Datasets:

WINGNUS
/

ACL-OCL

	{
	"paper_id": "2021",
	"header": {
	"generated_with": "S2ORC 1.0.0",
	"date_generated": "2023-01-19T02:11:03.070117Z"
	},
	"title": "On the Effectiveness of Dataset Embeddings in Mono-lingual, Multi-lingual and Zero-shot Conditions",
	"authors": [
	{
	"first": "Rob",
	"middle": [],
	"last": "Van Der Goot",
	"suffix": "",
	"affiliation": {},
	"email": "robv@itu.dk"
	},
	{
	"first": "",
	"middle": [],
	"last": "Ahmet\u00fcst\u00fcn",
	"suffix": "",
	"affiliation": {},
	"email": ""
	},
	{
	"first": "Barbara",
	"middle": [],
	"last": "Plank",
	"suffix": "",
	"affiliation": {},
	"email": ""
	}
	],
	"year": "",
	"venue": null,
	"identifiers": {},
	"abstract": "Recent complementary strands of research have shown that leveraging information on the data source through encoding their properties into embeddings can lead to performance increase when training a single model on heterogeneous data sources. However, it remains unclear in which situations these dataset embeddings are most effective, because they are used in a large variety of settings, languages and tasks. Furthermore, it is usually assumed that gold information on the data source is available, and that the test data is from a distribution seen during training. In this work, we compare the effect of dataset embeddings in mono-lingual settings, multilingual settings, and with predicted data source label in a zeroshot setting. We evaluate on three morphosyntactic tasks: morphological tagging, lemmatization, and dependency parsing, and use 104 datasets, 66 languages, and two different dataset grouping strategies. Performance increases are highest when the datasets are of the same language, and we know from which distribution the test-instance is drawn. In contrast, for setups where the data is from an unseen distribution, performance increase vanishes. 1",
	"pdf_parse": {
	"paper_id": "2021",
	"_pdf_hash": "",
	"abstract": [
	{
	"text": "Recent complementary strands of research have shown that leveraging information on the data source through encoding their properties into embeddings can lead to performance increase when training a single model on heterogeneous data sources. However, it remains unclear in which situations these dataset embeddings are most effective, because they are used in a large variety of settings, languages and tasks. Furthermore, it is usually assumed that gold information on the data source is available, and that the test data is from a distribution seen during training. In this work, we compare the effect of dataset embeddings in mono-lingual settings, multilingual settings, and with predicted data source label in a zeroshot setting. We evaluate on three morphosyntactic tasks: morphological tagging, lemmatization, and dependency parsing, and use 104 datasets, 66 languages, and two different dataset grouping strategies. Performance increases are highest when the datasets are of the same language, and we know from which distribution the test-instance is drawn. In contrast, for setups where the data is from an unseen distribution, performance increase vanishes. 1",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Abstract",
	"sec_num": null
	}
	],
	"body_text": [
	{
	"text": "The performance of natural language processing systems is dependent on the amount of training data, which is often scarce. To complement existing training data, supplementary data sources can be used. Especially data annotated for the same task from other sources can be beneficial to exploit. However, because of heterogeneity in language or domain this might lead to sub-optimal performance. In early work on combining training sources, data was selected at training time (Plank and van Noord, 2011; Khan et al., 2013) for a given test set. A more nuanced way to exploit heterogeneous data is to encode properties of the language as features (Naseem et al., 2012) .",
	"cite_spans": [
	{
	"start": 474,
	"end": 501,
	"text": "(Plank and van Noord, 2011;",
	"ref_id": "BIBREF6"
	},
	{
	"start": 502,
	"end": 520,
	"text": "Khan et al., 2013)",
	"ref_id": null
	},
	{
	"start": 644,
	"end": 665,
	"text": "(Naseem et al., 2012)",
	"ref_id": "BIBREF2"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "Recently, Ammar et al. (2016) showed that encoding the language of an instance as an embedding in a neural model is beneficial for multi-lingual learning. 2 Follow-up work found that multiple datasets within the same language can also be combined by encoding their origin Ust\u00fcn et al., 2019) , thereby implicitly learning useful commonalities, while still encoding datasetspecific knowledge. These dataset embeddings are employed in groups of datasets which usually range in size from 2 to 10 datasets. However, it remains unclear in which situations these dataset embeddings thrive best.",
	"cite_spans": [
	{
	"start": 10,
	"end": 29,
	"text": "Ammar et al. (2016)",
	"ref_id": "BIBREF0"
	},
	{
	"start": 272,
	"end": 291,
	"text": "Ust\u00fcn et al., 2019)",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "Furthermore, two often overseen issues with dataset embeddings are that they are commonly learned from the gold data-source labels attached to each training and test instance and it is assumed that the test data is from a distribution which is seen during training. In many real world situations these assumptions are clearly violated. A common strategy when the test data is drawn from a different distribution as the training datasets (zero-shot), is to use a manually assigned proxy treebank Barry et al., 2019; Meechan-Maddon and Nivre, 2019) . Recent work showed that for unseen datasets in mono-lingual setups (Wagner et al., 2020) , interpolated dataset embeddings can be used to improve performance for zero-shot settings. We use automatically predicted proxy data sources instead, and focus on mono-linugal as well as cross-lingual setups.",
	"cite_spans": [
	{
	"start": 495,
	"end": 514,
	"text": "Barry et al., 2019;",
	"ref_id": null
	},
	{
	"start": 515,
	"end": 546,
	"text": "Meechan-Maddon and Nivre, 2019)",
	"ref_id": "BIBREF1"
	},
	{
	"start": 616,
	"end": 637,
	"text": "(Wagner et al., 2020)",
	"ref_id": "BIBREF11"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "In this paper, we provide an extensive evaluation of the usefulness of dataset embeddings in existing e(wi) e(d) Dataset e(ci) <w> c a t </w> Figure 1 : Overview of the model where a BiLSTM encodes the word \"cat\" in a sentence. Dataset embeddings (blue) are concatenated with the character-based word representation (green) which feeds into the contextual encoder (red). setups and beyond. More concretely, we ask: 1) What are good indicators to predict the usefulness of dataset embeddings? 2) Can we effectively use dataset embeddings in the absence of gold datasource information?",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 142,
	"end": 150,
	"text": "Figure 1",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "Dataset embeddings enable conditioning of inputs on some property of the data when training on multiple sources. They are vector representations learned during model training, with the aim to capture distinctive properties of the sources into a continuous vector, without losing their heterogeneous characteristics. Given D data sources, technically, we learn a vector representation e(d) for each data source d \u2208 D while training a single model from a group of sources. Every input instance marked with its dataset source d. For each word w i with i = 0, ..., n, the word embedding e(w i ) is concatenated with the dataset embedding e(d), and both are updated during training. Figure 1 shows the overall architecture of the model which employs a contextual encoder that uses the resulting embedding as input and outputs e(c i ) to be used for prediction.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 678,
	"end": 686,
	"text": "Figure 1",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Dataset Embeddings",
	"sec_num": "2"
	},
	{
	"text": "In this work, we copy the exact setups from the UU-Parser 2.3 and the Multi-Team tagger (\u00dcst\u00fcn et al., 2019) because they were high ranking systems in two recent shared tasks (Zeman et al., 2018; McCarthy et al., 2019) and they both showed large gains by using dataset embeddings. The UUParser is an Arc-Hybrid (Kuhlmann et al., 2011) BiLSTM (Graves and Schmidhuber, 2005) dependency parser, which exploits a dynamic oracle (Goldberg and Nivre, 2013) and supports nonprojective parsing through the use of a swap action (de Lhoneux et al., 2017) . The Multi-Team tagger performs morphological tagging (Kirov et al., 2018) and lemmatization jointly; to this end, they use a shared BiLSTM encoder and feed the output of the tagging as input for the lemmatization, which is predicted as a sequence of characters. For efficiency reasons and simplicity, we disabled the use of external embeddings 3 as well as POS embeddings for the UUParser.",
	"cite_spans": [
	{
	"start": 175,
	"end": 195,
	"text": "(Zeman et al., 2018;",
	"ref_id": "BIBREF13"
	},
	{
	"start": 196,
	"end": 218,
	"text": "McCarthy et al., 2019)",
	"ref_id": null
	},
	{
	"start": 300,
	"end": 353,
	"text": "Arc-Hybrid (Kuhlmann et al., 2011) BiLSTM (Graves and",
	"ref_id": null
	},
	{
	"start": 354,
	"end": 372,
	"text": "Schmidhuber, 2005)",
	"ref_id": null
	},
	{
	"start": 438,
	"end": 450,
	"text": "Nivre, 2013)",
	"ref_id": null
	},
	{
	"start": 519,
	"end": 544,
	"text": "(de Lhoneux et al., 2017)",
	"ref_id": null
	},
	{
	"start": 600,
	"end": 620,
	"text": "(Kirov et al., 2018)",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Experimental Setup",
	"sec_num": "2.1"
	},
	{
	"text": "It should be noted that besides the differences in the models and tasks, the setups also differ among several aspects; the version of UD data (2.3, 2.2) (Nivre et al., 2020) , type of dataset splitting used (Multi-Team always has train-devtest), and most interestingly, the dataset grouping strategies. manually designed dataset groups based on typological information, language-relatedness and empirical evidence;\u00dcst\u00fcn et al. (2019) instead propose pairs: every dataset is matched with one other dataset based on word overlap. For both of the models, we copy the exact language grouping as in the original papers 4 . For comparison of different grouping strategies, we refer to Lin et al. (2019).",
	"cite_spans": [
	{
	"start": 153,
	"end": 173,
	"text": "(Nivre et al., 2020)",
	"ref_id": "BIBREF3"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Experimental Setup",
	"sec_num": "2.1"
	},
	{
	"text": "In this work, we predict data source on the sentence level, because it matches the language switches at test-time and it improves the accuracy of the classification. 5 We use a linear support vector classifier based on word and character n-grams (without tokenization). We use this approach here because of simplicity, efficiency and they have shown to reach competitive performance for text classification tasks (Zampieri et al., 2017; Medvedeva et al., 2017; \u00c7\u00f6ltekin and Rama, 2018; Basile et al., 2018) . We performed a grid search with n \u2208 [1\u22127] and all sequential combinations (1-2, 1-3, etc.) for n-grams. For this hyper-parameter tuning, we used the eight datasets from\u00dcst\u00fcn et al. (2019), and found the most robust parameters to be 1-2 for words and 1-5 for characters. The obtained macro Table 1 : Results per task: overall average, and monolingual vs cross-lingual aggregates. #SRC: number of datasets sources; BASE: single dataset baseline, CONCAT: concatenation of datasets in group, GOLD: gold dataset ids, PRED: predicted dataset ids. The intensity of colors indicate the difference to the baseline performance.",
	"cite_spans": [
	{
	"start": 166,
	"end": 167,
	"text": "5",
	"ref_id": null
	},
	{
	"start": 413,
	"end": 436,
	"text": "(Zampieri et al., 2017;",
	"ref_id": "BIBREF12"
	},
	{
	"start": 437,
	"end": 460,
	"text": "Medvedeva et al., 2017;",
	"ref_id": null
	},
	{
	"start": 461,
	"end": 485,
	"text": "\u00c7\u00f6ltekin and Rama, 2018;",
	"ref_id": null
	},
	{
	"start": 486,
	"end": 506,
	"text": "Basile et al., 2018)",
	"ref_id": null
	}
	],
	"ref_spans": [
	{
	"start": 798,
	"end": 805,
	"text": "Table 1",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Data Source Prediction",
	"sec_num": "2.2"
	},
	{
	"text": "average F1 on all data pairs from\u00dcst\u00fcn et al. 2019is 95.42, and on all data groups from is 91.76. The performance difference can be explained by the number of datasets per group, which in the former setup is always two. To match the setup during testing, we obtain predicted dataset identifiers for the training data with 5-fold jackknifing, and use these during training.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Data Source Prediction",
	"sec_num": "2.2"
	},
	{
	"text": "We report results for all tasks in two main settings: in-dataset, for setups where we assume that input data is from a distribution present during training (3.1); and zero-shot, (3.2), a setup where this is not the case. For all reported experiments, we use Labeled Attachment Score (LAS) for parsing (Zeman et al., 2018) , F1 score for morphological tagging, and accuracy for lemmatization. We do not perform any tuning, and thus only report results on development data (if no dev-split is available we use test). As a control, we compare dataset embeddings to a simple CONCAT, training on concatenation of all the data sources from a dataset group without dataset embeddings. Reported results are average over 3 runs for the UUparser, for the Multi-Team tagger we did only a single run because of the computational costs (see Appendix for more details).",
	"cite_spans": [
	{
	"start": 301,
	"end": 321,
	"text": "(Zeman et al., 2018)",
	"ref_id": "BIBREF13"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Results",
	"sec_num": "3"
	},
	{
	"text": "The average results over all datasets are shown in Table 1 , as well as the results for mono-lingual and multi-lingual dataset groups (the full results can be found in the appendix). These are the takeaways:",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 51,
	"end": 58,
	"text": "Table 1",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "In-dataset Evaluation",
	"sec_num": "3.1"
	},
	{
	"text": "Gold Overall, gold dataset embeddings provide substantial gains (Table 1 : all). They outperform both BASE and CONCAT on all 3 tasks, which confirms previous findings \u00dcst\u00fcn et al., 2019) . Gains are largest for dependency parsing, followed by lemmatization and finally morphological tagging, where the increase is only 0.71.",
	"cite_spans": [
	{
	"start": 167,
	"end": 186,
	"text": "\u00dcst\u00fcn et al., 2019)",
	"ref_id": null
	}
	],
	"ref_spans": [
	{
	"start": 64,
	"end": 72,
	"text": "(Table 1",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "In-dataset Evaluation",
	"sec_num": "3.1"
	},
	{
	"text": "10 2 10 3 10 4 10 5 10 6",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "In-dataset Evaluation",
	"sec_num": "3.1"
	},
	{
	"text": "Datasize (#words) Figure 2 : Absolute improvement in performance between BASE and GOLD in relation to data size (in number of words in the BASE training data) in log scale. Performance difference is absolute, and measured in the default metric for each task.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 18,
	"end": 26,
	"text": "Figure 2",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "In-dataset Evaluation",
	"sec_num": "3.1"
	},
	{
	"text": "Dataset group composition Comparing the mono-lingual dataset groups with the multi-lingual groups, we can see that gold dataset embeddings improve results in both settings for 2/3 tasks. The only setup where gold dataset embeddings are not beneficial is for morphological tagging in multilingual groups, where the gains for lemmatization are also only marginal (+0.63 abs. compared to base). This may be attributed to the nature of the tasks, morphological tagging and lemmatization are more language-specific, making it difficult to transfer relevant information from another language. In Figure 2 , we plot the performance increase from BASE to using gold dataset ID's in relation to its dataset size. Unsurprisingly, the largest gains are obtained in smaller datasets (<50,000 words) for all tasks. However, especially for the morphological tasks, the largest drops are also observed in this range, and mainly happen for lowresource languages which are paired with a distant language (e.g. Akkadian (akk pisandub) and Irish (ga idt)).",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 590,
	"end": 598,
	"text": "Figure 2",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "In-dataset Evaluation",
	"sec_num": "3.1"
	},
	{
	"text": "Gold vs Predicted The PRED columns in Table 1 shows that dataset embeddings are only beneficial (i.e. outperforming BASE) for lemmatization and dependency parsing when we do not have access to the gold dataset ids but use predicted ids instead. For lemmatization, the average increase compared to BASE is only 0.31 Acc., whereas for dependency parsing, it is +1.60 LAS. PRED is mostly beneficial in multi-lingual dataset groups, which is probably because the performance of the dataset classifier (Section 2.2) is higher in these cases.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 38,
	"end": 45,
	"text": "Table 1",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "In-dataset Evaluation",
	"sec_num": "3.1"
	},
	{
	"text": "In many real-world situations, the basic assumption that data instances originate from the training data distribution is violated, and it becomes essential to find a good way of using data from other sources, like finding the best proxy source. To test whether dataset embeddings are still useful in a zero-shot setup, we perform experiments where we hold out the target dataset during training and then classify all development sentences into the other sets. We run this experiment only for dataset groups containing more than 2 datasets (11 groups and 53 datasets, for dependency parsing). As baseline we compare to a model trained on the concatenation of all other datasets from the group. This zero-shot setup is challenging. Results are expected to be overall lower. Detailed results are reported in Appendix C- Table 2 summarizes the main results. 6 We aggregate over datasets for which another in-language dataset is available within its dataset group (\"\u2203 same-lang\"), and those where this is not the case. Overall, the performance increase has almost vanished, having only an 0.07 absolute increase in LAS ('all'). This increase is void in cases no in-language data exists in the group. Only for datasets for which a same language dataset exists (\u2203 same-lang, that is, a treebank exists for the language but it comes from another distribution/domain), slight improvements are obtained. We conclude that dataset embeddings are not useful in setups when the test instances are from another distribution. For demonstration purposes, we highlight the full results of two dataset groups in Table 3 . The first (multi-lingual) group shows that datasetembeddings are mainly beneficial for languages with multiple datasets, both in the in-domain and zero-shot setting. In this particular dataset group, prediction of the embeddings performs on-par with the gold labels, probably because of the high performance of the classifier. In contrast, in the second (mono-lingual) group, the classifier scores lower, and PRED prediction performances are lower compared to GOLD. For this group, predicted dataset embeddings are outperformed by a simple dataset concatenation.",
	"cite_spans": [
	{
	"start": 854,
	"end": 855,
	"text": "6",
	"ref_id": null
	}
	],
	"ref_spans": [
	{
	"start": 817,
	"end": 824,
	"text": "Table 2",
	"ref_id": "TABREF2"
	},
	{
	"start": 1593,
	"end": 1600,
	"text": "Table 3",
	"ref_id": "TABREF4"
	}
	],
	"eq_spans": [],
	"section": "Zero-shot evaluation",
	"sec_num": "3.2"
	},
	{
	"text": "We provide an extensive evaluation of dataset embeddings in two large-scale settings where they were used successfully \u00dcst\u00fcn et al., 2019) . In setups where in-distribution training data is available, we found dataset embeddings more useful in monolingual dataset groups, compared to cross-lingual ones. In general, performance gains were the largest for 1) datasets for which another same-language datasets was available during training 2) small datasets 3) datasets which were part of a large dataset group. However, with predicted id's, their benefit is limited, contrary to gold information. When moving to zero-shot setups, the performance increases become negligible (except for some particular datasets). In particular, without in-source training data, dataset embeddings work in some cases when another treebank for the language exists; but this gain is not consistent and often small. Overall, we find dataset embeddings fail to be a viable adaptation method when no insource data is available. Hence, in many realistic out-of-distribution setups, their benefit vanishes. A Reproducability report",
	"cite_spans": [
	{
	"start": 119,
	"end": 138,
	"text": "\u00dcst\u00fcn et al., 2019)",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Conclusion",
	"sec_num": "4"
	},
	{
	"text": "The UUParser was run on two E5-2660 v3's (40 threads total, we used only 30), and took on average approximately 20 hours on a single thread per model. For three random seeds, 16 dataset groups, and approximately 5 settings (4 from Section 3.1 +1 from Section 3.2 were we only used half of the groups in two settings), the total number of models is 240. So the total computation walltime was 160 hours (approximately a week). The Multi-Team tagger was run on two Tesla V100's (one model per V100, so two models were trained in parallel). On average this took approximately 3 hours. For this setup we had 104 dataset pairs, of which 82 were unique (if the pairs consist of the same two languages, we only trained one model). For the Multi-Team tagger, we ran four setups (Section 3.1), totaling to 328 models. The total computation time was 984 hours, which divided by 2 gpus resulted in a walltime of 492 hours (approximately three weeks).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Conclusion",
	"sec_num": "4"
	},
	{
	"text": "Regarding the differences in the settings (BASE, CONCAT, GOLD and PRED), there were no clear trends in differences in run-time, even the BASE settings (where multiple models where trained for 1 dataset group) was equal in runtime compared to the other settings where one large model was trained.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Conclusion",
	"sec_num": "4"
	},
	{
	"text": "It should be noted that the UUParser uses a maximum of 15,000 sentences per epoch, and the Multi-Team tagger 500,000 words (default settings), which makes training times substantially shorter (especially for the UUParser), and reduced the memory usage. Excluding external embeddings helped us reduce the running time and memory usage even further. For the UUParser, a maximum of 8GB ram is enough for training a single model ( 4GB on average), and the Multi-Team tagger requires a minimum of 8GB of GPU RAM.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Conclusion",
	"sec_num": "4"
	},
	{
	"text": "For all the other settings and hyperparameters, we exactly replicated the original code from the Smith et al. (2018) and\u00dcst\u00fcn et al. 2019, and thus refer to their papers for experimental details. The only adaptation we made to the systems is that for the UUParser we added support for supplying the dataset information in the connlu misc-column (the adapted version is available in our repo).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Conclusion",
	"sec_num": "4"
	},
	{
	"text": "For easier analysis, we provide average scores over aggregates of datasets. To this end, we propose to use data-filters, and report average scores over specific subsets of the data. The results are shown in Table 4 . The filters show aggregates over a) whether the training portion of the dataset is small (< 30,000 words) or large b) whether the dataset group to which this dataset belongs is mono-lingual or multi-lingual c) whether another dataset with the same language is available in the dataset group d) whether the svm classifier predicts the datasource id's with an accuracy of >95% accuracy e) whether the word overlap is larger then 10%. Table 5 shows the results of the UUParser for each dataset, grouped by dataset groups. All results are the average over three runs. We do not report scores for datasets without insource training data in the 'parser setting' columns (which corresponds to Section 3.1 of the paper), as training data is necessary for those settings.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 207,
	"end": 214,
	"text": "Table 4",
	"ref_id": "TABREF7"
	},
	{
	"start": 649,
	"end": 656,
	"text": "Table 5",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "B Aggregates over all results",
	"sec_num": null
	},
	{
	"text": "For the 'without train' columns (corresponding to Section 3.2 of the paper), we do not include results for dataset groups of size two; this is because we leave one training set out, and try to predict for the corresponding development set in which set it belongs. For groups of size two, this classification is trivial and non-informative, as there is only one dataset left. The left-out datasets are not taken into account for the averages. The reported results are on development splits, except for datasets which did not have a development split available, there we used test (indicated with * in the table) as we do not perform any tuning. (Pearson, 1901) . This is done with the default sklearn settings (Pedregosa et al., 2011) . Results of the gold spaces are shown in Figure 3 and the predicted spaces are plotted in Figure 4 . For some groups, there are some clear differences, however for others the plots are highly similar. There seems to be no clear trend in the amount of differences and the performance shifts in Table 5 . Table 5 : LAS scores on all development splits of the dependency parser. *: datasets for which no development data was available, we report results on test data instead. The 'in-dataset setting' results corresponds to Section 3, and 'zero-shot' to Section 3.2, where we assume no in-source training data. SIZE: size of training data in words. SVM: F1 score of svm classifier on dataset prediction. BASE: single dataset baseline. CONCAT: concatenation of datasets. GOLD: gold dataset embeddings. PRED: predicted dataset embeddings. Table 6 : Joint morphological tagging and lemmatization results for all datasets. First column is the dataset for which the results are reported, the second column is the 'helper' dataset. SIZE: size of training data for the target dataset in words. WO: % word overlap with 'helper' dataset. SVM: F1 score of svm classifier on dataset prediction. BASE: single dataset baseline. CONCAT: simple concatenation of datasets. GOLD: using gold dataset embeddings. PRED: predicted dataset embeddings. ",
	"cite_spans": [
	{
	"start": 644,
	"end": 659,
	"text": "(Pearson, 1901)",
	"ref_id": "BIBREF4"
	},
	{
	"start": 709,
	"end": 733,
	"text": "(Pedregosa et al., 2011)",
	"ref_id": "BIBREF5"
	}
	],
	"ref_spans": [
	{
	"start": 776,
	"end": 784,
	"text": "Figure 3",
	"ref_id": "FIGREF2"
	},
	{
	"start": 825,
	"end": 833,
	"text": "Figure 4",
	"ref_id": "FIGREF3"
	},
	{
	"start": 1028,
	"end": 1035,
	"text": "Table 5",
	"ref_id": null
	},
	{
	"start": 1038,
	"end": 1045,
	"text": "Table 5",
	"ref_id": null
	},
	{
	"start": 1569,
	"end": 1576,
	"text": "Table 6",
	"ref_id": "TABREF6"
	}
	],
	"eq_spans": [],
	"section": "C Full results for dependency parsing",
	"sec_num": null
	},
	{
	"text": "More recently, (Conneau and Lample, 2019) showed that embedding the language can also be beneficial for training contextualized embeddings with masked language modeling.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "3\u00dc st\u00fcn et al.(2019) showed that performance gains from external embeddings are highly complementary to performance gains from dataset embeddings.4 The full groups can be seen in Appendix C and D.5 However, Bhat et al. (2017) andRavishankar (2018) have shown the usefulness of word-level language labels for processing code-switched data.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "Note that BASE and GOLD are not reported here, because no data from the target dataset is included in this experiments.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	}
	],
	"back_matter": [
	{
	"text": "We would like to thank Gertjan van Noord for feedback on an earlier version of this paper. We thank NVIDIA, the HPC cluster at the ITU and the University of Groningen for computing resources. This research was supported by an Amazon Research Award, an STSM in the Multi3Generation COST action (CA18231), and grant 9063-00077B (Danmarks Frie Forskningsfond).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Acknowledgements",
	"sec_num": null
	}
	],
	"bib_entries": {
	"BIBREF0": {
	"ref_id": "b0",
	"title": "Many languages, one parser",
	"authors": [
	{
	"first": "Waleed",
	"middle": [],
	"last": "Ammar",
	"suffix": ""
	},
	{
	"first": "George",
	"middle": [],
	"last": "Mulcaire",
	"suffix": ""
	},
	{
	"first": "Miguel",
	"middle": [],
	"last": "Ballesteros",
	"suffix": ""
	},
	{
	"first": "Chris",
	"middle": [],
	"last": "Dyer",
	"suffix": ""
	},
	{
	"first": "Noah",
	"middle": [
	"A"
	],
	"last": "Smith",
	"suffix": ""
	}
	],
	"year": 2016,
	"venue": "Transactions of the Association for Computational Linguistics",
	"volume": "4",
	"issue": "",
	"pages": "431--444",
	"other_ids": {
	"DOI": [
	"10.1162/tacl_a_00109"
	]
	},
	"num": null,
	"urls": [],
	"raw_text": "Waleed Ammar, George Mulcaire, Miguel Ballesteros, Chris Dyer, and Noah A. Smith. 2016. Many lan- guages, one parser. Transactions of the Association for Computational Linguistics, 4:431-444.",
	"links": null
	},
	"BIBREF1": {
	"ref_id": "b1",
	"title": "How to parse low-resource languages: Cross-lingual parsing, target language annotation, or both?",
	"authors": [
	{
	"first": "Ailsa",
	"middle": [],
	"last": "Meechan-Maddon",
	"suffix": ""
	},
	{
	"first": "Joakim",
	"middle": [],
	"last": "Nivre",
	"suffix": ""
	}
	],
	"year": 2019,
	"venue": "Proceedings of the Fifth International Conference on Dependency Linguistics (Depling, SyntaxFest 2019)",
	"volume": "",
	"issue": "",
	"pages": "112--120",
	"other_ids": {
	"DOI": [
	"10.18653/v1/W19-7713"
	]
	},
	"num": null,
	"urls": [],
	"raw_text": "Ailsa Meechan-Maddon and Joakim Nivre. 2019. How to parse low-resource languages: Cross-lingual pars- ing, target language annotation, or both? In Pro- ceedings of the Fifth International Conference on Dependency Linguistics (Depling, SyntaxFest 2019), pages 112-120, Paris, France. Association for Com- putational Linguistics.",
	"links": null
	},
	"BIBREF2": {
	"ref_id": "b2",
	"title": "Selective sharing for multilingual dependency parsing",
	"authors": [
	{
	"first": "Tahira",
	"middle": [],
	"last": "Naseem",
	"suffix": ""
	},
	{
	"first": "Regina",
	"middle": [],
	"last": "Barzilay",
	"suffix": ""
	},
	{
	"first": "Amir",
	"middle": [],
	"last": "Globerson",
	"suffix": ""
	}
	],
	"year": 2012,
	"venue": "Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics",
	"volume": "1",
	"issue": "",
	"pages": "629--637",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Tahira Naseem, Regina Barzilay, and Amir Globerson. 2012. Selective sharing for multilingual dependency parsing. In Proceedings of the 50th Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers), pages 629-637, Jeju Is- land, Korea. Association for Computational Linguis- tics.",
	"links": null
	},
	"BIBREF3": {
	"ref_id": "b3",
	"title": "Universal dependencies v2: An evergrowing multilingual treebank collection",
	"authors": [
	{
	"first": "Joakim",
	"middle": [],
	"last": "Nivre",
	"suffix": ""
	},
	{
	"first": "Marie-Catherine",
	"middle": [],
	"last": "De Marneffe",
	"suffix": ""
	},
	{
	"first": "Filip",
	"middle": [],
	"last": "Ginter",
	"suffix": ""
	},
	{
	"first": "Jan",
	"middle": [],
	"last": "Hajic",
	"suffix": ""
	},
	{
	"first": "Christopher",
	"middle": [
	"D"
	],
	"last": "Manning",
	"suffix": ""
	},
	{
	"first": "Sampo",
	"middle": [],
	"last": "Pyysalo",
	"suffix": ""
	},
	{
	"first": "Sebastian",
	"middle": [],
	"last": "Schuster",
	"suffix": ""
	},
	{
	"first": "Francis",
	"middle": [],
	"last": "Tyers",
	"suffix": ""
	},
	{
	"first": "Daniel",
	"middle": [],
	"last": "Zeman",
	"suffix": ""
	}
	],
	"year": 2020,
	"venue": "Proceedings of The 12th Language Resources and Evaluation Conference",
	"volume": "",
	"issue": "",
	"pages": "4034--4043",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Joakim Nivre, Marie-Catherine de Marneffe, Filip Gin- ter, Jan Hajic, Christopher D. Manning, Sampo Pyysalo, Sebastian Schuster, Francis Tyers, and Daniel Zeman. 2020. Universal dependencies v2: An evergrowing multilingual treebank collection. In Proceedings of The 12th Language Resources and Evaluation Conference, pages 4034-4043, Mar- seille, France. European Language Resources Asso- ciation.",
	"links": null
	},
	"BIBREF4": {
	"ref_id": "b4",
	"title": "LIII. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin Philosophical Magazine",
	"authors": [
	{
	"first": "Karl",
	"middle": [],
	"last": "Pearson",
	"suffix": ""
	}
	],
	"year": 1901,
	"venue": "Journal of Science",
	"volume": "2",
	"issue": "11",
	"pages": "559--572",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Karl Pearson. 1901. LIII. on lines and planes of clos- est fit to systems of points in space. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 2(11):559-572.",
	"links": null
	},
	"BIBREF5": {
	"ref_id": "b5",
	"title": "Scikit-learn: Machine learning in Python",
	"authors": [
	{
	"first": "F",
	"middle": [],
	"last": "Pedregosa",
	"suffix": ""
	},
	{
	"first": "G",
	"middle": [],
	"last": "Varoquaux",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [],
	"last": "Gramfort",
	"suffix": ""
	},
	{
	"first": "V",
	"middle": [],
	"last": "Michel",
	"suffix": ""
	},
	{
	"first": "B",
	"middle": [],
	"last": "Thirion",
	"suffix": ""
	},
	{
	"first": "O",
	"middle": [],
	"last": "Grisel",
	"suffix": ""
	},
	{
	"first": "M",
	"middle": [],
	"last": "Blondel",
	"suffix": ""
	},
	{
	"first": "P",
	"middle": [],
	"last": "Prettenhofer",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [],
	"last": "Weiss",
	"suffix": ""
	},
	{
	"first": "V",
	"middle": [],
	"last": "Dubourg",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Vanderplas",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [],
	"last": "Passos",
	"suffix": ""
	},
	{
	"first": "D",
	"middle": [],
	"last": "Cournapeau",
	"suffix": ""
	},
	{
	"first": "M",
	"middle": [],
	"last": "Brucher",
	"suffix": ""
	},
	{
	"first": "M",
	"middle": [],
	"last": "Perrot",
	"suffix": ""
	},
	{
	"first": "E",
	"middle": [],
	"last": "Duchesnay",
	"suffix": ""
	}
	],
	"year": 2011,
	"venue": "Journal of Machine Learning Research",
	"volume": "12",
	"issue": "",
	"pages": "2825--2830",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duch- esnay. 2011. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825-2830.",
	"links": null
	},
	"BIBREF6": {
	"ref_id": "b6",
	"title": "Effective measures of domain similarity for parsing",
	"authors": [
	{
	"first": "Barbara",
	"middle": [],
	"last": "Plank",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Gertjan Van Noord",
	"suffix": ""
	}
	],
	"year": 2011,
	"venue": "Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies",
	"volume": "",
	"issue": "",
	"pages": "1566--1576",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Barbara Plank and Gertjan van Noord. 2011. Effec- tive measures of domain similarity for parsing. In Proceedings of the 49th Annual Meeting of the As- sociation for Computational Linguistics: Human Language Technologies, pages 1566-1576, Portland, Oregon, USA. Association for Computational Lin- guistics.",
	"links": null
	},
	"BIBREF7": {
	"ref_id": "b7",
	"title": "Master's thesis, Institute of Formal and Applied Linguisti",
	"authors": [
	{
	"first": "Vinit",
	"middle": [],
	"last": "Ravishankar",
	"suffix": ""
	}
	],
	"year": 2018,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Vinit Ravishankar. 2018. Parsing of texts with code- switching. Master's thesis, Institute of Formal and Applied Linguisti, Prague.",
	"links": null
	},
	"BIBREF8": {
	"ref_id": "b8",
	"title": "82 treebanks, 34 models: Universal dependency parsing with multi-treebank models",
	"authors": [
	{
	"first": "Aaron",
	"middle": [],
	"last": "Smith",
	"suffix": ""
	},
	{
	"first": "Bernd",
	"middle": [],
	"last": "Bohnet",
	"suffix": ""
	},
	{
	"first": "Joakim",
	"middle": [],
	"last": "Miryam De Lhoneux",
	"suffix": ""
	},
	{
	"first": "Yan",
	"middle": [],
	"last": "Nivre",
	"suffix": ""
	},
	{
	"first": "Sara",
	"middle": [],
	"last": "Shao",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Stymne",
	"suffix": ""
	}
	],
	"year": 2018,
	"venue": "Proceedings of the CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies",
	"volume": "",
	"issue": "",
	"pages": "113--123",
	"other_ids": {
	"DOI": [
	"10.18653/v1/K18-2011"
	]
	},
	"num": null,
	"urls": [],
	"raw_text": "Aaron Smith, Bernd Bohnet, Miryam de Lhoneux, Joakim Nivre, Yan Shao, and Sara Stymne. 2018. 82 treebanks, 34 models: Universal dependency pars- ing with multi-treebank models. In Proceedings of the CoNLL 2018 Shared Task: Multilingual Pars- ing from Raw Text to Universal Dependencies, pages 113-123, Brussels, Belgium. Association for Com- putational Linguistics.",
	"links": null
	},
	"BIBREF9": {
	"ref_id": "b9",
	"title": "Parser training with heterogeneous treebanks",
	"authors": [
	{
	"first": "Sara",
	"middle": [],
	"last": "Stymne",
	"suffix": ""
	},
	{
	"first": "Aaron",
	"middle": [],
	"last": "Miryam De Lhoneux",
	"suffix": ""
	},
	{
	"first": "Joakim",
	"middle": [],
	"last": "Smith",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Nivre",
	"suffix": ""
	}
	],
	"year": 2018,
	"venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics",
	"volume": "2",
	"issue": "",
	"pages": "619--625",
	"other_ids": {
	"DOI": [
	"10.18653/v1/P18-2098"
	]
	},
	"num": null,
	"urls": [],
	"raw_text": "Sara Stymne, Miryam de Lhoneux, Aaron Smith, and Joakim Nivre. 2018. Parser training with hetero- geneous treebanks. In Proceedings of the 56th An- nual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 619- 625, Melbourne, Australia. Association for Compu- tational Linguistics.",
	"links": null
	},
	"BIBREF10": {
	"ref_id": "b10",
	"title": "Multi-team: A multiattention, multi-decoder approach to morphological analysis",
	"authors": [
	{
	"first": "Rob",
	"middle": [],
	"last": "Ahmet\u00fcst\u00fcn",
	"suffix": ""
	},
	{
	"first": "Gosse",
	"middle": [],
	"last": "Van Der Goot",
	"suffix": ""
	},
	{
	"first": "Gertjan",
	"middle": [],
	"last": "Bouma",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Van Noord",
	"suffix": ""
	}
	],
	"year": 2019,
	"venue": "Proceedings of the 16th Workshop on Computational Research in Phonetics, Phonology, and Morphology",
	"volume": "",
	"issue": "",
	"pages": "35--49",
	"other_ids": {
	"DOI": [
	"10.18653/v1/W19-4206"
	]
	},
	"num": null,
	"urls": [],
	"raw_text": "Ahmet\u00dcst\u00fcn, Rob van der Goot, Gosse Bouma, and Gertjan van Noord. 2019. Multi-team: A multi- attention, multi-decoder approach to morphological analysis. In Proceedings of the 16th Workshop on Computational Research in Phonetics, Phonology, and Morphology, pages 35-49, Florence, Italy. As- sociation for Computational Linguistics.",
	"links": null
	},
	"BIBREF11": {
	"ref_id": "b11",
	"title": "Treebank embedding vectors for out-ofdomain dependency parsing",
	"authors": [
	{
	"first": "Joachim",
	"middle": [],
	"last": "Wagner",
	"suffix": ""
	},
	{
	"first": "James",
	"middle": [],
	"last": "Barry",
	"suffix": ""
	},
	{
	"first": "Jennifer",
	"middle": [],
	"last": "Foster",
	"suffix": ""
	}
	],
	"year": 2020,
	"venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
	"volume": "",
	"issue": "",
	"pages": "8812--8818",
	"other_ids": {
	"DOI": [
	"10.18653/v1/2020.acl-main.778"
	]
	},
	"num": null,
	"urls": [],
	"raw_text": "Joachim Wagner, James Barry, and Jennifer Foster. 2020. Treebank embedding vectors for out-of- domain dependency parsing. In Proceedings of the 58th Annual Meeting of the Association for Compu- tational Linguistics, pages 8812-8818, Online. As- sociation for Computational Linguistics.",
	"links": null
	},
	"BIBREF12": {
	"ref_id": "b12",
	"title": "Findings of the VarDial evaluation campaign 2017",
	"authors": [
	{
	"first": "Marcos",
	"middle": [],
	"last": "Zampieri",
	"suffix": ""
	},
	{
	"first": "Shervin",
	"middle": [],
	"last": "Malmasi",
	"suffix": ""
	},
	{
	"first": "Nikola",
	"middle": [],
	"last": "Ljube\u0161i\u0107",
	"suffix": ""
	},
	{
	"first": "Preslav",
	"middle": [],
	"last": "Nakov",
	"suffix": ""
	},
	{
	"first": "Ahmed",
	"middle": [],
	"last": "Ali",
	"suffix": ""
	},
	{
	"first": "J\u00f6rg",
	"middle": [],
	"last": "Tiedemann",
	"suffix": ""
	},
	{
	"first": "Yves",
	"middle": [],
	"last": "Scherrer",
	"suffix": ""
	},
	{
	"first": "No\u00ebmi",
	"middle": [],
	"last": "Aepli",
	"suffix": ""
	}
	],
	"year": 2017,
	"venue": "Proceedings of the Fourth Workshop on NLP for Similar Languages, Varieties and Dialects (VarDial)",
	"volume": "",
	"issue": "",
	"pages": "1--15",
	"other_ids": {
	"DOI": [
	"10.18653/v1/W17-1201"
	]
	},
	"num": null,
	"urls": [],
	"raw_text": "Marcos Zampieri, Shervin Malmasi, Nikola Ljube\u0161i\u0107, Preslav Nakov, Ahmed Ali, J\u00f6rg Tiedemann, Yves Scherrer, and No\u00ebmi Aepli. 2017. Findings of the VarDial evaluation campaign 2017. In Proceedings of the Fourth Workshop on NLP for Similar Lan- guages, Varieties and Dialects (VarDial), pages 1- 15, Valencia, Spain. Association for Computational Linguistics.",
	"links": null
	},
	"BIBREF13": {
	"ref_id": "b13",
	"title": "CoNLL 2018 shared task: Multilingual parsing from raw text to universal dependencies",
	"authors": [
	{
	"first": "Daniel",
	"middle": [],
	"last": "Zeman",
	"suffix": ""
	},
	{
	"first": "Jan",
	"middle": [],
	"last": "Haji\u010d",
	"suffix": ""
	},
	{
	"first": "Martin",
	"middle": [],
	"last": "Popel",
	"suffix": ""
	},
	{
	"first": "Martin",
	"middle": [],
	"last": "Potthast",
	"suffix": ""
	},
	{
	"first": "Milan",
	"middle": [],
	"last": "Straka",
	"suffix": ""
	},
	{
	"first": "Filip",
	"middle": [],
	"last": "Ginter",
	"suffix": ""
	},
	{
	"first": "Joakim",
	"middle": [],
	"last": "Nivre",
	"suffix": ""
	},
	{
	"first": "Slav",
	"middle": [],
	"last": "Petrov",
	"suffix": ""
	}
	],
	"year": 2018,
	"venue": "Proceedings of the CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies",
	"volume": "",
	"issue": "",
	"pages": "1--21",
	"other_ids": {
	"DOI": [
	"10.18653/v1/K18-2001"
	]
	},
	"num": null,
	"urls": [],
	"raw_text": "Daniel Zeman, Jan Haji\u010d, Martin Popel, Martin Pot- thast, Milan Straka, Filip Ginter, Joakim Nivre, and Slav Petrov. 2018. CoNLL 2018 shared task: Mul- tilingual parsing from raw text to universal depen- dencies. In Proceedings of the CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Univer- sal Dependencies, pages 1-21, Brussels, Belgium. Association for Computational Linguistics.",
	"links": null
	}
	},
	"ref_entries": {
	"FIGREF2": {
	"uris": null,
	"type_str": "figure",
	"text": "PCA-projection of the gold dataset embeddings of the dataset groups.",
	"num": null
	},
	"FIGREF3": {
	"uris": null,
	"type_str": "figure",
	"text": "PCA-projection of the predicted dataset embeddings of the dataset groups.",
	"num": null
	},
	"TABREF0": {
	"num": null,
	"text": "concat gold pred base concat gold pred #src base concat gold pred All 104 92.04 91.43 92.75 91.85 91.10 91.02 92.55 91.41 58 72.92 74.07 75.53 74.52 Single-lang 59 94.14 93.94 95.84 94.13 93.66 93.83 95.73 93.84 10 80.48 79.84 82.74 80.29 Multi-lang 45 89.30 88.14 88.69 88.88 87.75 87.33 88.38 88.22 48 71.35 72.87 74.03 73.32",
	"html": null,
	"content": "<table><tr><td/><td>Morphological Tagging (F1) Lemmatization (Accuracy)</td><td>Dependency Parsing (LAS)</td></tr><tr><td>Filtering</td><td>#src base</td></tr></table>",
	"type_str": "table"
	},
	"TABREF2": {
	"num": null,
	"text": "LAS scores for zero-shot experiments.",
	"html": null,
	"content": "<table/>",
	"type_str": "table"
	},
	"TABREF3": {
	"num": null,
	"text": "In-dataset training zero-shot dataset size svm base concat gold pred concat pred nl alp 186k 0.96 84.10 84.41 84.97 84.74 72.03 73.69 af afri 34k 1.00 79.57 78.95 79.98 80.44 35.86 34.84 nl lassy 75k 0.88 76.76 81.59 81.89 81.52 74.70 75.48 de gsd 268k 1.00 79.76 78.96 79.39 79.35 14.30 15.09 81.90 81.89 en ewt 205k 0.91 82.43 82.60 83.42 82.82 71.11 71.23 en lines 50k 0.77 76.15 75.06 79.20 77.14 74.71 74.68 en gum 54k 0.71 78.18 80.32 82.77 80.58 80.12 79.98",
	"html": null,
	"content": "<table><tr><td>en pud</td><td>0 0.00</td><td>-</td><td>-</td><td>-</td><td>-</td></tr></table>",
	"type_str": "table"
	},
	"TABREF4": {
	"num": null,
	"text": "Full results for the AF-DE-NL and EN dataset group (LAS). SIZE refers to training size in number of words. SVM accuracy of SVM language predictions.",
	"html": null,
	"content": "<table/>",
	"type_str": "table"
	},
	"TABREF5": {
	"num": null,
	"text": "Markus Dickinson, and Sandra K\u00fcbler. 2013. Towards domain adaptation for parsing web data. In Proceedings of the International Conference Recent Advances in Natural Language Processing RANLP 2013, pages 357-364, Hissar, Bulgaria. INCOMA Ltd. Shoumen, BULGARIA.",
	"html": null,
	"content": "<table><tr><td/><td>and other neural network architectures. Neural net-</td></tr><tr><td/><td>works, 18(5-6):602-610.</td></tr><tr><td/><td>Mohammad Khan, Christo Kirov, Ryan Cotterell, John Sylak-Glassman,</td></tr><tr><td/><td>G\u00e9raldine Walther, Ekaterina Vylomova, Patrick</td></tr><tr><td/><td>Xia, Manaal Faruqui, Sabrina J. Mielke, Arya Mc-</td></tr><tr><td/><td>Carthy, Sandra K\u00fcbler, David Yarowsky, Jason Eis-</td></tr><tr><td/><td>ner, and Mans Hulden. 2018. UniMorph 2.0: Uni-</td></tr><tr><td/><td>versal Morphology. In Proceedings of the Eleventh</td></tr><tr><td/><td>International Conference on Language Resources</td></tr><tr><td/><td>and Evaluation (LREC 2018), Miyazaki, Japan. Eu-</td></tr><tr><td/><td>ropean Language Resources Association (ELRA).</td></tr><tr><td>James Barry, Joachim Wagner, and Jennifer Foster.</td><td/></tr><tr><td>2019. Cross-lingual parsing with polyglot training</td><td/></tr><tr><td>and multi-treebank learning: A Faroese case study.</td><td/></tr><tr><td>In Proceedings of the 2nd Workshop on Deep Learn-</td><td/></tr><tr><td>ing Approaches for Low-Resource NLP (DeepLo</td><td/></tr><tr><td>2019), pages 163-174, Hong Kong, China. Associ-</td><td/></tr><tr><td>ation for Computational Linguistics.</td><td/></tr><tr><td>Angelo Basile, Gareth Dwyer, Maria Medvedeva, Jo-</td><td/></tr><tr><td>sine Rawee, Hessel Haagsma, and Malvina Nissim.</td><td/></tr><tr><td>2018. Simply the best: Minimalist system trumps</td><td/></tr><tr><td>complex models in author profiling. In Experimen-</td><td/></tr><tr><td>tal IR Meets Multilinguality, Multimodality, and In-</td><td/></tr><tr><td>teraction, pages 143-156, Cham. Springer Interna-</td><td/></tr><tr><td>tional Publishing.</td><td/></tr><tr><td>Irshad Bhat, Riyaz A. Bhat, Manish Shrivastava, and</td><td/></tr><tr><td>Dipti Sharma. 2017. Joining hands: Exploiting</td><td/></tr><tr><td>monolingual treebanks for parsing of code-mixing</td><td/></tr><tr><td>data. In Proceedings of the 15th Conference of the</td><td/></tr><tr><td>European Chapter of the Association for Computa-</td><td/></tr><tr><td>tional Linguistics: Volume 2, Short Papers, pages</td><td/></tr><tr><td>324-330, Valencia, Spain. Association for Computa-</td><td/></tr><tr><td>tional Linguistics.</td><td/></tr><tr><td>\u00c7 agr\u0131 \u00c7\u00f6ltekin and Taraka Rama. 2018. T\u00fcbingen-</td><td/></tr><tr><td>Oslo at SemEval-2018 task 2: SVMs perform better</td><td/></tr><tr><td>than RNNs in emoji prediction. In Proceedings of</td><td>, and</td></tr><tr><td>The 12th International Workshop on Semantic Eval-</td><td>Mans Hulden. 2019. The SIGMORPHON 2019</td></tr><tr><td>uation, pages 34-38, New Orleans, Louisiana. Asso-</td><td>shared task: Morphological analysis in context and</td></tr><tr><td>ciation for Computational Linguistics.</td><td>cross-lingual transfer for inflection. In Proceedings</td></tr><tr><td/><td>of the 16th Workshop on Computational Research in</td></tr><tr><td>Alexis Conneau and Guillaume Lample. 2019. Cross-</td><td>Phonetics, Phonology, and Morphology, pages 229-</td></tr><tr><td>lingual language model pretraining. In Advances</td><td>244, Florence, Italy. Association for Computational</td></tr><tr><td>in Neural Information Processing Systems, pages</td><td>Linguistics.</td></tr><tr><td>7059-7069.</td><td/></tr><tr><td>Yoav Goldberg and Joakim Nivre. 2013. Training de-</td><td/></tr><tr><td>terministic parsers with non-deterministic oracles.</td><td/></tr><tr><td>Transactions of the Association for Computational</td><td/></tr><tr><td>Linguistics, 1:403-414.</td><td/></tr><tr><td>Alex Graves and J\u00fcrgen Schmidhuber. 2005. Frame-</td><td/></tr><tr><td>wise phoneme classification with bidirectional lstm</td><td/></tr></table>",
	"type_str": "table"
	},
	"TABREF6": {
	"num": null,
	"text": ".95 92.04 91.43 92.75 91.85 91.10 91.02 92.55 91.41 58 161 0.91 72.92 74.07 75.53 74.52 Large 65 186 0.28 0.94 95.40 94.78 95.72 94.98 95.92 95.08 96.00 95.19 46 200 0.92 80.80 79.62 80.98 80.06 Small 39 13 0.35 0.98 86.46 85.85 87.78 86.65 83.07 84.24 86.79 85.11 12 10 0.88 42.75 52.79 54.65 53.30 Multi-lang 45 77 0.13 0.99 89.30 88.14 88.69 88.88 87.75 87.33 88.38 88.22 48 163 0.92 71.35 72.87 74.03 73.32 Single-lang 59 154 0.44 0.92 94.14 93.94 95.84 94.13 93.66 93.83 95.73 93.84 10 149 0.86 80.48 79.84 82.74 80.29 \u2203 same-lang 59 154 0.44 0.92 94.14 93.94 95.84 94.13 93.66 93.83 95.73 93.84 35 187 0.86 77.32 77.33 79.29 77.76",
	"html": null,
	"content": "<table><tr><td>shows the results of the Multi-Team tag-</td></tr><tr><td>ger (\u00dcst\u00fcn et al., 2019) on the development data for</td></tr><tr><td>each dataset. Because of the computational costs,</td></tr><tr><td>results are over a single run. The second column</td></tr><tr><td>shows the 'help-dataset' that each dataset is paired</td></tr><tr><td>with, based on word overlap.</td></tr><tr><td>Note that data sizes are different compared to</td></tr><tr><td>Table 5 due to a re-split of the data by (McCarthy</td></tr><tr><td>et al., 2019), and different UD versions (\u00dcst\u00fcn et al.</td></tr><tr><td>(2019) used 2.3 whereas Smith et al. (2018) used</td></tr><tr><td>2.2). Another effect of this re-split is that for all</td></tr><tr><td>datasets, a train, development and test split is avail-</td></tr><tr><td>able. Also note that dataset prediction (SVM) scores</td></tr><tr><td>reported are on the train data; so if the score is 1.00,</td></tr><tr><td>PRED and GOLD can still have different scores be-</td></tr></table>",
	"type_str": "table"
	},
	"TABREF7": {
	"num": null,
	"text": "Results per tasks, with averages over different dataset filters. #SETS: number of datasets, SIZE: training data size of dataset (in 1,000 words), SVM: F1 score of dataset classifier, BASE: single dataset baseline, CONCAT: concatenation of datasets in cluster, GOLD: gold dataset ids, PRED: predicted dataset ids.",
	"html": null,
	"content": "<table><tr><td>cause the dataset prediction was not equally accu-</td></tr><tr><td>rate on the development data.</td></tr><tr><td>E PCA-analysis of gold versus predicted</td></tr><tr><td>treebank embeddings</td></tr><tr><td>To gain deeper insights in what is represented in</td></tr><tr><td>treebank embeddings, we plotted the eight largest</td></tr><tr><td>dataset groups of the UUParser setup into a PCA</td></tr><tr><td>space</td></tr></table>",
	"type_str": "table"
	},
	"TABREF8": {
	"num": null,
	"text": ".96 84.10 84.41 84.97 84.74 72.03 73.69 af afribooms 33,894 1.00 79.57 78.95 79.98 80.44 35.86 34.84 nl lassysmall 75,134 0.88 76.76 81.59 81.89 81.52 74.70 75.48 de gsd 268,414 1.00 79.76 78.96 79.39 79.35 14.30 15.09 e-sla uk iu 75,098 0.99 79.86 79.49 80.63 80.53 35.69 35.99 ru taiga 10,479 0.35 56.47 69.32 71.73 69.30 68.81 68.81 ru syntagrus 871,521 0.99 87.26 87.07 87.13 87.24 59.52 59.34",
	"html": null,
	"content": "<table><tr><td/><td/><td/><td/><td colspan=\"2\">In-dataset training</td><td/><td>Zero-shot</td></tr><tr><td>cluster</td><td>dataset</td><td>size svm</td><td colspan=\"2\">base concat</td><td>gold</td><td colspan=\"2\">pred concat</td><td>pred</td></tr><tr><td colspan=\"3\">af-de-nl nl alpino 186,046 0en en pud 0 0.00</td><td>-</td><td>-</td><td>-</td><td colspan=\"2\">-81.90 81.89</td></tr><tr><td/><td>en ewt</td><td colspan=\"6\">204,607 0.91 82.43 82.60 83.42 82.82 71.11 71.23</td></tr><tr><td/><td>en lines</td><td colspan=\"6\">50,096 0.77 76.15 75.06 79.20 77.14 74.71 74.68</td></tr><tr><td/><td>en gum</td><td colspan=\"6\">53,686 0.71 78.18 80.32 82.77 80.58 80.12 79.98</td></tr><tr><td>es-ca</td><td>ca ancora</td><td colspan=\"5\">418,494 1.00 87.97 88.41 88.55 88.51</td><td>-</td><td>-</td></tr><tr><td/><td>es ancora</td><td colspan=\"5\">446,145 1.00 87.44 87.74 87.97 88.07</td><td>-</td><td>-</td></tr><tr><td>finno</td><td>et edt</td><td colspan=\"6\">287,859 1.00 79.48 77.47 77.79 77.66 14.38 14.32</td></tr><tr><td/><td>fi tdt</td><td colspan=\"6\">162,827 0.77 79.48 71.43 78.27 70.45 47.21 47.44</td></tr><tr><td/><td>fi ftb</td><td colspan=\"6\">127,845 0.69 79.58 70.48 79.05 70.88 51.46 52.16</td></tr><tr><td/><td>sme giella</td><td colspan=\"5\">16,835 1.00 63.17 53.08 56.23 55.32</td><td>6.58</td><td>6.45</td></tr><tr><td/><td>fi pud</td><td>0 0.00</td><td>-</td><td>-</td><td>-</td><td colspan=\"2\">-74.58 74.94</td></tr><tr><td>fr</td><td>fr spoken</td><td colspan=\"6\">14,952 0.93 71.39 76.15 77.48 76.72 53.50 53.35</td></tr><tr><td/><td>fr gsd</td><td colspan=\"6\">366,372 0.96 88.23 88.09 88.43 88.30 75.65 76.36</td></tr><tr><td/><td>fr sequoia</td><td colspan=\"6\">51,906 0.69 85.72 85.35 88.75 87.13 80.62 80.78</td></tr><tr><td>indic</td><td>ur udtb</td><td colspan=\"5\">108,690 1.00 78.15 78.43 78.58 78.58</td><td>-</td><td>-</td></tr><tr><td/><td>hi hdtb</td><td colspan=\"5\">281,057 1.00 89.20 89.27 89.38 89.38</td><td>-</td><td>-</td></tr><tr><td colspan=\"2\">iranian fa seraji</td><td colspan=\"5\">122,180 1.00 82.45 82.26 82.41 82.48</td><td>-</td><td>-</td></tr><tr><td/><td>kmr mg</td><td colspan=\"5\">242 0.99 12.24 34.76 34.96 35.38</td><td>-</td><td>-</td></tr><tr><td>it</td><td>it isdt</td><td colspan=\"5\">294,397 0.99 87.71 87.58 87.78 87.89</td><td>-</td><td>-</td></tr><tr><td/><td>it postwita</td><td colspan=\"5\">103,553 0.98 75.17 77.72 78.15 77.83</td><td>-</td><td>-</td></tr><tr><td>ko</td><td>ko gsd</td><td colspan=\"5\">56,687 0.68 76.70 65.62 78.42 63.56</td><td>-</td><td>-</td></tr><tr><td/><td>ko kaist</td><td colspan=\"5\">296,446 0.95 83.08 79.90 83.03 80.93</td><td>-</td><td>-</td></tr><tr><td>n-ger</td><td>no nynorsklia</td><td colspan=\"6\">3,583 0.90 50.05 62.27 62.87 62.91 52.89 53.27</td></tr><tr><td/><td>fo oft</td><td>0 0.00</td><td>-</td><td>-</td><td>-</td><td colspan=\"2\">-39.57 40.87</td></tr><tr><td/><td>sv talbanken</td><td colspan=\"6\">66,673 0.94 77.37 76.35 78.37 77.57 70.49 71.97</td></tr><tr><td/><td>no bokmaal</td><td colspan=\"6\">243,887 0.97 87.21 87.67 87.97 87.76 76.79 76.12</td></tr><tr><td/><td>sv pud</td><td>0 0.00</td><td>-</td><td>-</td><td>-</td><td colspan=\"2\">-77.89 77.65</td></tr><tr><td/><td>sv lines</td><td colspan=\"6\">48,325 0.91 76.48 77.71 78.95 78.37 71.66 72.18</td></tr><tr><td/><td>no nynorsk</td><td colspan=\"6\">245,330 0.98 85.67 85.49 86.27 85.93 73.50 74.36</td></tr><tr><td/><td>da ddt</td><td colspan=\"6\">80,378 0.97 76.97 73.79 76.84 76.06 52.04 52.27</td></tr><tr><td>old</td><td>cu proiel</td><td colspan=\"5\">37,432 1.00 76.62 73.91 73.12 74.22</td><td>5.36</td><td>4.95</td></tr><tr><td/><td>got proiel</td><td colspan=\"5\">35,024 1.00 71.46 69.02 68.29 69.06</td><td>8.53</td><td>8.25</td></tr><tr><td/><td>grc proiel</td><td colspan=\"6\">187,049 1.00 76.09 74.32 74.03 74.35 53.90 53.51</td></tr><tr><td/><td>la perseus</td><td colspan=\"6\">18,184 0.88 42.55 50.45 53.51 52.61 42.61 41.12</td></tr><tr><td/><td>la proiel</td><td colspan=\"6\">171,928 0.99 71.18 67.03 66.22 66.94 42.68 43.68</td></tr><tr><td/><td>grc perseus</td><td colspan=\"6\">159,895 1.00 61.78 61.46 61.46 61.36 46.42 47.87</td></tr><tr><td/><td>la ittb</td><td colspan=\"6\">270,403 1.00 79.40 74.29 74.14 74.69 41.66 40.61</td></tr><tr><td>pt-gl</td><td>gl ctg</td><td colspan=\"6\">86,676 0.95 80.21 79.85 81.11 80.56 63.78 64.19</td></tr><tr><td/><td>pt bosque</td><td colspan=\"6\">222,069 1.00 87.68 87.11 87.54 87.59 49.85 50.26</td></tr><tr><td/><td>gl treegal</td><td colspan=\"6\">16,707 0.62 69.15 67.24 75.76 69.03 61.62 61.71</td></tr><tr><td>sw-sla</td><td>sl sst</td><td colspan=\"6\">19,473 0.96 58.65 65.65 66.42 66.31 52.06 51.81</td></tr><tr><td/><td>sr set</td><td colspan=\"6\">65,764 0.86 83.91 84.07 86.42 85.91 75.63 75.66</td></tr><tr><td/><td>hr set</td><td colspan=\"6\">154,055 0.94 80.66 80.22 81.23 81.07 65.74 66.08</td></tr><tr><td/><td>sl ssj</td><td colspan=\"6\">112,530 0.99 85.27 84.89 85.46 85.23 65.28 65.83</td></tr><tr><td>turkic</td><td>ug udt</td><td colspan=\"5\">19,262 1.00 61.43 60.86 61.45 60.88</td><td>1.88</td><td>2.80</td></tr><tr><td/><td>bxr bdt</td><td>153 0.98</td><td colspan=\"4\">9.95 17.99 17.92 17.04</td><td>4.85</td><td>5.37</td></tr><tr><td/><td>tr imst</td><td colspan=\"5\">39,169 1.00 57.01 55.51 56.29 56.63</td><td>9.51</td><td>9.38</td></tr><tr><td/><td>kk ktb</td><td colspan=\"5\">547 1.00 11.54 30.52 31.16 29.14</td><td>7.03</td><td>6.19</td></tr><tr><td>w-sla</td><td>sk snk</td><td colspan=\"6\">80,575 0.98 80.39 82.49 83.07 82.51 59.77 59.57</td></tr><tr><td/><td>cs pud</td><td>0 0.00</td><td>-</td><td>-</td><td>-</td><td colspan=\"2\">-83.79 83.77</td></tr><tr><td/><td>cs pdt</td><td colspan=\"6\">1,175,374 0.92 87.92 87.41 87.40 87.39 79.07 79.33</td></tr><tr><td/><td>pl sz</td><td colspan=\"6\">63,070 0.46 85.31 80.88 82.31 81.49 67.40 67.88</td></tr><tr><td/><td>hsb ufal</td><td>460 0.90</td><td colspan=\"5\">6.40 45.24 46.30 44.99 39.10 38.55</td></tr><tr><td/><td>pl lfg</td><td colspan=\"6\">104,750 0.71 90.98 86.51 87.96 87.53 70.76 70.49</td></tr><tr><td/><td>cs fictree</td><td colspan=\"6\">134,059 0.78 85.77 86.92 87.14 87.08 83.75 82.91</td></tr><tr><td/><td>cs cac</td><td colspan=\"6\">473,622 0.83 86.86 87.40 87.32 87.40 83.73 83.57</td></tr></table>",
	"type_str": "table"
	},
	"TABREF9": {
	"num": null,
	"text": "96.50 96.42 96.96 96.87 95.25 94.97 96.01 96.93 94.59 96.10 96.31 95.96 98.06 98.81 98.77 98.77 ja modern ja gsd 12,213 0.29 99.78 95.93 95.69 95.65 95.86 93.67 95.96 95.55 96.59 ja pud ja gsd 22,450 0.64 92.59 95.85 97.91 97.66 97.29 96.08 99.30 99.34 99.01 kmr mg es gsd 8,680 0.03 100.00 86.43 86.93 88.64 87.61 88.38 90.72 91.19 90.44 89.49 92.91 93.47 93.75 91.76 95.64 95.93 96.38 sme giella no nynorsk 23,877 0.02 99.89 91.20 90.44 91.94 91.85 87.31 86.19 88.69 88.69 sr set hr set 72,045 0.62 88.98 95.52 95.24 97.47 96.09 95.77 95.96 97.02 95.96 sv lines sv talbanken 67,016 0.31 90.90 94.33 94.98 95.30 95.14 95.29 94.73 95.12 94.57 sv pud sv talbanken 15,758 0.39 93.08 89.65 93.78 95.12 94.20 83.86 88.30 93.11 90.35 sv talbanken sv lines 82,088 0.25 90.90 96.28 96.62 96.97 96.81 96.67 95.93 95.34 95.41 tl trg es gsd 274 0.13 99.98 74.73 82.22 76.92 83.15 60.00 72.00 80.00 76.00 tr imst tr pud 50,925 0.13 93.10 92.59 90.83 93.79 90.79 92.94 92.57 94.22 92.20 tr pud tr imst 14,180 0.33 93.10 91.82 86.88 93.83 87.66 84.80 84.92 86.32 84.34 uk iu ru syntagrus 98,865 0.10 99.55 92.69 91.14 92.41 92.13 93.67 91.59 93.07 92.86 ur udtb fa seraji 114,786 0.16 100.00 91.69 91.30 91.64 91.40 96.20 95.51 95.68 95.93 vi vtb en ewt 37,637 0.02 99.87 89.82 88.84 89.02 89.39 99.18 99.83 99.83 99.90 yo ytb es gsd 2,238 0.06 99.99 87.46 89.49 85.04 91.00 94.00 95.20 93.60 96.00 yue hk zh gsd 5,641 0.42 99.93 86.32 87.42 89.77 88.27 92.97 98.97 98.97 98.97 zh cfl zh gsd 6,048 0.34 99.93 86.15 88.92 88.13 89.72 91.00 95.57 96.26 95.98 zh gsd ja gsd 102,731 0.15 99.99 89.62 91.40 90.87 91.24 98.46 99.05 98.97 99.09 Average 121,049 0.31 95.42 92.04 91.43 92.75 91.85 91.10 91.02 92.55 91.41",
	"html": null,
	"content": "<table><tr><td colspan=\"4\">Morphological Tagging (F1) base concat gold pred 92.29 93.53 86.94 94.97 87.35 89.88 89.30 91.47 85.59 Lemmatization (Accuracy) svm base concat gold pred 99.85 akk pisandub cs pdt dataset additional size WO af afribooms nl alpino 40,390 0.20 ja gsd ja pud 154,453 0.14 92.59 ko gsd ko kaist 69,382 0.33 1,505 0.02 100.00 81.96 67.99 69.49 71.96 41.33 34.67 34.22 35.11 ar padt ar pud 231,625 0.18 ko kaist ko gsd 302,384 0.12 92.29 95.86 95.48 95.97 95.39 94.30 94.20 93.88 92.37 96.64 95.29 95.46 96.17 95.44 90.79 91.52 95.08 91.80 ar pud ar padt 17,645 0.56 ko pud ko kaist 14,106 0.55 97.97 93.41 82.62 95.62 83.29 92.36 75.41 98.08 75.41 96.64 89.97 87.49 92.24 87.20 77.36 62.83 84.39 57.39 be hse ru syntagrus 6,855 0.08 kpv ikdp ru syntagrus 916 0.26 99.95 61.38 48.34 63.10 55.79 56.63 55.42 61.45 61.45 99.98 80.51 80.61 82.01 83.49 78.48 75.67 79.75 81.58 bg btb ru syntagrus 133,659 0.12 kpv lattice ru syntagrus 1,805 0.09 99.96 75.26 66.32 67.50 67.36 57.69 58.24 64.29 61.54 99.60 97.85 96.67 96.89 96.97 96.95 94.54 94.70 94.95 bm crb cs pdt 12,025 0.09 la ittb la proiel 298,460 0.37 99.98 97.08 97.23 97.03 97.36 98.54 97.98 98.65 98.54 99.98 94.03 89.49 90.64 89.92 87.86 78.40 80.40 80.09 br keb no bokmaal 8,772 0.07 la perseus la proiel 25,157 0.49 99.89 82.04 87.64 88.67 87.95 80.99 87.44 88.76 87.37 99.87 90.89 90.53 88.45 89.67 88.75 89.77 88.86 88.35 bxr bdt ru syntagrus 8,770 0.04 la proiel la ittb 174,977 0.20 99.98 95.30 95.15 94.83 95.10 96.65 94.54 96.03 95.47 99.94 83.46 80.65 78.94 79.96 82.71 80.65 80.56 81.63 ca ancora es ancora 441,014 0.18 lt hse lv lvtb 4,511 0.05 99.83 72.44 77.53 78.12 77.94 72.96 77.25 78.11 77.68 99.71 98.61 98.48 98.57 98.56 98.35 98.42 98.56 98.69 cs cac cs pdt 414,810 0.57 lv lvtb hr set 129,982 0.02 99.81 95.13 94.58 94.34 94.54 93.78 93.41 92.60 93.85 90.62 97.19 96.71 96.98 96.72 98.03 97.05 97.27 97.24 cs cltt cs pdt 29,549 0.81 mr ufal hi hdtb 3,427 0.16 100.00 76.63 77.98 76.28 78.61 70.12 72.47 74.12 72.00 99.87 94.62 96.65 97.16 96.62 94.02 97.56 97.35 97.56 cs fictree cs pdt 143,508 0.58 nl alpino nl lassysmall 178,169 0.23 93.27 95.88 95.49 96.18 95.98 95.60 94.57 95.61 95.43 95.01 95.89 94.89 96.54 95.36 95.21 97.07 97.65 97.15 cs pdt cs cac 1,278,252 0.27 nl lassysmall nl alpino 84,612 0.41 93.27 93.67 95.15 96.04 95.60 93.34 94.24 95.61 94.99 90.62 96.65 96.78 97.05 96.76 97.50 97.10 97.12 97.19 cs pud cs pdt 15,614 0.79 no bokmaal no nynorsk 264,958 0.24 96.34 97.24 97.41 97.25 97.65 98.23 97.86 98.31 98.01 98.98 87.38 96.26 94.88 96.38 87.03 96.84 96.03 97.00 cu proiel ru syntagrus 50,963 0.04 no nynorsk no bokmaal 255,088 0.25 96.34 96.68 97.10 97.03 97.31 96.59 97.50 98.02 97.66 99.98 94.71 92.94 91.62 92.12 95.17 92.66 91.09 91.28 da ddt no bokmaal 85,373 0.25 no nynorsklia no nynorsk 11,959 0.65 99.18 92.16 95.44 95.77 95.11 92.93 98.09 97.64 97.72 97.38 92.58 95.21 95.52 95.26 92.27 95.60 96.25 95.30 de gsd fr gsd 246,633 0.06 pcm nsc en ewt 11,038 0.74 99.99 92.95 93.28 94.26 94.15 98.55 99.92 99.84 99.76 99.94 93.48 92.18 93.16 93.05 96.55 94.95 95.94 95.60 el gdt grc proiel 52,583 0.04 pl lfg pl sz 118,526 0.41 60.73 95.71 94.43 96.43 93.91 95.93 94.67 95.49 94.96 99.96 96.51 96.25 96.36 96.45 95.00 94.52 93.93 95.15 en ewt en gum 218,154 0.30 pl sz pl lfg 73,011 0.53 60.73 92.68 88.62 94.82 89.33 95.78 93.96 95.79 94.81 89.27 95.73 95.75 95.49 95.32 97.02 96.79 96.81 96.46 en gum en ewt 67,381 0.57 pt bosque pt gsd 188,265 0.48 87.08 96.38 88.48 96.84 92.15 97.43 86.31 97.84 90.18 89.27 94.45 94.39 95.11 94.29 96.94 93.94 96.37 94.63 en lines en ewt 70,079 0.56 pt gsd pt bosque 265,352 0.41 87.08 97.63 94.92 98.03 93.49 97.53 94.37 98.36 94.38 91.20 95.14 93.33 95.88 94.71 97.39 94.98 97.34 96.59 en partut en ewt 40,974 0.63 ro nonstandard ro rrt 164,375 0.24 97.63 95.19 95.83 95.55 96.14 94.49 96.06 95.71 96.02 95.66 93.45 90.75 94.03 91.83 97.58 96.53 97.27 95.83 en pud en ewt 17,727 0.68 ro rrt ro nonstandard 182,366 0.11 97.63 97.25 97.16 97.19 97.30 97.09 97.11 96.47 97.36 95.14 90.87 95.08 95.40 94.91 93.55 95.62 96.17 94.68 es ancora es gsd 454,069 0.47 ru gsd ru syntagrus 84,013 0.55 96.08 93.72 91.02 94.67 91.26 95.69 91.96 96.90 92.36 87.10 98.34 97.46 98.48 97.97 98.60 97.15 98.59 97.80 es gsd es ancora 358,355 0.42 ru pud ru syntagrus 16,233 0.74 98.69 88.38 87.34 93.93 87.87 86.92 94.09 93.46 93.51 87.10 97.35 96.15 97.55 96.97 98.59 95.81 98.43 97.65 et edt cs pdt 371,564 0.02 ru syntagrus ru gsd 937,395 0.13 96.08 96.69 96.84 97.36 96.40 96.73 96.42 97.50 95.68 99.75 96.60 94.65 94.86 94.68 94.73 89.62 88.80 89.09 eu bdt es ancora 104,530 0.05 ru taiga ru syntagrus 18,173 0.66 98.05 82.67 92.42 92.38 92.23 83.24 92.07 92.47 92.98 99.96 95.10 93.62 94.41 94.38 96.32 95.35 95.47 95.52 fa seraji ur udtb sa ufal hi hdtb 1,634 0.10 99.98 69.59 68.81 69.60 70.68 52.58 62.89 64.95 66.49 127,371 0.12 100.00 97.15 97.24 97.26 97.31 95.00 94.24 95.29 94.97 fi ftb fi tdt 142,514 0.37 sk snk cs pdt 93,740 0.22 99.22 94.69 92.39 93.30 93.51 95.13 91.31 93.31 92.61 71.46 95.25 94.94 96.18 94.94 92.15 90.99 92.75 90.85 fi pud fi tdt 13,356 0.49 sl ssj hr set 118,536 0.12 99.33 95.38 94.98 95.09 95.65 96.13 95.58 95.97 96.32 94.58 91.56 96.97 97.41 96.81 78.62 86.06 88.98 86.78 fi tdt fi ftb 173,899 0.30 71.46 96.54 95.32 97.07 95.18 92.42 90.27 92.22 90.03 sl sst sl ssj 26,309 0.55 99.20</td></tr><tr><td>fo oft</td><td>no nynorsk</td><td>8,960 0.12</td><td>99.83 90.36 86.49 91.46 90.38 83.87 81.59 88.52 86.66</td></tr><tr><td>fr gsd</td><td>fr sequoia</td><td>333,477 0.15</td><td>93.23 97.71 97.65 98.07 97.40 97.74 96.52 97.50 96.14</td></tr><tr><td>fr partut</td><td>fr gsd</td><td>23,443 0.82</td><td>97.28 94.72 96.33 97.51 96.13 94.20 95.13 96.78 95.10</td></tr><tr><td>fr sequoia</td><td>fr gsd</td><td>58,963 0.66</td><td>93.23 96.74 96.79 98.07 96.21 96.99 96.90 98.17 95.81</td></tr><tr><td>fr spoken</td><td>fr gsd</td><td>30,410 0.77</td><td>99.33 95.81 97.10 97.64 97.44 96.77 98.79 98.98 99.07</td></tr><tr><td>ga idt</td><td>cs pdt</td><td>19,812 0.04</td><td>99.95 83.38 76.03 75.27 77.94 84.63 76.48 76.91 79.01</td></tr><tr><td>gl ctg</td><td>es ancora</td><td>114,228 0.40</td><td>99.88 97.33 97.22 97.38 97.36 98.12 98.14 98.16 98.37</td></tr><tr><td>gl treegal</td><td>gl ctg</td><td>21,366 0.53</td><td>92.26 91.56 82.76 94.64 84.77 92.69 94.69 96.66 95.57</td></tr><tr><td>got proiel</td><td>no nynorsk</td><td>48,980 0.01</td><td>99.91 95.20 94.61 93.97 93.94 95.35 94.60 94.58 94.23</td></tr><tr><td>grc perseus</td><td>grc proiel</td><td>173,299 0.25</td><td>99.95 94.86 94.74 95.05 95.05 93.24 92.67 93.15 93.19</td></tr><tr><td>grc proiel</td><td>grc perseus</td><td>185,142 0.31</td><td>99.95 96.92 96.98 97.08 97.11 95.85 95.90 96.46 96.30</td></tr><tr><td>he htb</td><td>ru gsd</td><td colspan=\"2\">134,397 0.00 100.00 96.26 96.17 96.07 96.23 96.62 96.52 96.44 96.71</td></tr><tr><td>hi hdtb</td><td>mr ufal</td><td colspan=\"2\">295,265 0.01 100.00 96.70 96.87 96.77 96.89 98.57 98.34 98.52 98.40</td></tr><tr><td>hr set</td><td>sr set</td><td>164,557 0.28</td><td>88.98 95.14 94.85 95.50 94.59 95.81 94.53 95.25 94.52</td></tr><tr><td>hsb ufal</td><td>cs pdt</td><td>9,475 0.08</td><td>99.95 79.92 77.49 79.09 76.84 82.68 74.70 78.39 76.76</td></tr><tr><td>hu szeged</td><td>et edt</td><td>34,903 0.03</td><td>99.95 92.27 91.05 91.15 89.75 90.09 88.37 87.94 84.66</td></tr><tr><td>hy armtdp</td><td>ru pud</td><td colspan=\"2\">19,419 0.00 100.00 90.34 91.31 90.81 91.13 90.00 92.15 91.87 92.01</td></tr><tr><td>id gsd</td><td>es gsd</td><td>101,687 0.11</td><td>99.92 92.72 92.87 93.30 93.39 98.77 98.87 98.97 98.81</td></tr><tr><td>it isdt</td><td>it partut</td><td>250,714 0.28</td><td>79.42 98.07 97.91 98.26 97.77 97.52 96.70 97.87 96.85</td></tr><tr><td>it partut</td><td>it isdt</td><td>46,228 0.94</td><td>79.42 96.11 98.47 98.73 98.30 96.07 97.78 98.68 97.21</td></tr><tr><td>it postwita</td><td>it isdt</td><td>104,437 0.46</td><td>98.49 95.76 96.13 96.15 96.53 94.15 96.14 94.89 95.20</td></tr><tr><td>it pud</td><td>it isdt</td><td>19,634 0.69</td><td>94.02 94.30 85.62 96.69 84.00 93.32 96.64 97.00 95.22</td></tr></table>",
	"type_str": "table"
	}
	}
	}
	}