ImanAndrea/citation_annotations · Datasets at Hugging Face

file stringclasses 75 values	start int64 12 8.27k	end int64 49 8.32k	label stringclasses 4 values	user stringclasses 5 values	text stringlengths 2 4.16k
paper_100.txt	594	966	Coherence	Ed	Zhang et al. (2019) improves an LSTM- based encoder-decoder model with online vocabulary adaptation. For abbreviated pinyin, CoCAT (Huang et al., 2015) uses machine translation technology to reduce the number of the typing letters. Huang and Zhao (2018) propose an LSTM-based encoder-decoder approach with the concatenation of context words and abbreviated pinyin as input
paper_100.txt	1,110	1,646	Coherence	Ed	In addition, there are some works handling pinyin with typing errors. Chen and Lee (2000) investigate a typing model which handles spelling correction in sentence-based pinyin input method. CHIME (Zheng et al., 2011) is a error-tolerant Chinese pinyin input method. It finds similar pinyin which will be further ranked with Chinese specific features. Jia and Zhao (2014) propose a joint graph model to globally optimize the tasks of pinyin input method and typo correction. We leave error-tolerant pinyin input method as a future work.
paper_100.txt	2,088	2,529	Coherence	Ed	Zhang et al. (2021a) add a pinyin embedding layer and learns to predict characters from similarly pronounced candidates. PLOME (Liu et al., 2021) add two embedding layers implemented with two GRU networks to inject both pinyin and shape of characters, respectively. Xu et al. (2021) add a hierarchical encoder to inject the pinyin letters at character and sentence levels, and add a ResNet encoder to use graphic features of character image.
paper_100.txt	2,471	2,478	Unsupported claim	Ed	ResNet
paper_100.txt	1,929	1,934	Unsupported claim	Ed	BERT
paper_100.txt	1,815	1,820	Unsupported claim	Ed	BERT
paper_11.txt	797	960	Unsupported claim	Ed	alternative end-to-end approach that can tackle the problem purely cross-lingually, i.e., without involving MT, would clearly be more efficient and cost-effective
paper_13.txt	3,365	3,395	Unsupported claim	Ed	In contrast to most prior work
paper_13.txt	14	105	Unsupported claim	Ed	Few-shot learning is the problem of learning classifiers with only a few training examples.
paper_13.txt	445	932	Lacks synthesis	Ed	In recent years, there has been a surge in zeroshot and few-shot approaches to text classification. One approach (Yin et al., 2019, 2020; Halder et al., 2020;Wang et al., 2021) makes use of entailment models. Textual entailment (Dagan et al., 2006), also known as natural language inference (NLI) (Bowman et al., 2015), is the problem of predicting whether a textual premise implies a textual hypothesis in a logical sense. For example, Emma loves apples implies that Emma likes apples.
paper_13.txt	933	1,190	Lacks synthesis	Ed	The entailment approach for text classification sets the input text as the premise and the text repre-senting the label as the hypothesis. A NLI model is applied to each input pair and the entailment probability is used to identify the best matching label.
paper_13.txt	1,713	1,881	Unsupported claim	Ed	In contrast, the models typically applied in the entailment approach are Cross Attention (CA) models which need to be executed for every combination of text and label.
paper_14.txt	182	310	Unsupported claim	Ed	Unfortunately, for many languages, and especially low-resource languages, such taskspecific labelled data is often not available
paper_14.txt	2,549	2,644	Unsupported claim	Ed	as this is the only task for which high-quality data is available in a large number of language
paper_14.txt	2,833	2,980	Unsupported claim	Ed	a base understanding of syntactic structure in both the source and target language is necessary for any meaningful natural language processing task
paper_15.txt	897	902	Format	Ed	2021)
paper_15.txt	14	105	Unsupported claim	Ed	Multimodal machine translation is a cross-domain task in the filed of machine translation.
paper_16.txt	713	1,264	Lacks synthesis	Ed	Researchers recently explore the peer review domain data for a few tasks, such as PeerRead (Kang et al., 2018) for paper decision predictions, AM-PERE for proposition classification in reviews, and RR (Cheng et al., 2020) for paired-argument extraction from review-rebuttal pairs. Additionally, a meta-review dataset is introduced by Bhatia et al. (2020) without any annotation. There are also some explorations on research articles (Teufel et al., 1999;Liakata et al., 2010;Lauscher et al., 2018), which differ in nature from the peer review domain.
paper_16.txt	13	416	Coherence	Ed	To facilitate the study of text summarization, earlier datasets are mostly in the news domain with relatively short input passages, such as NYT (Sandhaus, 2008), Gigaword (Napoles et al., 2012), CNN/Daily Mail (Hermann et al., 2015), NEWSROOM (Grusky et al., 2018) and XSUM (Narayan et al., 2018). Datasets for long docu-ments include Sharma et al. (2019), Cohan et al. (2018), andFisas et al. (2016).
paper_17.txt	14	1,286	Lacks synthesis	Ed	Fully supervised event extraction. Event extraction has been studied for over a decade (Ahn, 2006;Ji and Grishman, 2008) and most traditional event extraction works follow the fully supervised setting (Nguyen et al., 2016;Sha et al., 2018;Nguyen and Nguyen, 2019;Yang et al., 2019;Lin et al., 2020;Li et al., 2020). Many of them use classification-based models and use pipeline-style frameworks to extract events (Nguyen et al., 2016;Yang et al., 2019;Wadden et al., 2019). To better leverage shared knowledge in event triggers and arguments, some works propose to incorporate global features to jointly decide triggers and arguments (Lin et al., 2020;Li et al., 2013;Yang and Mitchell, 2016). Recently, few generation-based event extraction models have been proposed. TANL (Paolini et al., 2021) treats event extraction as translation tasks between augmented natural languages. Their predicted targetaugmented language embed labels into the input passage via using brackets and vertical bar symbols, hindering the model from fully leveraging label semantics. BART-Gen is also a generation-based model focusing on documentlevel event argument extraction. Yet, similar to TANL, they solve event extraction with a pipeline, which prevents knowledge sharing across subtasks.
paper_17.txt	1,726	2,044	Lacks synthesis	Ed	Liu et al. (2020) uses a machine reading comprehension formulation to conduct event extraction in a low-resource regime. Text2Event (Lu et al., 2021), a sequence-to-structure generation paradigm, first presents events in a linearized format, and then trains a generative model to generate the linearized event sequence
paper_17.txt	1,726	2,044	Coherence	Ed	Liu et al. (2020) uses a machine reading comprehension formulation to conduct event extraction in a low-resource regime. Text2Event (Lu et al., 2021), a sequence-to-structure generation paradigm, first presents events in a linearized format, and then trains a generative model to generate the linearized event sequence
paper_17.txt	1,074	1,169	Unsupported claim	Ed	BART-Gen is also a generation-based model focusing on documentlevel event argument extraction.
paper_17.txt	1,397	1,587	Unsupported claim	Ed	However, their designs are not specific for low-resource scenarios, hence, these models can not enjoy all the benefits that DEGREE obtains for low-resource event extraction at the same time,
paper_18.txt	952	970	Format	Ed	(Li et al., 2020a;
paper_18.txt	3,495	3,574	Unsupported claim	Ed	three large-scale benchmark datasets (OntoNotes V4.0, OntoNotes V5.0, and MSRA)
paper_18.txt	3,769	3,792	Unsupported claim	Ed	medical dataset (CBLUE)
paper_19.txt	1,566	1,654	Unsupported claim	Ed	shown promising for AL in NLP due to its good qualitative and computational performance
paper_19.txt	1,801	1,824	Format	Ed	Shelmanov et al. (2021
paper_20.txt	252	631	Coherence	Ed	Following Chen et al. (2020c), other works adopt PLMs for few-shot D2T generation (Chang et al., 2021b;Su et al., 2021a). Kale and Rastogi (2020b) and Ribeiro et al. (2020) showed that PLMs using linearized representations of data can outperform graph neural networks on graph-to-text datasets, recently surpassed again by graph-based models (Ke et al., 2021;Chen et al., 2020a)
paper_20.txt	3,514	3,533	Format	Ed	Jiang et al., 2020)
paper_20.txt	1,781	1,870	Unsupported claim	Ed	Recently, have shown that using a content plan leads to improved quality of PLM outputs.
paper_20.txt	39	631	Lacks synthesis	Ed	Large neural language models pretrained on self-supervised tasks (Lewis et al., 2020;Liu et al., 2019;Devlin et al., 2019) have recently gained a lot of traction in D2T generation research (Ferreira et al., 2020). Following Chen et al. (2020c), other works adopt PLMs for few-shot D2T generation (Chang et al., 2021b;Su et al., 2021a). Kale and Rastogi (2020b) and Ribeiro et al. (2020) showed that PLMs using linearized representations of data can outperform graph neural networks on graph-to-text datasets, recently surpassed again by graph-based models (Ke et al., 2021;Chen et al., 2020a)
paper_20.txt	1,003	1,051	Format	Ed	(Heidari et al., 2021;Kale and Rastogi, 2020a;.
paper_20.txt	2,107	2,491	Lacks synthesis	Ed	Sentence ordering is the task of organizing a set of natural language sentences to increase the coherence of a text (Barzilay et al., 2001;Lapata, 2003). Several neural methods for this task were proposed, using either interactions between pairs of sentences Li and Jurafsky, 2017), global interactions (Gong et al., 2016;Wang and Wan, 2019), or combination of both (Cui et al., 2020)
paper_20.txt	4,026	4,047	Format	Ed	(Botha et al., 2018;.
paper_22.txt	169	179	Unsupported claim	Ed	STS tasks
paper_22.txt	635	964	Coherence	Ed	Specifically, Reimers and Gurevych (2019) mainly use the classification objective for an NLI dataset, and Wu et al. (2020) adopt contrastive learning to utilize self-supervision from a large corpus. Yan et al. (2021); Gao et al. (2021) incorporate a parallel corpus such as NLI datasets into their contrastive learning framework.
paper_22.txt	1,101	1,201	Unsupported claim	Ed	One related task is interpretable STS, which aims to predict chunk alignment between two sentences .
paper_22.txt	1,291	1,313	Format	Ed	(Konopík et al., 2016;
paper_22.txt	1,863	1,881	Format	Ed	(Li et al., 2020;
paper_22.txt	2,408	2,746	Coherence	Ed	. To get the solution efficiently, Cuturi (2013) provides a regularizer inspired by a probabilistic theory and then uses Sinkhorn's algorithm. Kusner et al. (2015) relax the problem to get the quadratic-time solution by removing one of the constraints, and Wu et al. (2018) introduce a kernel method to approximate the optimal transport.
paper_23.txt	1,377	1,380	Unsupported claim	Ed	GPT
paper_23.txt	1,382	1,387	Unsupported claim	Ed	GPT-3
paper_23.txt	1,854	1,967	Unsupported claim	Ed	Discriminator-based methods alleviate the training cost problem, as discriminators are easier to train than a LM.
paper_23.txt	3,910	3,915	Unsupported claim	Ed	GPT-3
paper_23.txt	4,275	4,299	Unsupported claim	Ed	myopic decoding strategy
paper_23.txt	2,272	2,283	Unsupported claim	Ed	beam search
paper_23.txt	3,593	3,626	Unsupported claim	Ed	pre-trained transformer-based LM
paper_24.txt	767	1,355	Lacks synthesis	Ed	It is crucial to understand human morality to develop beneficial AI (Soares and Fallenstein, 2017;Russell, 2019). As artificial agents live and operate among humans (Akata et al., 2020), they must be able to comprehend and recognize the moral values that drive the differences in human behavior (Gabriel, 2020). The ability to understand moral rhetoric can be instrumental for, e.g., facilitating human-agent trust (Chhogyal et al., 2019;Mehrotra et al., 2021) and engineering value-aligned sociotechnical systems (Murukannaiah et al., 2020;Serramia et al., 2020;Montes and Sierra, 2021).
paper_24.txt	1,357	1,804	Lacks synthesis	Ed	There are survey instruments to estimate individual value profiles (Schwartz, 2012;Graham et al., 2013). However, reasoning about moral values is challenging for humans (Le Dantec et al., 2009;Pommeranz et al., 2012). Further, in practical applications, e.g., to conduct meaningful conversations (Tigunova et al., 2019) or to identify online trends (Mooijman et al., 2018), artificial agents should be able to understand moral rhetoric on the fly.
paper_24.txt	1,922	1,945	Format	Ed	Mooijman et al., 2018;
paper_24.txt	1,806	2,218	Lacks synthesis	Ed	The growing capabilities of natural language processing (NLP) enable the estimation of moral rhetoric from discourse Mooijman et al., 2018;Rezapour et al., 2019;Hoover et al., 2020;Araque et al., 2020). Value classifiers can be used to identify the moral values underlying a piece of text on the fly. For instance, Mooijman et al. (2018) show that detecting moral values from tweets can predict violent protests.
paper_24.txt	2,220	2,353	Unsupported claim	Ed	Existing value classifiers are evaluated on a specific dataset, without re-training or testing the classifier on a different dataset.
paper_24.txt	2,354	2,494	Unsupported claim	Ed	This shows the ability of the classifier to predict values from text, but not the ability to transfer the learned knowledge across datasets.
paper_24.txt	4,396	4,424	Format	Ed	(BERT Devlin et al. (2019))
paper_37.txt	1,213	1,280	Format	Ed	Radford et al., 2021;Schick and Schütze, 2020a,b;Brown et al., 2020
paper_37.txt	1,544	1,572	Format	Ed	Schick and Schütze, 2020a,b)
paper_37.txt	992	1,071	Unsupported claim	Ed	they are impractical to use in real-world applications due to their model sizes
paper_37.txt	1,100	1,692	Lacks synthesis	Ed	Providing prompts or task descriptions play an vital role in improving pre-trained language models in many tasks Radford et al., 2021;Schick and Schütze, 2020a,b;Brown et al., 2020). Among them, GPT models (Radford et al., 2019;Brown et al., 2020) achieved great success in prompting or task demonstrations in NLP tasks. In light of this direction, prompt-based approaches improve small pre-trained models in few-shot text classification tasks Schick and Schütze, 2020a,b). CLIP (Radford et al., 2021) also explores prompt templates for image classification which affect zero-shot performance
paper_37.txt	49	932	Lacks synthesis	Ed	Recently, several few-shot learners on vision-language tasks were proposed including GPT (Radford et al., 2019;Brown et al., 2020), Frozen (Tsimpoukelli et al., 2021), PICa , and SimVLM . Frozen (Tsimpoukelli et al., 2021) is a large language model based on GPT-2 (Radford et al., 2019), and is transformed into a multimodal few-shot learner by extending the soft prompting to incorporate a set of images and text. Their approach shows the fewshot capability on visual question answering and image classification tasks. Similarly, PICa uses GPT-3 (Brown et al., 2020) to solve VQA tasks in a few-shot manner by providing a few in-context VQA examples. It converts images into textual descriptions so that GPT-3 can understand the images. SimVLM is trained with prefix language modeling on weakly-supervised datasets. It demonstrates its effectiveness on a zero-shot captioning task
paper_38.txt	24	794	Lacks synthesis	Ed	pre-training a transformer model on a large corpus with language modeling tasks and finetuning it on different downstream tasks has become the main transfer learning paradigm in natural language processing (Devlin et al., 2019). Notably, this paradigm requires updating and storing all the model parameters for every downstream task. As the model size proliferates (e.g., 330M parameters for BERT (Devlin et al., 2019) and 175B for GPT-3 (Brown et al., 2020)), it becomes computationally expensive and challenging to fine-tune the entire pre-trained language model (LM). Thus, it is natural to ask the question of whether we can transfer the knowledge of a pre-trained LM into downstream tasks by tuning only a small portion of its parameters with most of them freezing.
paper_38.txt	872	1,390	Lacks synthesis	Ed	One line of research (Li and Liang, 2021) suggests to augment the model with a few small trainable mod-ules and freeze the original transformer weight. Take Adapter (Houlsby et al., 2019;Pfeiffer et al., 2020a,b) and Compacter (Mahabadi et al., 2021) for example, both of them insert a small set of additional modules between each transformer layer. During fine-tuning, only these additional and taskspecific modules are trained, reducing the trainable parameters to ∼ 1-3% of the original transformer model per task.
paper_38.txt	1,434	1,921	Lacks synthesis	Ed	The GPT-3 models (Brown et al., 2020;Schick and Schütze, 2020) find that with proper manual prompts, a pre-trained LM can successfully match the fine-tuning performance of BERT models. LM-BFF (Gao et al., 2020), EFL (Wang et al., 2021), and AutoPrompt (Shin et al., 2020) further this direction by insert prompts in the input embedding layer. However, these methods rely on grid-search for a natural language-based prompt from a large search space, resulting in difficulties to optimize.
paper_38.txt	2,461	2,602	Unsupported claim	Ed	all existing prompt-tuning methods have thus far focused on task-specific prompts, making them incompatible with the traditional LM objective
paper_38.txt	2,617	2,711	Unsupported claim	Ed	it is unlikely to see many different sentences with the same prefix in the pre-training corpus
paper_39.txt	129	213	Format	Ed	(Eric et al., 2017;Wu et al., 2019; and collections of largescale annotation corpora
paper_39.txt	355	377	Format	Ed	(El Asri et al., 2017
paper_39.txt	530	533	Unsupported claim	Ed	SGD
paper_39.txt	874	909	Format	Ed	Quan et al., 2020;Lin et al., 2021)
paper_39.txt	998	1,151	Unsupported claim	Ed	vast majority of existing multilingual ToD datasets do not consider the real use cases when using a ToD system to search for local entities in a country.
paper_40.txt	396	437	Format	Ed	[Levy et al., 2017, Elsahar et al., 2018
paper_41.txt	1,890	2,370	Lacks synthesis	Ed	Previous work has shown that SNLI (Bowman et al., 2015) and MNLI (Williams et al., 2018) have annotation artifacts (e.g., negation is a strong indicator of contradictions) (Gururangan et al., 2018). The literature has also shown that simple adversarial attacks including negation cues are very effective (Naik et al., 2018;Wallace et al., 2019). Kovatchev et al. (2019) analyze 11 paraphrasing systems and show that they obtain substantially worse results when negation is present
paper_41.txt	3,201	3,272	Format	Ed	Bar-Haim et al., 2006;Giampiccolo et al., 2007;Bentivogli et al., 2009)
paper_43.txt	465	524	Unsupported claim	Ed	Machine Translation (MT) is the mainstream approach for GEC
paper_43.txt	918	966	Unsupported claim	Ed	recent powerful Transformer-based Seq2Seq model
paper_44.txt	75	92	Format	Ed	(Li et al., 2015;
paper_44.txt	521	656	Unsupported claim	Ed	As external knowledge supplements the background to the inputs and decides what to say, knowledge selection is a key ingredient in KGC.
paper_44.txt	1,224	1,592	Unsupported claim	Ed	A crucial point is, they often make assumption that the golden knowledge is distinguishable as long as the dialogue context is known, yet this is not always held true because there exists a one-to-many relationship in conversation and the past utterance history in a dialogue session is insufficient to decide the knowledge selection or the future trend of a dialogue.
paper_44.txt	2,046	2,143	Unsupported claim	Ed	In other words, there exists a mapping from one's personal memory to its selection of knowledge.
paper_45.txt	533	969	Coherence	Ed	Mathew et al. (2019) collect and handcode 6,898 counter hate comments from YouTube videos targeting Jews, Blacks and LGBT communities. Ziems et al. (2020) use a collection of hate and counter hate keywords relevant to COVID-19 and create a dataset containing 359 counter hate tweets targeting Asians. Garland et al. (2020) work with German tweets and define hate and counter speech based on the communities to which the authors belong.
paper_45.txt	1,244	1,339	Unsupported claim	Ed	Even if it were, conclusions and models from synthetic data may not transfer to the real world.
paper_45.txt	1,834	2,159	Coherence	Ed	Gao and Huang (2017) annotate hateful comments in the nested structures of 10 Fox News discussion threads. Vidgen et al. (2021) Utilizing conversational context has also been explored in text classification tasks such as sentiment analysis (Ren et al., 2016), stance (Zubiaga et al., 2018) and sarcasm (Ghosh et al., 2020).
paper_46.txt	410	415	Unsupported claim	Ed	LSTM
paper_46.txt	201	231	Unsupported claim	Ed	conditional random field (CRF)
paper_46.txt	709	824	Unsupported claim	Ed	Nested NER allows a token to belong to multiple entities, which conflicts with the plain sequence tagging framework
paper_46.txt	826	1,280	Coherence	Ed	Ju et al. (2018) proposed to use stacked LSTM-CRFs to predict from inner to outer entities. Straková et al. (2019) concatenated the BILOU tags for each token inside the nested entities, which allows the LSTM-CRF to work as for flat entities. Li et al. (2020b) reformulated nested NER as a machine reading comprehension task. Shen et al. (2021) proposed to recognize nested entities by the two-stage object detection method widely used in computer vision.
paper_46.txt	2,065	2,736	Lacks synthesis	Ed	Label Smoothing Szegedy et al. (2016) proposed the label smoothing as a regularization technique to improve the accuracy of the Inception networks on ImageNet. By explicitly assigning a small probability to non-ground-truth labels, label smoothing can prevent the models from becoming too confident about the predictions, and thus improve generalization. It turned out to be a useful alternative to the standard cross entropy loss, and has been widely adopted to fight against the over-confidence (Zoph et al., 2018;Chorowski and Jaitly, 2017;Vaswani et al., 2017), improve the model calibration (Müller et al., 2019), and denoise incorrect labels (Lukasik et al., 2020).
paper_46.txt	2,844	2,969	Unsupported claim	Ed	This is driven by the observation that entity boundaries are more ambiguous and inconsistent to annotate in NER engineering.
paper_47.txt	633	637	Format	Ed	2018
paper_49.txt	1,006	1,043	Format	Ed	Wang et al., 2016aDai and Song, 2019
paper_49.txt	177	1,179	Lacks synthesis	Ed	In contrast, Aspect-based Sentiment Analysis (ABSA) is an aspect or entity oriented fine-grained sentiment analysis task. The most three basic subtasks are Aspect Term Extraction (ATE) (Hu and Liu, 2004;Yin et al., 2016;Li et al., 2018b;Xu et al., 2018;Ma et al., 2019;Chen and Qian, 2020;, Aspect Sentiment Classification (ASC) (Wang et al., 2016b;Tang et al., 2016;Ma et al., 2017;Fan et al., 2018;Li et al., 2018a;Li et al., 2021) and Opinion Term Extraction (OTE) Cardie, 2012, 2013;Fan et al., 2019;Wu et al., 2020b). The studies solve these tasks separately and ignore the dependency between these subtasks. Therefore, some efforts devoted to couple the two subtasks and proposed effective models to jointly extract aspect-based pairs. This kind of work mainly has two tasks: Aspect and Opinion Term Co-Extraction (AOTE) (Wang et al., 2016aDai and Song, 2019; Wang and Pan, 2019;Wu et al., 2020a) and Aspect-Sentiment Pair Extraction (ASPE) (Ma et al., 2018;Li et al., 2019a,b;He et al., 2019).
paper_49.txt	1,535	1,540	Unsupported claim	Ed	BERT
paper_49.txt	1,739	1,756	Format	Ed	Wu et al., 2020a
paper_49.txt	1,865	1,944	Unsupported claim	Ed	limitations related to existing works by enriching the expressiveness of labels
paper_49.txt	1,181	2,322	Lacks synthesis	Ed	Most recently, Peng et al. (2020) first proposed the ASTE task and developed a two-stage pipeline framework to couple together aspect extraction, aspect sentiment classification and opinion extraction. To further explore this task, (Mao et al., 2021;Chen et al., 2021a) transformed ASTE to a machine reading comprehension problem and utilized the shared BERT encoder to obatin the triplets after multiple stages decoding. Another line of research focuses on designing a new tagging scheme that makes the model can extract the triplets in an endto-end fashion Wu et al., 2020a;Xu et al., 2021;Yan et al., 2021). For instance, proposed a positionaware tagging scheme, which solves the limitations related to existing works by enriching the expressiveness of labels. Wu et al. (2020a) proposed a grid tagging scheme, similar to table filling (Miwa and Sasaki, 2014;Gupta et al., 2016), to solve this task in an end-to-end manner. Yan et al. (2021) converted ASTE task into a generative formulation. However, these approaches generally ignore the relations between words and linguistic features which effectively promote the triplet extraction.
paper_50.txt	13	134	Unsupported claim	Ed	Grapheme-to-phoneme conversion (G2P) is the task of converting grapheme sequences into corresponding phoneme sequences.
paper_50.txt	134	278	Unsupported claim	Ed	Many languages have the difficulty that some grapheme sequences correspond to more than one different phoneme sequence depending on the context.
paper_50.txt	280	384	Unsupported claim	Ed	G2P plays a key role in speech and text processing systems, especially in text-to-speech (TTS) systems.
paper_50.txt	746	835	Unsupported claim	Ed	each syllable is composed of characters following the orthography rules of that language.

paper_100.txt

594

966

Coherence

Ed

Zhang et al. (2019) improves an LSTM- based encoder-decoder model with online vocabulary adaptation. For abbreviated pinyin, CoCAT (Huang et al., 2015) uses machine translation technology to reduce the number of the typing letters. Huang and Zhao (2018) propose an LSTM-based encoder-decoder approach with the concatenation of context words and abbreviated pinyin as input

paper_100.txt

1,110

1,646

Coherence

Ed

In addition, there are some works handling pinyin with typing errors. Chen and Lee (2000) investigate a typing model which handles spelling correction in sentence-based pinyin input method. CHIME (Zheng et al., 2011) is a error-tolerant Chinese pinyin input method. It finds similar pinyin which will be further ranked with Chinese specific features. Jia and Zhao (2014) propose a joint graph model to globally optimize the tasks of pinyin input method and typo correction. We leave error-tolerant pinyin input method as a future work.

paper_100.txt

2,088

2,529

Coherence

Ed

Zhang et al. (2021a) add a pinyin embedding layer and learns to predict characters from similarly pronounced candidates. PLOME (Liu et al., 2021) add two embedding layers implemented with two GRU networks to inject both pinyin and shape of characters, respectively. Xu et al. (2021) add a hierarchical encoder to inject the pinyin letters at character and sentence levels, and add a ResNet encoder to use graphic features of character image.

paper_100.txt

2,471

2,478

Unsupported claim

Ed

ResNet

paper_100.txt

1,929

1,934

Unsupported claim

Ed

BERT

paper_100.txt

1,815

1,820

Unsupported claim

Ed

BERT

paper_11.txt

797

960

Unsupported claim

Ed

alternative end-to-end approach that can tackle the problem purely cross-lingually, i.e., without involving MT, would clearly be more efficient and cost-effective

paper_13.txt

3,365

3,395

Unsupported claim

Ed

In contrast to most prior work

paper_13.txt

14

105

Unsupported claim

Ed

Few-shot learning is the problem of learning classifiers with only a few training examples.

paper_13.txt

445

932

Lacks synthesis

Ed

In recent years, there has been a surge in zeroshot and few-shot approaches to text classification. One approach (Yin et al., 2019, 2020; Halder et al., 2020;Wang et al., 2021) makes use of entailment models. Textual entailment (Dagan et al., 2006), also known as natural language inference (NLI) (Bowman et al., 2015), is the problem of predicting whether a textual premise implies a textual hypothesis in a logical sense. For example, Emma loves apples implies that Emma likes apples.

paper_13.txt

933

1,190

Lacks synthesis

Ed

The entailment approach for text classification sets the input text as the premise and the text repre-senting the label as the hypothesis. A NLI model is applied to each input pair and the entailment probability is used to identify the best matching label.

paper_13.txt

1,713

1,881

Unsupported claim

Ed

In contrast, the models typically applied in the entailment approach are Cross Attention (CA) models which need to be executed for every combination of text and label.

paper_14.txt

182

310

Unsupported claim

Ed

Unfortunately, for many languages, and especially low-resource languages, such taskspecific labelled data is often not available

paper_14.txt

2,549

2,644

Unsupported claim

Ed

as this is the only task for which high-quality data is available in a large number of language

paper_14.txt

2,833

2,980

Unsupported claim

Ed

a base understanding of syntactic structure in both the source and target language is necessary for any meaningful natural language processing task

paper_15.txt

897

902

Format

Ed

2021)

paper_15.txt

14

105

Unsupported claim

Ed

Multimodal machine translation is a cross-domain task in the filed of machine translation.

paper_16.txt

713

1,264

Lacks synthesis

Ed

Researchers recently explore the peer review domain data for a few tasks, such as PeerRead (Kang et al., 2018) for paper decision predictions, AM-PERE for proposition classification in reviews, and RR (Cheng et al., 2020) for paired-argument extraction from review-rebuttal pairs. Additionally, a meta-review dataset is introduced by Bhatia et al. (2020) without any annotation. There are also some explorations on research articles (Teufel et al., 1999;Liakata et al., 2010;Lauscher et al., 2018), which differ in nature from the peer review domain.

paper_16.txt

13

416

Coherence

Ed

To facilitate the study of text summarization, earlier datasets are mostly in the news domain with relatively short input passages, such as NYT (Sandhaus, 2008), Gigaword (Napoles et al., 2012), CNN/Daily Mail (Hermann et al., 2015), NEWSROOM (Grusky et al., 2018) and XSUM (Narayan et al., 2018). Datasets for long docu-ments include Sharma et al. (2019), Cohan et al. (2018), andFisas et al. (2016).

paper_17.txt

14

1,286

Lacks synthesis

Ed

Fully supervised event extraction. Event extraction has been studied for over a decade (Ahn, 2006;Ji and Grishman, 2008) and most traditional event extraction works follow the fully supervised setting (Nguyen et al., 2016;Sha et al., 2018;Nguyen and Nguyen, 2019;Yang et al., 2019;Lin et al., 2020;Li et al., 2020). Many of them use classification-based models and use pipeline-style frameworks to extract events (Nguyen et al., 2016;Yang et al., 2019;Wadden et al., 2019). To better leverage shared knowledge in event triggers and arguments, some works propose to incorporate global features to jointly decide triggers and arguments (Lin et al., 2020;Li et al., 2013;Yang and Mitchell, 2016). Recently, few generation-based event extraction models have been proposed. TANL (Paolini et al., 2021) treats event extraction as translation tasks between augmented natural languages. Their predicted targetaugmented language embed labels into the input passage via using brackets and vertical bar symbols, hindering the model from fully leveraging label semantics. BART-Gen is also a generation-based model focusing on documentlevel event argument extraction. Yet, similar to TANL, they solve event extraction with a pipeline, which prevents knowledge sharing across subtasks.

paper_17.txt

1,726

2,044

Lacks synthesis

Ed

Liu et al. (2020) uses a machine reading comprehension formulation to conduct event extraction in a low-resource regime. Text2Event (Lu et al., 2021), a sequence-to-structure generation paradigm, first presents events in a linearized format, and then trains a generative model to generate the linearized event sequence

paper_17.txt

1,726

2,044

Coherence

Ed

Liu et al. (2020) uses a machine reading comprehension formulation to conduct event extraction in a low-resource regime. Text2Event (Lu et al., 2021), a sequence-to-structure generation paradigm, first presents events in a linearized format, and then trains a generative model to generate the linearized event sequence

paper_17.txt

1,074

1,169

Unsupported claim

Ed

BART-Gen is also a generation-based model focusing on documentlevel event argument extraction.

paper_17.txt

1,397

1,587

Unsupported claim

Ed

However, their designs are not specific for low-resource scenarios, hence, these models can not enjoy all the benefits that DEGREE obtains for low-resource event extraction at the same time,

paper_18.txt

952

970

Format

Ed

(Li et al., 2020a;

paper_18.txt

3,495

3,574

Unsupported claim

Ed

three large-scale benchmark datasets (OntoNotes V4.0, OntoNotes V5.0, and MSRA)

paper_18.txt

3,769

3,792

Unsupported claim

Ed

medical dataset (CBLUE)

paper_19.txt

1,566

1,654

Unsupported claim

Ed

shown promising for AL in NLP due to its good qualitative and computational performance

paper_19.txt

1,801

1,824

Format

Ed

Shelmanov et al. (2021

paper_20.txt

252

631

Coherence

Ed

Following Chen et al. (2020c), other works adopt PLMs for few-shot D2T generation (Chang et al., 2021b;Su et al., 2021a). Kale and Rastogi (2020b) and Ribeiro et al. (2020) showed that PLMs using linearized representations of data can outperform graph neural networks on graph-to-text datasets, recently surpassed again by graph-based models (Ke et al., 2021;Chen et al., 2020a)

paper_20.txt

3,514

3,533

Format

Ed

Jiang et al., 2020)

paper_20.txt

1,781

1,870

Unsupported claim

Ed

Recently, have shown that using a content plan leads to improved quality of PLM outputs.

paper_20.txt

39

631

Lacks synthesis

Ed

Large neural language models pretrained on self-supervised tasks (Lewis et al., 2020;Liu et al., 2019;Devlin et al., 2019) have recently gained a lot of traction in D2T generation research (Ferreira et al., 2020). Following Chen et al. (2020c), other works adopt PLMs for few-shot D2T generation (Chang et al., 2021b;Su et al., 2021a). Kale and Rastogi (2020b) and Ribeiro et al. (2020) showed that PLMs using linearized representations of data can outperform graph neural networks on graph-to-text datasets, recently surpassed again by graph-based models (Ke et al., 2021;Chen et al., 2020a)

paper_20.txt

1,003

1,051

Format

Ed

(Heidari et al., 2021;Kale and Rastogi, 2020a;.

paper_20.txt

2,107

2,491

Lacks synthesis

Ed

Sentence ordering is the task of organizing a set of natural language sentences to increase the coherence of a text (Barzilay et al., 2001;Lapata, 2003). Several neural methods for this task were proposed, using either interactions between pairs of sentences Li and Jurafsky, 2017), global interactions (Gong et al., 2016;Wang and Wan, 2019), or combination of both (Cui et al., 2020)

paper_20.txt

4,026

4,047

Format

Ed

(Botha et al., 2018;.

paper_22.txt

169

179

Unsupported claim

Ed

STS tasks

paper_22.txt

635

964

Coherence

Ed

Specifically, Reimers and Gurevych (2019) mainly use the classification objective for an NLI dataset, and Wu et al. (2020) adopt contrastive learning to utilize self-supervision from a large corpus. Yan et al. (2021); Gao et al. (2021) incorporate a parallel corpus such as NLI datasets into their contrastive learning framework.

paper_22.txt

1,101

1,201

Unsupported claim

Ed

One related task is interpretable STS, which aims to predict chunk alignment between two sentences .

paper_22.txt

1,291

1,313

Format

Ed

(Konopík et al., 2016;

paper_22.txt

1,863

1,881

Format

Ed

(Li et al., 2020;

paper_22.txt

2,408

2,746

Coherence

Ed

. To get the solution efficiently, Cuturi (2013) provides a regularizer inspired by a probabilistic theory and then uses Sinkhorn's algorithm. Kusner et al. (2015) relax the problem to get the quadratic-time solution by removing one of the constraints, and Wu et al. (2018) introduce a kernel method to approximate the optimal transport.

paper_23.txt

1,377

1,380

Unsupported claim

Ed

GPT

paper_23.txt

1,382

1,387

Unsupported claim

Ed

GPT-3

paper_23.txt

1,854

1,967

Unsupported claim

Ed

Discriminator-based methods alleviate the training cost problem, as discriminators are easier to train than a LM.

paper_23.txt

3,910

3,915

Unsupported claim

Ed

GPT-3

paper_23.txt

4,275

4,299

Unsupported claim

Ed

myopic decoding strategy

paper_23.txt

2,272

2,283

Unsupported claim

Ed

beam search

paper_23.txt

3,593

3,626

Unsupported claim

Ed

pre-trained transformer-based LM

paper_24.txt

767

1,355

Lacks synthesis

Ed

It is crucial to understand human morality to develop beneficial AI (Soares and Fallenstein, 2017;Russell, 2019). As artificial agents live and operate among humans (Akata et al., 2020), they must be able to comprehend and recognize the moral values that drive the differences in human behavior (Gabriel, 2020). The ability to understand moral rhetoric can be instrumental for, e.g., facilitating human-agent trust (Chhogyal et al., 2019;Mehrotra et al., 2021) and engineering value-aligned sociotechnical systems (Murukannaiah et al., 2020;Serramia et al., 2020;Montes and Sierra, 2021).

paper_24.txt

1,357

1,804

Lacks synthesis

Ed

There are survey instruments to estimate individual value profiles (Schwartz, 2012;Graham et al., 2013). However, reasoning about moral values is challenging for humans (Le Dantec et al., 2009;Pommeranz et al., 2012). Further, in practical applications, e.g., to conduct meaningful conversations (Tigunova et al., 2019) or to identify online trends (Mooijman et al., 2018), artificial agents should be able to understand moral rhetoric on the fly.

paper_24.txt

1,922

1,945

Format

Ed

Mooijman et al., 2018;

paper_24.txt

1,806

2,218

Lacks synthesis

Ed

The growing capabilities of natural language processing (NLP) enable the estimation of moral rhetoric from discourse Mooijman et al., 2018;Rezapour et al., 2019;Hoover et al., 2020;Araque et al., 2020). Value classifiers can be used to identify the moral values underlying a piece of text on the fly. For instance, Mooijman et al. (2018) show that detecting moral values from tweets can predict violent protests.

paper_24.txt

2,220

2,353

Unsupported claim

Ed

Existing value classifiers are evaluated on a specific dataset, without re-training or testing the classifier on a different dataset.

paper_24.txt

2,354

2,494

Unsupported claim

Ed

This shows the ability of the classifier to predict values from text, but not the ability to transfer the learned knowledge across datasets.

paper_24.txt

4,396

4,424

Format

Ed

(BERT Devlin et al. (2019))

paper_37.txt

1,213

1,280

Format

Ed

Radford et al., 2021;Schick and Schütze, 2020a,b;Brown et al., 2020

paper_37.txt

1,544

1,572

Format

Ed

Schick and Schütze, 2020a,b)

paper_37.txt

992

1,071

Unsupported claim

Ed

they are impractical to use in real-world applications due to their model sizes

paper_37.txt

1,100

1,692

Lacks synthesis

Ed

Providing prompts or task descriptions play an vital role in improving pre-trained language models in many tasks Radford et al., 2021;Schick and Schütze, 2020a,b;Brown et al., 2020). Among them, GPT models (Radford et al., 2019;Brown et al., 2020) achieved great success in prompting or task demonstrations in NLP tasks. In light of this direction, prompt-based approaches improve small pre-trained models in few-shot text classification tasks Schick and Schütze, 2020a,b). CLIP (Radford et al., 2021) also explores prompt templates for image classification which affect zero-shot performance

paper_37.txt

49

932

Lacks synthesis

Ed

Recently, several few-shot learners on vision-language tasks were proposed including GPT (Radford et al., 2019;Brown et al., 2020), Frozen (Tsimpoukelli et al., 2021), PICa , and SimVLM . Frozen (Tsimpoukelli et al., 2021) is a large language model based on GPT-2 (Radford et al., 2019), and is transformed into a multimodal few-shot learner by extending the soft prompting to incorporate a set of images and text. Their approach shows the fewshot capability on visual question answering and image classification tasks. Similarly, PICa uses GPT-3 (Brown et al., 2020) to solve VQA tasks in a few-shot manner by providing a few in-context VQA examples. It converts images into textual descriptions so that GPT-3 can understand the images. SimVLM is trained with prefix language modeling on weakly-supervised datasets. It demonstrates its effectiveness on a zero-shot captioning task

paper_38.txt

24

794

Lacks synthesis

Ed

pre-training a transformer model on a large corpus with language modeling tasks and finetuning it on different downstream tasks has become the main transfer learning paradigm in natural language processing (Devlin et al., 2019). Notably, this paradigm requires updating and storing all the model parameters for every downstream task. As the model size proliferates (e.g., 330M parameters for BERT (Devlin et al., 2019) and 175B for GPT-3 (Brown et al., 2020)), it becomes computationally expensive and challenging to fine-tune the entire pre-trained language model (LM). Thus, it is natural to ask the question of whether we can transfer the knowledge of a pre-trained LM into downstream tasks by tuning only a small portion of its parameters with most of them freezing.

paper_38.txt

872

1,390

Lacks synthesis

Ed

One line of research (Li and Liang, 2021) suggests to augment the model with a few small trainable mod-ules and freeze the original transformer weight. Take Adapter (Houlsby et al., 2019;Pfeiffer et al., 2020a,b) and Compacter (Mahabadi et al., 2021) for example, both of them insert a small set of additional modules between each transformer layer. During fine-tuning, only these additional and taskspecific modules are trained, reducing the trainable parameters to ∼ 1-3% of the original transformer model per task.

paper_38.txt

1,434

1,921

Lacks synthesis

Ed

The GPT-3 models (Brown et al., 2020;Schick and Schütze, 2020) find that with proper manual prompts, a pre-trained LM can successfully match the fine-tuning performance of BERT models. LM-BFF (Gao et al., 2020), EFL (Wang et al., 2021), and AutoPrompt (Shin et al., 2020) further this direction by insert prompts in the input embedding layer. However, these methods rely on grid-search for a natural language-based prompt from a large search space, resulting in difficulties to optimize.

paper_38.txt

2,461

2,602

Unsupported claim

Ed

all existing prompt-tuning methods have thus far focused on task-specific prompts, making them incompatible with the traditional LM objective

paper_38.txt

2,617

2,711

Unsupported claim

Ed

it is unlikely to see many different sentences with the same prefix in the pre-training corpus

paper_39.txt

129

213

Format

Ed

(Eric et al., 2017;Wu et al., 2019; and collections of largescale annotation corpora

paper_39.txt

355

377

Format

Ed

(El Asri et al., 2017

paper_39.txt

530

533

Unsupported claim

Ed

SGD

paper_39.txt

874

909

Format

Ed

Quan et al., 2020;Lin et al., 2021)

paper_39.txt

998

1,151

Unsupported claim

Ed

vast majority of existing multilingual ToD datasets do not consider the real use cases when using a ToD system to search for local entities in a country.

paper_40.txt

396

437

Format

Ed

[Levy et al., 2017, Elsahar et al., 2018

paper_41.txt

1,890

2,370

Lacks synthesis

Ed

Previous work has shown that SNLI (Bowman et al., 2015) and MNLI (Williams et al., 2018) have annotation artifacts (e.g., negation is a strong indicator of contradictions) (Gururangan et al., 2018). The literature has also shown that simple adversarial attacks including negation cues are very effective (Naik et al., 2018;Wallace et al., 2019). Kovatchev et al. (2019) analyze 11 paraphrasing systems and show that they obtain substantially worse results when negation is present

paper_41.txt

3,201

3,272

Format

Ed

Bar-Haim et al., 2006;Giampiccolo et al., 2007;Bentivogli et al., 2009)

paper_43.txt

465

524

Unsupported claim

Ed

Machine Translation (MT) is the mainstream approach for GEC

paper_43.txt

918

966

Unsupported claim

Ed

recent powerful Transformer-based Seq2Seq model

paper_44.txt

75

92

Format

Ed

(Li et al., 2015;

paper_44.txt

521

656

Unsupported claim

Ed

As external knowledge supplements the background to the inputs and decides what to say, knowledge selection is a key ingredient in KGC.

paper_44.txt

1,224

1,592

Unsupported claim

Ed

A crucial point is, they often make assumption that the golden knowledge is distinguishable as long as the dialogue context is known, yet this is not always held true because there exists a one-to-many relationship in conversation and the past utterance history in a dialogue session is insufficient to decide the knowledge selection or the future trend of a dialogue.

paper_44.txt

2,046

2,143

Unsupported claim

Ed

In other words, there exists a mapping from one's personal memory to its selection of knowledge.

paper_45.txt

533

969

Coherence

Ed

Mathew et al. (2019) collect and handcode 6,898 counter hate comments from YouTube videos targeting Jews, Blacks and LGBT communities. Ziems et al. (2020) use a collection of hate and counter hate keywords relevant to COVID-19 and create a dataset containing 359 counter hate tweets targeting Asians. Garland et al. (2020) work with German tweets and define hate and counter speech based on the communities to which the authors belong.

paper_45.txt

1,244

1,339

Unsupported claim

Ed

Even if it were, conclusions and models from synthetic data may not transfer to the real world.

paper_45.txt

1,834

2,159

Coherence

Ed

Gao and Huang (2017) annotate hateful comments in the nested structures of 10 Fox News discussion threads. Vidgen et al. (2021) Utilizing conversational context has also been explored in text classification tasks such as sentiment analysis (Ren et al., 2016), stance (Zubiaga et al., 2018) and sarcasm (Ghosh et al., 2020).

paper_46.txt

410

415

Unsupported claim

Ed

LSTM

paper_46.txt

201

231

Unsupported claim

Ed

conditional random field (CRF)

paper_46.txt

709

824

Unsupported claim

Ed

Nested NER allows a token to belong to multiple entities, which conflicts with the plain sequence tagging framework

paper_46.txt

826

1,280

Coherence

Ed

Ju et al. (2018) proposed to use stacked LSTM-CRFs to predict from inner to outer entities. Straková et al. (2019) concatenated the BILOU tags for each token inside the nested entities, which allows the LSTM-CRF to work as for flat entities. Li et al. (2020b) reformulated nested NER as a machine reading comprehension task. Shen et al. (2021) proposed to recognize nested entities by the two-stage object detection method widely used in computer vision.

paper_46.txt

2,065

2,736

Lacks synthesis

Ed

Label Smoothing Szegedy et al. (2016) proposed the label smoothing as a regularization technique to improve the accuracy of the Inception networks on ImageNet. By explicitly assigning a small probability to non-ground-truth labels, label smoothing can prevent the models from becoming too confident about the predictions, and thus improve generalization. It turned out to be a useful alternative to the standard cross entropy loss, and has been widely adopted to fight against the over-confidence (Zoph et al., 2018;Chorowski and Jaitly, 2017;Vaswani et al., 2017), improve the model calibration (Müller et al., 2019), and denoise incorrect labels (Lukasik et al., 2020).

paper_46.txt

2,844

2,969

Unsupported claim

Ed

This is driven by the observation that entity boundaries are more ambiguous and inconsistent to annotate in NER engineering.

paper_47.txt

633

637

Format

Ed

2018

paper_49.txt

1,006

1,043

Format

Ed

Wang et al., 2016aDai and Song, 2019

paper_49.txt

177

1,179

Lacks synthesis

Ed

In contrast, Aspect-based Sentiment Analysis (ABSA) is an aspect or entity oriented fine-grained sentiment analysis task. The most three basic subtasks are Aspect Term Extraction (ATE) (Hu and Liu, 2004;Yin et al., 2016;Li et al., 2018b;Xu et al., 2018;Ma et al., 2019;Chen and Qian, 2020;, Aspect Sentiment Classification (ASC) (Wang et al., 2016b;Tang et al., 2016;Ma et al., 2017;Fan et al., 2018;Li et al., 2018a;Li et al., 2021) and Opinion Term Extraction (OTE) Cardie, 2012, 2013;Fan et al., 2019;Wu et al., 2020b). The studies solve these tasks separately and ignore the dependency between these subtasks. Therefore, some efforts devoted to couple the two subtasks and proposed effective models to jointly extract aspect-based pairs. This kind of work mainly has two tasks: Aspect and Opinion Term Co-Extraction (AOTE) (Wang et al., 2016aDai and Song, 2019; Wang and Pan, 2019;Wu et al., 2020a) and Aspect-Sentiment Pair Extraction (ASPE) (Ma et al., 2018;Li et al., 2019a,b;He et al., 2019).

paper_49.txt

1,535

1,540

Unsupported claim

Ed

BERT

paper_49.txt

1,739

1,756

Format

Ed

Wu et al., 2020a

paper_49.txt

1,865

1,944

Unsupported claim

Ed

limitations related to existing works by enriching the expressiveness of labels

paper_49.txt

1,181

2,322

Lacks synthesis

Ed

Most recently, Peng et al. (2020) first proposed the ASTE task and developed a two-stage pipeline framework to couple together aspect extraction, aspect sentiment classification and opinion extraction. To further explore this task, (Mao et al., 2021;Chen et al., 2021a) transformed ASTE to a machine reading comprehension problem and utilized the shared BERT encoder to obatin the triplets after multiple stages decoding. Another line of research focuses on designing a new tagging scheme that makes the model can extract the triplets in an endto-end fashion Wu et al., 2020a;Xu et al., 2021;Yan et al., 2021). For instance, proposed a positionaware tagging scheme, which solves the limitations related to existing works by enriching the expressiveness of labels. Wu et al. (2020a) proposed a grid tagging scheme, similar to table filling (Miwa and Sasaki, 2014;Gupta et al., 2016), to solve this task in an end-to-end manner. Yan et al. (2021) converted ASTE task into a generative formulation. However, these approaches generally ignore the relations between words and linguistic features which effectively promote the triplet extraction.

paper_50.txt

13

134

Unsupported claim

Ed

Grapheme-to-phoneme conversion (G2P) is the task of converting grapheme sequences into corresponding phoneme sequences.

paper_50.txt

134

278

Unsupported claim

Ed

Many languages have the difficulty that some grapheme sequences correspond to more than one different phoneme sequence depending on the context.

paper_50.txt

280

384

Unsupported claim

Ed

G2P plays a key role in speech and text processing systems, especially in text-to-speech (TTS) systems.

paper_50.txt

746

835

Unsupported claim

Ed

each syllable is composed of characters following the orthography rules of that language.