When Time Makes Sense: A Historically-Aware Approach to Targeted Sense Disambiguation

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2751–2761 August 1–6, 2021. ©2021 Association for Computational Linguistics 2751 When Time Makes Sense: A Historically-Aware Approach to Targeted Sense Disambiguation Kaspar Beelen∗ The Alan Turing Institute Queen Mary University of London kbeelen@turing.ac.uk Federico Nanni The Alan Turing Institute fnanni@turing.ac.uk Mariona Coll Ardanuy The Alan Turing Institute Queen Mary University of London mcollardanuy@turing.ac.uk Kasra Hosseini The Alan Turing Institute khosseini@turing.ac.uk Giorgia Tolfo The British Library giorgia.tolfo@bl.uk Barbara McGillivray The Alan Turing Institute University of Cambridge bmcgillivray@turing.ac.uk Abstract As languages evolve historically, making com- putational approaches sensitive to time can improve performance on specific tasks. In this work, we assess whether applying histor- ical language models and time-aware meth- ods help with determining the correct sense of polysemous words. We outline the task of time-sensitive Targeted Sense Disambigua- tion (TSD), which aims to detect instances of a sense or set of related senses in historical and time-stamped texts, and address two main goals: 1) we scrutinize the effect of applying historical language models on the performance of several TSD methods and 2) we assess dif- ferent disambiguation methods that take into account the year in which a text was produced. We train historical BERT models on a corpus of nineteenth-century English books and draw on the Oxford English Dictionary (and its His- torical Thesaurus) to create historically evolv- ing sense representations. Our results show that using historical language models consis- tently improves performance whereas time- sensitive disambiguation helps especially with older documents. ∗ Contributions of each author (in alphabetical order): Conceptualization: KB, BMcG, FN; Data curation: KB, GT; Formal Analysis: MCA; Funding acquisition: BMcG; Method- ology: KB, MCA, KH, BMcG, FN; Project management: KB, BMcG, FN; Software: KB, MCA, KH, FN; Supervision: BMcG; Reproducibility: KB, MCA, FN; Writing: KB, MCA, BMcG, FN. 1 Introduction As language is in continuous flux, the question arises as to whether (and how) we should adapt Natural Language Processing (NLP) methods to the changing context in which texts are produced.1 This paper offers a novel contribution by assessing the extent to which making NLP models sensitive to time (or historically-aware) actually improves their performance. We present the task of time- sensitive Targeted Sense Disambiguation (TSD), which determines whether or not a token in a given time-stamped text is related to a specific sense of a lemma. TSD is a variation on Word Sense Disam- biguation (WSD), but (as we argue) it is of more practical relevance to research in digital history and cultural analysis. For example, if a historian wants to investigate the lemma machine2 in the sense of “anything that transmits force or directs its applica- tion”, TSD classifies whether the token machine (or one of its synonyms, e.g. machine) in a given text expresses this sense. This task assists researchers with tracking the evolution of very specific senses across time, instead of just words, as in Michel et al. (2011). 1Code and models used in this paper are accessible via Github and Zenodo, see …

Content truncated to 500 words