INTEGRATIVE NEUROSCIENCE ORIGINAL RESEARCH ARTICLE

These two problems were solved at once deriving the high-dimensional semantic space from a lexical co-occurrence matrix (Deerwester et al., 1990; Lund and Burgess, 1996). The main assumption for this derivation is that semantic proximity can be inferred by analyzing the statistical regularities in a text corpus. For instance, if the word “giraffe ” is mentioned in a text, it is likely that the words, “neck”, “zebra ” and “zoo ” (which are semantically related) will also be mentioned. The conditional probabilities of co-occurrence (i.e. the probability of fi nding “dog ” given that “cat” was mentioned) are thus likely to determine a good measure of semantic proximity. Lund and Burgess demonstrated this relation determining that: (1) Near neighbors in the co-ocurrence space correspond to related meanings (2) Clusters in this space cor-responded to semantic categories and (3) Similarity in co-occur-rence space determined the effect on RT in a semantic priming experiment. This result was in line with prior evidence which had shown a correlation between frequency of co-occurrence and free-word association probabilities (Spence and Owens, 1990), since the asso-ciative structure of words plays a central role in recall (Bousfi eld, 1953; Jenkins et al., 1958; Deese, 1959, 1965), cued recall (Nelson et al., 1992), and recognition (Nelson et al., 2001). Recently, Steyvers and colleagues showed that free word-association metrics (Nelson et al., 1999; Steyvers et al., 2004) performed better than corpus based metrics to predict human performance in various experi-ments of semantic memory. Assuming that the semantic space has been fully characterized, how should a model of retrieval and free-association be deter-mined from this space? The underlying hypothesis of most curren