| Among computational models of lexical semantic representation, semantic space models are unique in their invocation of the distributional hypothesis of word meaning: word meanings are correlated with the distribution of word forms in language. By recording words' co-occurrence patterns in a large corpus, the models attempt to produce a representation of lexical similarity structure that reflects human semantic representation.; A variety of semantic space models have been proposed, and many of them have proven successful at accounting for a broad range of semantic data, in particular semantic priming. However, finding tasks where the models make different predictions, and narrowing the space of plausible models, has been very difficult.; The present work uses child-directed speech as a realistic data domain for comparing the models. A number of the models from the cognitive scientific literature were re-implemented and trained on an age-stratified corpus of CDS from the CHILDES database.; Following a corpus analysis to determine the suitability of the corpus for supporting semantic space models, the models were compared through a series of 10 simulations on three increasingly rigorous semantic tasks. First, the models were compared on their abilities to discriminate between related and artificially created word pairs. All models showed significant discriminative capacity, but there was wide variation in discrimination. Next, the models were applied to modeling association strengths from word association norms and semantic distance in the WordNet lexicon. The performance of families of models varied with the type of semantic data, and not all models were reasonably successful on each task. Finally, the remaining models were given the task of modeling two sets of age-of-acquisition norms. The models performed comparably on this task.; This study represents the first time semantic space models have been systematically compared on a common corpus with common evaluation tasks. The results suggest that, first, there are nontrivial differences in the architectures of families of models, such that no one model is likely to perform optimally on all tasks. Second, the total variance explained by the models in each of the tasks was low, suggesting that there are limitations on the distributional hypothesis of word meaning. |
