Grammatical Profiling for Semantic Change Detection

Proceedings of the 25th Conference on Computational Natural Language Learning (CoNLL), pages 423–434 November 10–11, 2021. ©2021 Association for Computational Linguistics 423 Grammatical Profiling for Semantic Change Detection Mario Giulianelli∗ ILLC, University of Amsterdam m.giulianelli@uva.nl Andrey Kutuzov∗ University of Oslo andreku@ifi.uio.no Lidia Pivovarova∗ University of Helsinki first.last@helsinki.fi Abstract Semantics, morphology and syntax are strongly interdependent. However, the major- ity of computational methods for semantic change detection use distributional word rep- resentations which encode mostly semantics. We investigate an alternative method, gram- matical profiling, based entirely on changes in the morphosyntactic behaviour of words. We demonstrate that it can be used for semantic change detection and even outperforms some distributional semantic methods. We present an in-depth qualitative and quantitative analysis of the predictions made by our grammatical profiling system, showing that they are plausible and interpretable. 1 Introduction Lexical semantic change detection has recently be- come a well-represented field in NLP, with several shared tasks conducted for English, German, Latin and Swedish (Schlechtweg et al., 2020), Italian (Basile et al., 2020) and Russian (Kutuzov and Pivovarova, 2021a). The overwhelming majority of solutions employ either static word embeddings like word2vec (Mikolov et al., 2013) or more re- cent contextualised language models like ELMo (Peters et al., 2018) and BERT (Devlin et al., 2019). These models build upon the distributional seman- tics hypothesis and can capture lexical meaning, at least to some extent (e.g., Iacobacci et al., 2016; Pilehvar and Camacho-Collados, 2019; Yenicelik et al., 2020). Thus, they are naturally equipped to model semantic change. Yet it has long been known for linguists that semantics, morphology and syntax are strongly interrelated (Langacker, 1987; Hock and Joseph, 2019). Semantic change is consequently often ac- companied by morphosyntactic drifts. Consider the English noun ‘lass’: in the 20th century, its ‘SWEETHEART’ meaning became more dominant ∗Equal contribution, the authors listed alphabetically. Figure 1: Changes in the number category distribution for the English noun ‘lass’ over time, calculated on the English corpora of the SemEval 2020 shared task 1 (Schlechtweg et al., 2020). ‘Lass’ is annotated as se- mantically changed in the SemEval dataset. over the older sense of ‘YOUNG WOMAN’. This was accompanied by a sharp decrease in plural us- ages (‘lasses’), as shown in Figure 1. Exploiting distributions of grammatical pro- files—i.e., morphological and syntactic features— to detect lexical semantic change is the focus of this paper. We investigate to what extent lexical semantic change can be detected using only mor- phosyntax. Our main hypothesis is that significant changes in the distribution of morphosyntactic cat- egories can reveal useful information on the degree of the word’s semantic change, even without help from any lexical or explicitly semantic features. Due to the interdependence of semantics and morphosyntax, it is often difficult to determine which type of change occurred first, and whether it triggered the other. Establishing the correct causal direction is outside the scope of this study; it is sufficient for us to know that semantic and mor- phosyntactic changes often co-occur. By proposing this functionalist …

Content truncated to 500 words