A Comparison of Statistical Models for the Extraction of Lexical Information from Text Corpora

The Syntagmatic Paradigmatic model (SP; Dennis & Harrington 2001, Dennis submitted) and the Pooled Adjacent Context model (PAC; Redington, Chater & Finch 1998) are compared on their ability to extract syntactic, semantic and associative information from a corpus of text. On a measure of syntactic class (and subclass) information based on the WordNet lexical database (Miller 1990), the models performed similarly with a small advantage for the PAC model. On a measure of semantic structure based on the similarities produced by Latent Semantic Analysis (LSA; Landauer & Dumais 1997), the models performed equivalently with a small advantage for the SP model. On a measure of associative information based on the free association norms of Nelson, McEvoy & Schreiber (1999), the SP model shows a substantive advantage over the PAC model producing more than twice as many associates.