A Cognitive Model of Semantic Network Learning

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 244–254, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics A Cognitive Model of Semantic Network Learning Aida Nematzadeh, Afsaneh Fazly, and Suzanne Stevenson Department of Computer Science University of Toronto {aida,afsaneh,suzanne}@cs.toronto.edu Abstract Child semantic development includes learning the meaning of words as well as the semantic relations among words. presumed outcome of semantic develop-A ment is the formation of a semantic net- work that reflects this knowledge. We present an algorithm for simultaneously learning word meanings and gradually growing a semantic network, which ad- heres to the cognitive plausibility require- ments of incrementality and limited com- putations. We demonstrate that the seman- tic connections among words in addition to their context is necessary in forming a semantic network that resembles an adult’s semantic knowledge. 1 Introduction Child semantic development includes the acquisi- tion of word-to-concept mappings (part of word learning), and the formation of semantic connec- tions among words/concepts. There is consid- erable evidence that understanding the semantic properties of words improves child vocabulary ac- quisition. In particular, children are sensitive to commonalities of semantic categories, and this abstract knowledge facilitates subsequent word learning (Jones et al., 1991; Colunga and Smith, 2005). Furthermore, representation of semantic knowledge is significant as it impacts how word meanings are stored in, searched for, and retrieved from memory (Steyvers and Tenenbaum, 2005; Griffiths et al., 2007). Semantic knowledge is often represented as a graph (a semantic network) in which nodes cor- respond to words/concepts1, and edges specify 1Here we assume that the nodes of a semantic network are word forms and its edges are determined by the semantic features of those words. the semantic relations (Collins and Loftus, 1975; Steyvers and Tenenbaum, 2005). Steyvers and Tenenbaum (2005) demonstrated that a seman- tic network that encodes adult-level knowledge of words exhibits a small-world and scale-free struc- ture. That is, it is an overall sparse network with highly-connected local sub-networks, where these sub-networks are connected through high-degree hubs (nodes with many neighbours). Much experimental research has investigated the underlying mechanisms of vocabulary learn- ing and characteristics of semantic knowledge (Quine, 1960; Bloom, 1973; Carey and Bartlett, 1978; Gleitman, 1990; Samuelson and Smith, 1999; Jones et al., 1991; Jones and Smith, 2005). However, existing computational models focus on certain aspects of semantic acquisition: Some researchers develop computational models of word learning without considering the acqui- sition of semantic connections that hold among words, or how this semantic knowledge is struc- tured (Siskind, 1996; Regier, 2005; Yu and Bal- lard, 2007; Frank et al., 2009; Fazly et al., 2010). Another line of work is to model formation of semantic categories but this work does not take into account how word meanings/concepts are ac- quired (Anderson and Matessa, 1992; Griffiths et al., 2007; Fountain and Lapata, 2011). Our goal in this work is to provide a cognitively- plausible and unified account for both acquiring and representing semantic knowledge. The re- quirements for cognitive …

Content truncated to 500 words