TERMinator: A System for Scientific Texts Processing

Proceedings of the 29th International Conference on Computational Linguistics, pages 3420–3426 October 12–17, 2022. 3420 TERMinator: A system for scientific texts processing Elena Bruches A.P. Ershov Institute of Informatics Systems / Russia bruches@bk.ru Olga Tikhobaeva Novosibirsk State University / Russia o.tikhobaeva@g.nsu.ru Yana Dementyeva Novosibirsk State University / Russia y.dementeva@g.nsu.ru Tatiana Batura A.P. Ershov Institute of Informatics Systems / Russia tbatura@iis.nsk.su Abstract This paper is devoted to the extraction of en- tities and semantic relations between them from scientific texts, where we consider sci- entific terms as entities. In this paper, we present a dataset that includes annotations for two tasks and develop a system called TER- Minator for the study of the influence of lan- guage models on term recognition and com- parison of different approaches for relation extraction. Experiments show that language models pre-trained on the target language are not always show the best performance. Also adding some heuristic approaches may im- prove the overall quality of the particular task. The developed tool and the annotated corpus are publicly available at https://github.com/iis- research-team/terminator and may be useful for other researchers. 1 Introduction Nowadays the amount of scientific publications is constantly growing. In this regard, the processing of scientific texts becomes especially relevant in relation to rapidly developing scientific fields, for example, computer science. Information extrac- tion from scientific texts can be useful in domain- specific areas, for completion of knowledge graphs, in search and question-answering systems. This paper describes the study on entity recognition and relation extraction from scientific texts on computer science in Russian. Currently, there are a number of datasets with annotations of entities and relations in a general domain (Doddington et al., 2004; Roth and Yih, 2004; Loukachevitch et al., 2021), biomedical do- main (Kim et al., 2003; Gurulingappa et al., 2012; Li et al., 2016), or even multi-domains (Terryn et al., 2020). Still it is more difficult to find a pub- licly available dataset such as SciERC (Luan et al., 2018) for scientific fields other than biomedical, and especially in languages other than English. Despite that the named entity recognition task is well studied, it still faces multiple challenges (Li et al., 2022a), namely, NER in domain-specific areas (Weber et al., 2021), NER from noisy data (Derczynski et al., 2017), code-mixed data (Fe- tahu et al., 2021), and detection of fine-grained and nested named entities (Kim and Kim, 2021; Ring- land et al., 2019; Loukachevitch et al., 2021). This is caused by several issues: defining boundaries of compound terms; recognition of whether a lexical unit is part of a compound term; identification of a lexical unit as a term depending on the context and topic of the text in which this lexical unit is used etc. The relation extraction task also remains an unsolved problem, as it often requires the use of knowledge outside the text (for example, from knowledge bases or obtained in another way), and also due to the lack of a large amount of labeled data in different languages. Selection of the …

Content truncated to 500 words