Distributed meanings and senses within discrimination learning framework – proof of concept

Abstract

Discrimination learning (DL), a simple learning mechanism has proven to be a powerful model for describing language processing. In this paper we contribute to describing the semantic phenomena in the light of DL. We do so by continuing to focus on polysemy advantage and homonymy disadvantage. Whereas the former is most dominantly explained by a higher activation through the shared semantic core of multiple related senses, the latter seems to be a consequence of sharing activation across unrelated semantic features of multiple meanings. Filipović Đurđević and Kostić (2021) crossed DL and distributional semantics to demonstrate that the relatedness of polysemous senses could be operationalised as the overlap at the level of the outcomes. Our aim was to test whether this activation pattern can arise in endstate of learning when we simulate learning of polysemous, homonymous, and unambiguous words using a small scale model over which we have full control. We used a toy lexicon containing two entries from each of the three groups. Cues were bigrams made from nine-letter strings randomly generated from five letters, to introduce cue competition. To simulate all aspects of ambiguity, each word had four outcomes. Each unambiguous word and each homonym meaning had four unique outcomes, and each sense of a polyseme had one unique outcome, one shared by only one other sense, and two outcomes shared by all senses. The results revealed that cue-outcome weights were the highest for the polysemous words, thus corroborating the findings of Filpović-Đurđević and Kostić (2021). However, no difference in cue-outcome weights was observed between homonyms and unambiguous words. This simple simulation continues to inform future studies on how polysemous senses could be defined when corpus data is used. The distributional hypothesis (Harris, 1954) states that similar words, or in this case senses, appear in similar contexts. Our simulation suggests that outcomes should be defined in a way where homonym meanings do not share any outcomes, and polysemous words do. However, further simulations on toy corpora are needed in order to more precisely understand the supposed structure of distributed meanings/senses.