Modeling Word Meaning: Distributional Semantics and the Corpus Quality-Quantity Trade-Off

摘要

Dictionaries constructed using distributional models of lexical semantics have a wide range of applications in NLP and in the modeling of linguistic cognition. However when constructing such a model, we are faced with range of corpora to choose from. Often there is a choice between small carefully constructed corpora of well-edited text, and very large heterogeneous collections harvested automatically from the web. There may also be differences in the distribution of genres and registers in such corpora. In this paper we examine these trade-offs by constructing a simple SVD-reduced word-collocate model, using four English corpora: the Google Web 5-gram collection, the Google Book 5-gram collection, the English Wikipedia, and collection of short social messages harvested from Twitter. Since these models need to encode semantics in a way that approximates the mental lexicon, we evaluate the felicity of the resulting semantic representations using a set of behavioral and neural-activity benchmarks that depend on word-similarity. We find that the quality of the input text has a very strong effect on the performance of the output model, and that a corpus of high quality at a small size can outperform a corpus of poor quality that is many orders of magnitude larger. We also explore the semantic closeness of the models using their mutual information overlap to interpret the similarity of corpus texts.