Constructing semantic representations using the MDL principle

Abstract

Words receive a significant part of their meaning from use in communicative settings. The formal mechanisms of lexical acquisition, as they apply to rich situational settings,may also be studied in the limited case of corpora of written texts. This work constitutes an approach to deriving semantic representations for lexemes using techniques from statistical induction. In particular, a number of variations on the MDL principle were applied to selected sample sets and their influence on emerging theories of word meaning explored. We found that by changing the definition of description length for data and theory - which is equivalent to different encodings of data and theory - we may customize the emerging theory, augmenting and altering frequency effects. Also the influence of the stochastic properties of the size of the theory has been demonstrated. The results consist in a set of distributional properties of lexemes, which reflect cognitive distinctions in the meaning of words.