HerBERT Based Language Model Detects Quantifiers and Their Semantic Properties in Polish

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 7140–7146 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 7140 HerBERT Based Language Model Detects Quantifiers and Their Semantic Properties in Polish Marcin Woliński, Bartłomiej Nitoń, Witold Kieraś, Jakub Szymanik Institute of Computer Science Polish Academy of Sciences Institute for Logic, Language, and Computation, University of Amsterdam {marcin.wolinski, bartlomiej.niton, witold.kieras}@ipipan.waw.pl, j.k.szymanik@uva.nl Abstract The paper presents a tool for automatic marking up of quantifying expressions, their semantic features, and scopes. We explore the idea of using a BERT based neural model for the task (in this case HerBERT, a model trained specifically for Polish, is used). The tool is trained on a recent manually annotated Corpus of Polish Quantificational Expressions (Szymanik and Kieraś, 2022). We discuss how it performs against human annotation and present results of automatic annotation of 300 million sub-corpus of National Corpus of Polish. Our results show that language models can effectively recognise semantic category of quantification as well as identify key semantic properties of quantifiers, like monotonicity. Furthermore, the algorithm we have developed can be used for building semantically annotated quantifier corpora for other languages. Keywords: quantifiers, annotation, HerBERT 1. Introduction Quantifying expressions or quantifiers are understood in this text as language expressions which indicate quantity. These can be exact numbers of objects (‘three [books]’) or generalised quantifiers: ‘all’, ‘each’, ‘some’, ‘none’, ‘majority [of]’. Quantifiers can also count events or express their frequency: ‘never’, ‘al- ways’, ‘twice’, ‘often’, ‘repeatedly’, ‘each Tuesday’ (Szymanik and Kieraś, 2022). The task of detecting quantifiers in the text is super- ficially similar to named entity recognition. The set of possible quantifiers comprises numerals expressed with digits and a seemingly limited set of other phrases, which, with some effort, could be listed in a dictionary. However, unlike in the case of named entities, not all occurrences of these expressions are in fact quantifiers. For example the digit ‘3’ denotes a quantifier in the con- texts where it is read as ‘three’ but not when it means ‘third’. Similarly, nouns naming numbers, e.g. Pol- ish trójka, may express a similar concept as numerals (trójka chłopców ‘three boys’) or it may denote any ob- ject bearing the number 3 (a radio station, TV channel, military unit, squad). In this latter context the word is not a quantifier. For another example, consider the word most. It can denote a quantifier when followed by a plu- ral noun as in most men or a modifier when followed by an adjective as in most beautiful. Thus, the decision whether a given expression is a quantifier is to a large extent based on the semantic analysis of the linguistic context in which the expression appears. Beyond de- tecting quantifiers, we are also interested in two subor- dinate tasks: having a quantifier we want to characterise its semantic features (e.g., monotonicity) and to deter- mine its scope, that is to find in the text the expression specifying what is being quantified. Furthermore, …

Content truncated to 500 words