Speakers enhance contextually confusable words

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1991–2002 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 1991 Speakers Enhance Contextually Confusable Words Eric Meinhardt Department of Linguistics UC San Diego La Jolla, CA, USA 92093 emeinhardt@ucsd.edu Eric Baković Department of Linguistics UC San Diego La Jolla, CA, USA 92093 ebakovic@ucsd.edu Leon Bergen Department of Linguistics UC San Diego La Jolla, CA, USA 92093 lbergen@ucsd.edu Abstract Recent work has found evidence that natural languages are shaped by pressures for efficient communication — e.g. the more contextually predictable a word is, the fewer speech sounds or syllables it has (Piantadosi et al. 2011). Re- search on the degree to which speech and lan- guage are shaped by pressures for effective communication — robustness in the face of noise and uncertainty — has been more equiv- ocal. We develop a measure of contextual con- fusability during word recognition based on psychoacoustic data. Applying this measure to naturalistic speech corpora, we find evidence suggesting that speakers alter their productions to make contextually more confusable words easier to understand. 1 Introduction A major open question in the study of natural lan- guages is the extent to which pressures for effi- cient communication shape the online production choices of speakers or the system of forms and form-meaning mappings. Zipf (1936, 1949) fa- mously noted that highly frequent words tend to be shorter and hypothesized that this could be ex- plained in terms of pressures for efficient commu- nication: the average cost of producing a word is lower than it would be otherwise. More recent work has formalized hypotheses about the effect of communicative pressures on language usage and design using tools from infor- mation theory (Shannon 1948, Cover and Thomas 2012) and rational analysis (Anderson 1990, 1991). This work has found evidence that meanings are allocated to word types in a way that minimizes speaker effort (Piantadosi et al. 2011, 2012), and that this appears to be at least partly explainable by online production choices (Mahowald et al. 2013). While this research offers evidence that lexi- cons and the production choices of speakers are shaped by pressures for efficient communication, other work examining how much words and lexi- cons are shaped by pressures for ensuring effective communication in the face of noise and uncertainty has been more equivocal. This work has found evi- dence that words with greater neighborhood size or density — that is, words that have a greater num- ber of similar-sounding neighbors — have faster onset of production, and have lower overall dura- tions. Words with greater neighborhood density also take longer for listeners to recognize and com- prehend, and have less acoustically distinctive vow- els (Vitevitch 2002, Gahl et al. 2012; see Vitevitch and Luce 2016 for review). This work provides a challenge for communicatively-oriented models of production: words with greater numbers of similar-sounding neighbors seem likely to be more confusable, and therefore speakers would be predicted to decrease the likelihood of noise by, …

Content truncated to 500 words