How well can a corpus-derived co-occurrence network simulate human associative behavior?

Abstract

Free word associations are the words people spontaneously come up with in response to a stimulus word. Such information has been collected from test persons and stored in databases. A well known example is the Edinburgh Associative Thesaurus (EAT). We will show in this paper that this kind of knowledge can be acquired automatically from corpora, enabling the computer to produce similar associative responses as people do. While in the past test sets typically consisted of approximately 100 words, we will use here a large part of the EAT which, in total, comprises 8400 words. Apart from extending the test set, we consider different properties of words: saliency, frequency and part-of-speech. For each feature categorize our test set, and we compare the simulation results to those based on the EAT. It turns out that there are surprising similarities which supports our claim that a corpus-derived co-occurrence network can simulate human associative behavior, i.e. an important part of language acquisition and verbal behavior.

Proceedings of the 8th Workshop on Cognitive Modeling and Computational Linguistics (CMCL 2018)

“…Despite their 'localist' approach in which a word is simply represented by a node (rather than using distributed representations), such models are a useful tool in the study of lexical access and acquisition. In particular, they have successfully replicated patterns of human verbal behavior in free word association (Enguix et al, 2014;Gruenenfelder et al, 2015), semantic fluency tasks , lexical growth/acquisition (Stella et al, 2017;Bilson et al, 2015), assessment of semantic similarity (Jackson and Bolger, 2014;De Deyne et al, 2016), etc.…”

Graph‐Theoretic Properties of Networks Based on Word Association Norms: Implications for Models of Lexical Semantic Memory

“…There has been little research directly testing the assumption that word associations reflect associative relatedness rather than featural relatedness or contextual similarity, largely because controlling confounds in stimulus selection is practically impossible. That research which does exist (Enguix, Rapp, & Zock, ; Spence & Owens, ; Wettler et al., ) has found that co‐occurrence frequency was correlated with the probability that one word was produced as a response to the other in a word association task, a finding interpreted as supporting the hypothesis that word associations reflect associative similarity. None of these studies, however, attempted to control for the confound between associative similarity and featural or contextual similarity, making their results difficult to interpret.…”

Modeling bilingual word associations as connected monolingual networks

“…Despite their 'localist' approach in which a word is simply represented by a node (rather than using distributed representations), such models are a useful tool in the study of lexical access and acquisition. In particular, they have successfully replicated patterns of human verbal behavior in free word association (Enguix et al, 2014;Gruenenfelder et al, 2015), semantic fluency tasks (Abbott et al, 2015;Nematzadeh et al, 2016), lexical growth/acquisition (Stella et al, 2017;Bilson et al, 2015), assessment of semantic similarity (Jackson and Bolger, 2014;De Deyne et al, 2016), etc.…”