Abstract:

Unpredictable variation is rare in language. Explanations for this include a language-specific regularization bias (e.g. Reali & Griffiths, 2009), general constraints on memory (e.g. Hudson Kam & Newport, 2005), or both (e.g. Perfors, 2012; Ferdinand, Thompson, Kirby, & Smith, 2013). Experiments on lexical regularization typically study how words in free synonymous relationships become increasingly deterministic through use (e.g. some words drop from use). There is also experimental evidence that learners regularize homonymous relationships (Vouloumanos, 2010), but to date no experimental design has directly compared the relative regularization of synonyms versus homonyms. This is an important comparison to make because synonyms and homonyms have asymmetrical functional roles in communication (Hurford, 2003) and the jury is still out as to which of these two regularization biases are better for evolving effective communication systems. On one hand, Hurford proposes that there is less bias against homonyms because they are more common in language than synonyms and Piantadosi, Tily, and Gibson (2012) argue for the communicative function of ambiguous lexicons. On the other hand, Doherty (2004) demonstrates children’s difficulty in learning homonyms and Spike, Stadler, Kirby, and Smith (2013) show that self-organizing novel lexicons require a bias against homonymy but not synonymy.

We extend the experimental paradigm of Ferdinand et al. (2013) to investigate the relative regularization of synonyms versus homonyms. 128 participants were trained on one of two artificial mini-languages with identically matched distributions of variation. In the synonyms condition, this variation was over word forms and in the homonyms condition it was over referents. Regularization is quantified by the drop in entropy of the words and referents that participants produced when tested on their mini-language. Participants regularized 67% of the variation among homonyms (t(63) = −12.8169, p < .001) and 56% of the variation among synonyms (t(63) = −10.5526, p < .001). However, there was no significant difference between these conditions (t(126) = 1.3518, p = 0.18), suggesting that learners compress synonymous and homonymous variation similarly.

This experiment was repeated with non-linguistic stimuli, where participants learned the mappings between marbles and the different containers they were drawn from. Participants also regularized the non-linguistic stimuli, eliminating 42% of the variation among containers (t(63) = −7.277, p < .001) and 32% of the variation among marbles (t(63) = −6.6908, p < .001), again with no significant difference between conditions (t(126) = 1.5049, p = 0.13). This suggests a domain-general component to linguistic regularization. However, participants in the linguistic conditions regularized significantly more than those in the non-linguistic conditions (F(252) = 11.259,p < .001). We conclude that regularization results from general-purpose compression during learning, which can be ramped up for effective communication with linguistic stimuli, and operates similarly on synonyms and homonyms.

Download Paper (pdf)

Citation:

Ferdinand V. and Spike M. (2016). Word Learners Regularize Synonyms And Homonyms Similarly. In S.G. Roberts, C. Cuskley, L. McCrohon, L. Barceló-Coblijn, O. Fehér & T. Verhoef (eds.) The Evolution of Language: Proceedings of the 11th International Conference (EVOLANG11). Available online: http://evolang.org/neworleans/papers/82.html