On learning and representing social meaning in NLP: a sociolinguistic perspective

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 603–612 June 6–11, 2021. ©2021 Association for Computational Linguistics 603 On learning and representing social meaning in NLP: a sociolinguistic perspective Dong Nguyen Utrecht University The Netherlands d.p.nguyen@uu.nl Laura Rosseel Vrije Universiteit Brussel Belgium Laura.Rosseel@vub.be Jack Grieve University of Birmingham United Kingdom J.Grieve@bham.ac.uk Abstract The field of NLP has made substantial progress in building meaning representations. However, an important aspect of linguistic meaning, social meaning, has been largely overlooked. We introduce the concept of so- cial meaning to NLP and discuss how insights from sociolinguistics can inform work on rep- resentation learning in NLP. We also identify key challenges for this new line of research. 1 Introduction Variation is inherent to language. Any variety of language provides its users with a multitude of lin- guistic forms—e.g., speech sounds, words, gram- matical constructions—to express the same refer- ential meaning. Consider, for example, the many ways of pronouncing a given word or the variety of words that can refer to a given concept. Linguistic variation is the primary object of in- quiry of sociolinguistics, which has a long history of describing and explaining variation in linguistic form across society and across levels of linguistic analysis (Tagliamonte, 2015). Perhaps the most basic finding of the field is that linguistic variation allows for the expression of social meaning, infor- mation about the social background and identity of the language user. Such sociolinguistic varia- tion adds an additional layer of meaning onto the basic referential meaning communicated by any utterance or text. Understanding the expression of social meaning based on linguistic variation is a crucial part of the linguistic knowledge of any language user, drawn upon continuously during both the production and processing of natural lan- guage. The relationship between variation and so- cial meaning, however, has only begun to be ex- plored computationally (e.g., Pavalanathan et al. (2017)). Studies have shown, for example, that words, capitalisation, or the language variety used can index political identity (Shoemark et al., 2017; Stewart et al., 2018; Tatman et al., 2017). Despite general acceptance of the link between linguistic variation and social meaning in linguis- tics, NLP has largely ignored this relationship. Nev- ertheless, the importance of linguistic variation more generally is increasingly being acknowledged in NLP (Nguyen et al., 2016). NLP tools are usu- ally developed for standard varieties of language, and therefore tend to under-perform on texts writ- ten in varieties that diverge from the ‘standard’, including language identification (Blodgett et al., 2016), dependency parsing (Blodgett et al., 2016), and POS tagging (Hovy and Søgaard, 2015). One approach to overcoming the challenges posed by linguistic variation is text normalisation (Han and Baldwin, 2011; Liu et al., 2011). Normal- isation transforms non-standard texts into a more standardised form, which can then be analysed more accurately using NLP models trained on stan- dard language data. Text normalisation, however, removes rich social signals encoded via sociolin- guistic variation. Other …

Content truncated to 500 words