Selecting Corpus-Semantic Models for Neurolinguistic Decoding

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 114–123, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Selecting Corpus-Semantic Models for Neurolinguistic Decoding Brian Murphy Machine Learning Dept. Carnegie Mellon University Pittsburgh, USA brianmurphy@cmu.edu Partha Talukdar Machine Learning Dept. Carnegie Mellon University Pittsburgh, USA ppt@cs.cmu.edu Tom Mitchell Machine Learning Dept. Carnegie Mellon University Pittsburgh, USA tom.mitchell@cs.cmu.edu Abstract Neurosemantics aims to learn the mapping between concepts and the neural activity which they elicit during neuroimaging ex- periments. Different approaches have been used to represent individual concepts, but current state-of-the-art techniques require extensive manual intervention to scale to arbitrary words and domains. To over- come this challenge, we initiate a system- atic comparison of automatically-derived corpus representations, based on various types of textual co-occurrence. We find that dependency parse-based features are the most effective, achieving accuracies similar to the leading semi-manual ap- proaches and higher than any published for a corpus-based model. We also find that simple word features enriched with directional information provide a close-to- optimal solution at much lower computa- tional cost. 1 Introduction The cognitive plausibility of computational models of word meaning has typically been tested using behavioural benchmarks, such as identification of synonyms among close asso- ciates (the TOEFL task for language learners, see e.g. Landauer and Dumais, 1997); emulating elicited judgments of pairwise similarity (such as Rubenstein and Goodenough, 1965); judgments of category membership (e.g. Battig and Mon- tague, 1969); and word priming effects (Lund and Burgess, 1996). Mitchell et al. (2008) in- troduced a new task in neurosemantic decoding – using models of semantics to learn the map- ping between concepts and the neural activity which they elicit during neuroimaging experi- ments. This was achieved with a linear model which used training data to find neural basis im- ages that correspond to the assumed semantic dimensions (for instance, one such basis image might be the activity of the brain for words rep- resenting animate concepts), and subsequently used these general patterns and known seman- tic dimensions to infer the fMRI activity that should be elicited by an unseen stimulus con- cept. Follow-on work has experimented with other neuroimaging modalities (Murphy et al., 2009), and with a range of semantic models in- cluding elicited property norms (Chang et al., 2011), corpus derived models (Devereux and Kelly, 2010; Pereira et al., 2011) and structured ontologies (Jelodar et al., 2010). The current state-of-the-art performance on this task is achieved using models that are hand- tailored in some respect, whether using manual annotation tasks (Palatucci et al., 2009), use of a domain-appropriate curated corpus (Pereira et al., 2011), or selection of particular collocates to suit the concepts to be described (Mitchell et al., 2008). While these approaches are clearly very successful, it is questionable whether they are a general solution to describe the vari- ous parts-of-speech and semantic domains that make up a speaker’s vocabulary. The Mitchell et al. (2008) 25-verb model would probably have to be extended to describe the lexicon at large, and it …

Content truncated to 500 words