding of the original word, $cos(w, u)$ is the cosine similarity between $w$ and $u$, and $n=|V_i|$. By introducing the linear combination of $w$ and $u~\in ~V_i$ we enforce the similarity of sense vectors to the original word important for this task. In addition to that, we weight $u$ by their similarity to the original word, so that more similar neighbours contribute more to the sense vector. The shifting parameter $\lambda $ is set to $0.5$, following Remus:18.
A fastText model is able to generate a vector for each word even if it is not represented in the vocabulary, due to the use of subword information. However, our system cannot assemble sense vectors for out-of-vocabulary words, for such words it returns their original fastText vector. Still, the coverage of the benchmark datasets by our vocabulary is at least 85% and approaches 100% for some languages, so we do not have to resort to this back-off strategy very often.
We use the original fastText vectors as a baseline. In this case, we compute the relatedness scores of the two words as a cosine similarity of their vectors.
<<</Experimental Setup>>>
<<<Discussion of Results>>>
We compute the relatedness scores for all benchmark datasets using our sense vectors and compare them to cosine similarity scores of original fastText vectors. The results vary for different languages. Figure FIGREF28 shows the change in Pearson correlation score when switching from the baseline fastText embeddings to our sense vectors. The new vectors significantly improve the relatedness detection for German, Farsi, Russian, and Chinese, whereas for Italian, Dutch, and Swedish the score slightly falls behind the baseline. For other languages, the performance of sense vectors is on par with regular fastText.
<<</Discussion of Results>>>
<<</Lexical Similarity and Relatedness>>>
<<<Analysis>>>
In order to see how the separation of word contexts that we perform corresponds to actual senses of polysemous words, we visualise ego-graphs produced by our method. Figure FIGREF17 shows the nearest neighbours clustering for the word Ruby, which divides the graph into five senses: Ruby-related programming tools, e.g. RubyOnRails (orange cluster), female names, e.g. Josie (magenta cluster), gems, e.g. Sapphire (yellow cluster), programming languages in general, e.g. Haskell (red cluster). Besides, this is typical for fastText embeddings featuring sub-string similarity, one can observe a cluster of different spelling of the word Ruby in green.
Analogously, the word python (see Figure FIGREF35) is divided into the senses of animals, e.g. crocodile (yellow cluster), programming languages, e.g. perl5 (magenta cluster), and conference, e.g. pycon (red cluster).
In addition, we show a qualitative analysis of senses of mouse and apple. Table TABREF38 shows nearest neighbours of the original words separated into clusters (labels for clusters were assigned manually). These inventories demonstrate clear separation of different senses, although it can be too fine-grained. For example, the first and the second cluster for mouse both refer to computer mouse, but the first one addresses the different types of computer mice, and the second one is used in the context of mouse actions. Similarly, we see that iphone and macbook are separated into two clusters. Interestingly, fastText handles typos, code-switching, and emojis by correctly associating all non-standard variants to the word they refer, and our method is able to cluster them appropriately. Both inventories were produced with $K=200$, which ensures stronger connectivity of graph. However, we see that this setting still produces too many clusters. We computed the average numbers of clusters produced by our model with $K=200$ for words from the word relatedness datasets and compared these numbers with the number of senses in WordNet for English and RuWordNet BIBREF35 for Russian (see Table TABREF37). We can see that the number of senses extracted by our method is consistently higher than the real number of senses.
We also compute the average number of senses per word for all the languages and different values of $K$ (see Figure FIGREF36). The average across languages does not change much as we increase $K$. However, for larger $K$ the average exceed the median value, indicating that more languages have lower number of senses per word. At the same time, while at smaller $K$ the maximum average number of senses per word does not exceed 6, larger values of $K$ produce outliers, e.g. English with $12.5$ senses.
Notably, there are no languages with an average number of senses less than 2, while numbers on English and Russian WordNets are considerably lower. This confirms that our method systematically over-generates senses. The presence of outliers shows that this effect cannot be eliminated by further increasing $K$, because the $i$-th nearest neighbour of a word for $i>200$ can be only remotely related to this word, even if the word is rare. Thus, our sense clustering algorithm needs a method of merging spurious senses.
<<</Analysis>>>
<<</Evaluation>>>
<<<Conclusions and Future Work>>>
We present egvi, a new algorithm for word sense induction based on graph clustering that is fully unsupervised and relies on graph operations between word vectors. We apply this algorithm to a large collection of pre-trained fastText word embeddings, releasing sense inventories for 158 languages. These inventories contain all the necessary information for constructing sense vectors and using them in downstream tasks. The sense vectors for polysemous words can be directly retrofitted with the pre-trained word embeddings and do not need any external resources. As one application of these multilingual sense inventories, we present a multilingual word sense disambiguation system that performs unsupervised and knowledge-free WSD for 158 languages without the use of any dictionary or sense-labelled corpus.
The evaluation of quality of the produced sense inventories is performed on multilingual word similarity benchmarks, showing that our sense vectors improve the scores compared to non-disambiguated word embeddings. Therefore, our system in its present state can improve WSD and downstream tasks for languages where knowledge bases, taxonomies, and annotated corpora are not available and supervised WSD models cannot be trained.
A promising direction for future work is combining distributional information from the induced sense inventories with lexical knowledge bases to improve WSD performance. Besides, we encourage the use of the produced word sense inventories in other downstream tasks.
<<</Conclusions and Future Work>>>
<<</Title>>>
| {
"references": [
"Yes"
],
"type": "boolean"
} |
2003.06651 | "Please extract a concise answer without any additional explanation for the following question based(...TRUNCATED) | {"references":["The contexts are manually labelled with WordNet senses of the target words"],"type":(...TRUNCATED) |
2003.06651 | "Please answer the following question with yes or no based on the given text. You only need to outpu(...TRUNCATED) | {
"references": [
"Yes"
],
"type": "boolean"
} |
End of preview. Expand
in Data Studio
This repository contains the data presented in HiBench: Benchmarking LLMs Capability on Hierarchical Structure Reasoning.
- Downloads last month
- 38