Can a naive discrimination learning model classify inflected forms of polysemous nouns?

Abstract

The interactive approach in the modularity-of-syntax debate suggests that we simultaneously process a vast amount of information regarding both semantic and syntactic meaning of a word. When we add to the consideration the ubiquity of ambiguity phenomena such as polysemy (multiple related senses), the information that most words convey seems quite demanding. Knowing that multiple approaches demonstrated the interaction between semantic and syntactic meanings (Kostić et al., 2004; Mišić & Filipović Đurđević, 2021), the aim of this study was to investigate whether this complexity can arise from a simple and cognitively plausible learning mechanism such as Rescorla-Wagner rule (RW; Rescorla & Wagner, 1972) implemented in the ndl framework (Arppe et al., 2015). Starting from the fact ndl can perform as a classifier (Milin et al., 2017, Sering et al., 2018), we wanted to test whether it can predict the correct inflected form of a polysemous word based on both orthographic and semantic cues. For the purposes of this paper we conducted two simulation studies with two different networks. Orthographic network learned inflected forms of 150 Serbian polysemous words (outcomes) based on the orthographic cues (trigrams; bradom – #br, bra, rad, ado, dom, om#). Each appearance of a target word in corpus was considered a learning event. These events were used to train the network using the RW rule. For each of the unique learning events, we calculated the bottom-up support that the outcome gets from cues that appeared by summing cue-outcome connection weights. The network correctly classified 94.6% of the outcomes. The semantic network utilized co-occurring words from the corpus as cues with the same outcomes. We preselected 3000 most frequent nouns, verbs, and adjectives (Frequency Dictionary of Contemporary Serbian Language; Kostić, 1999) for context words. Learning events in this simulation were built by moving a seven-point window centered on the outcome through the corpus. Whenever any of the context words were found within the positions flanking the outcome word, they would be considered cues in that learning event. RW rule training and classification on 880975 unique learning events was the same as in the orthographic network, however the success rate was opposite – only 3.4% of the words were correctly classified. The success of the orthographic network is not surprising, considering the small sample of words which reduces the need for cue competition. On the other hand, semantic cues i.e. context words seem to have the exact opposite problem. When most frequent words are selected, they are shared between many of the outcomes. Fewer distinct cues make discrimination between outcomes harder. Therefore, this low classification rate in the semantic network does not reject the interactive approach, but suggests the cue selection that does not capture the complex semantic and syntactic information.