Short-Term Meaning Shift: A Distributional Exploration

Proceedings of NAACL-HLT 2019, pages 2069–2075 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 2069 Short-Term Meaning Shift: A Distributional Exploration Marco Del Tredici∗ Raquel Fern´andez∗ Gemma Boleda† ∗University of Amsterdam †Universitat Pompeu Fabra {m.deltredici|raquel.fernandez}@uva.nl gemma.boleda@upf.edu Abstract We present the first exploration of meaning shift over short periods of time in online com- munities using distributional representations. We create a small annotated dataset and use it to assess the performance of a standard model for meaning shift detection on short- term meaning shift. We find that the model has problems distinguishing meaning shift from referential phenomena, and propose a measure of contextual variability to remedy this. 1 Introduction Semantic change has received increasing attention in empirical Computational Linguistics / NLP in the last few years (Tang, 2018; Kutuzov et al., 2018). Almost all studies so far have focused on meaning shift in long periods of time—decades to centuries. However, the genesis of meaning shift and the mechanisms that produce it operate at much shorter time spans, ranging from the on- line agreement on words’ meaning in dyadic in- teractions (Brennan and Clark, 1996) to the rapid spread of new meanings in relatively small com- munities of people in (Wenger, 1998; Eckert and McConnell-Ginet, 1992). This paper is, to the best of our knowledge, the first exploration of the latter phenomenon—which we call short-term mean- ing shift—using distributional representations. More concretely, we focus on meaning shift arising within a period of 8 years, and explore it on data from an online community of speakers, be- cause there the adoption of new meanings happens at a fast pace (Clark, 1996; Hasan, 2009). Indeed, short-term shift is usually hard to observe in stan- dard language, such as the language of books or news, which has been the focus of long-term shift studies (e.g., Hamilton et al., 2016; Kulkarni et al., 2015), since it takes a long time for a new meaning to be widely accepted in the standard language. Our contribution is twofold. First, we create a small dataset of short-term shift for analysis and evaluation, and qualitatively analyze the types of meaning shift we find.1 This is necessary because, unlike studies of long-term shift, we cannot rely on material previously gathered by linguists or lex- icographers. Second, we test the behavior of a standard distributional model of semantic change when applied to short-term shift. Our results show that this model successfully detects most shifts in our data, but it overgeneralizes. Specifically, the model gets confused with contextual changes due to speakers in the community often talking about particular people and events, which are frequent on short time spans. We propose to use a measure of contextual variability to remedy this and show- case its potential to spot false positives of refer- ential nature like these. We thus make progress in understanding the nature of semantic shift and towards improving computational models thereof. 2 Related Work Distributional models of semantic change are based on the hypothesis that a …

Content truncated to 500 words