Dealing with Abbreviations in the Slovenian Biographical Lexicon

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 8715 - 8720 December 7-11, 2022 ©2022 Association for Computational Linguistics Dealing with Abbreviations in the Slovenian Biographical Lexicon Angel Daza1 Antske Fokkens1 Tomaž Erjavec2 1CLTL, Vrije Universiteit Amsterdam 2Department of Knowledge Technologies, Jožef Stefan Institute {j.a.dazaarevalo,antske.fokkens}@vu.nl, tomaz.erjavec@ijs.si Abstract Abbreviations present a significant challenge for NLP systems because they cause tokeniza- tion and out-of-vocabulary errors. They can also make the text less readable, especially in reference printed books, where they are exten- sively used. Abbreviations are especially prob- lematic in low-resource settings, where systems are less robust to begin with. In this paper, we propose a new method for addressing the problems caused by a high density of domain- specific abbreviations in a text. We apply this method to the case of a Slovenian biographical lexicon and evaluate it on a newly developed gold-standard dataset of 51 Slovenian biogra- phies. Our abbreviation identification method performs significantly better than commonly used ad-hoc solutions, especially at identify- ing unseen abbreviations. We also propose and present the results of a method for expanding the identified abbreviations in context. 1 Introduction Abbreviations such as "b." for "born", or "gr." for "graduated" are a common issue when dealing with digitized texts which use a large number of them for space-saving reasons. They are also a known prob- lem when processing technical documents (Park and Byrd, 2001) and biomedical texts (Jin et al., 2019). In this paper, we examine the case of bi- ographical dictionaries, i.e. collections of biogra- phies that have been digitized, and, in particular, the Slovenian Biographical Lexicon. To automatically extract facts from biographi- cal texts, Digital Humanities researchers normally rely on out-of-the-box NLP tools such as Stanza (Qi et al., 2020) or SpaCy.1 These tools are often adequate for identifying sentences which are then used as input for higher-level downstream tasks, for manual inspection, and for visualization purposes. However, out-of-the-box tools are designed to work 1https://spacy.io/ for the broadest possible text domains and cover the most common cases. This impacts performance sig- nificantly when dealing with domain-specific data, such as entries of biographical dictionaries, and more so when they contain a lot of abbreviations. The problem is even more pronounced when deal- ing with a relatively lower resource language, such as Slovenian. The performance bottleneck occurs already at the first step, i.e. tokenization: in order to perform good tokenization in domain-specific texts, we need to have a reliable method for iden- tifying abbreviations such that the tokenizer does not split them wrongly, generating faulty tokens and incomplete sentences. In this paper we: • Quantify the effect that abbreviations have on a downstream task such as NER on Slovenian biographical texts. • Propose a method for abbreviation identifica- tion, apply it to raw texts and compare it to straightforward baselines. • Analyze the feasibility of using contextually dependent word embeddings, in particular, the SloBERTa (Ulˇcar and Robnik-Šikonja, 2021) language model, to automatically expand ab- breviations in text and improve readability. …

Content truncated to 500 words