Multilingual Coreference Resolution with Harmonized Annotations

Proceedings of Recent Advances in Natural Language Processing, pages 1119–1123 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_125 1119 Multilingual Coreference Resolution with Harmonized Annotations Ondˇrej Praˇz´ak and Miloslav Konop´ık and Jakub Sido {ondfa,sidoj}@ntis.zcu.cz konopik@kiv.zcu.cz Department of Computer Science and Engineering, NTIS – New Technologies for the Information Society, Faculty of Applied Sciences, University of West Bohemia, Technick´a 8, 306 14 Plzeˇn Czech Republic Abstract In this paper, we present coreference resolu- tion experiments with a newly created multi- lingual corpus CorefUD (Nedoluzhko et al., 2021). We focus on the following languages: Czech, Russian, Polish, German, Spanish, and Catalan. In addition to monolingual experi- ments, we combine the training data in multi- lingual experiments and train two joined mod- els – for Slavic languages and for all the lan- guages together. We rely on an end-to-end deep learning model that we slightly adapted for the CorefUD corpus. Our results show that we can profit from harmonized annotations, and using joined models helps significantly for the languages with smaller training data. 1 Introduction Coreference resolution is the task of finding lan- guage expressions that refer to the same real-world entity (antecedent) of a given text. Sometimes the corefering expressions can come from a single sen- tence. However, the expressions can be one or more sentences apart as well. It is necessary to see the whole document in some hard cases to judge whether two expressions are corefering adequately. This task can be divided into two subtasks. Identi- fying entity mentions, and grouping the mentions together according to the real-world entity they re- fer to. The task of coreference resolution is closely related to anaphora resolution – see (Sukthanker et al., 2020) to compare these two tasks. One of the challenging difficulties of corefer- ence resolution lay in linguistically complicated annotations. Some examples of linguistic compli- cations are split antecedents (a mention refer to more than one real-word entities), near identity relations, anaphoric and cataphoric relations, etc (Nedoluzhko et al., 2021). In this paper, we rely on a CorefUD corpus (Nedoluzhko et al., 2021) of harmonized annota- tions. This corpus enables us to battle linguistic complications since it simply presents corefering mentions in clusters. Since the corpus is compiled from 11 different corpora in 8 different languages, we can conduct multilingual experiments in this work. Our research goal is to evaluate whether the harmonized annotations open the possibility to ob- tain some performance gain by joint learning on multiple languages. We aim to compare the har- monized annotations with the original corpora as well. 2 Related Work In agreement with many other NLP tasks, deep learning models prevail in the coreference resolu- tion task. Lee et al. (2017) were first to introduce the end-to-end approach that many following pa- pers adopted (they obtained an average of 67.2 of F1 score). The task experienced a big leap in per- formance with the introduction of large pre-trained models. BERT based models deliver the best re- sults; Kantor and Globerson (2019) F1 76.6, and Joshi et al. (2019) F1 …

Content truncated to 500 words