Does latent semantic analysis actually have a latent structure

Latent Semantic Analysis (LSA) is a statistical technique that computes similarity comparisons for terms and texts. LSA allegedly captures the “latent” structure of word usage by taking note of the context that words appear in. Researchers have repeatedly demonstrated impressive results of the technique and have even suggested that LSA might be used as a viable theory of human knowledge representation.; The goal of this dissertation was to test the ability of LSA to capture some basic conceptual relations. Three experiments were conducted in order to test LSA. Experiment 1 was a comparison of word pairs. Word pairs like bird - nest should have a higher cosine value (i.e., a measure of similarity) with pairs like bear - cave than with word pairs like monkey - stable; the first two word pairs both share the relation of “lives in”, whereas the second pair shares no obvious relation. The results indicate that LSA was able to recognize the underlying relationship between two sets of word pairs. Experiment 2 was a comparison of word pairs to relationship labels. Thus, word pairs like bird - nest should have a higher cosine to the relation of “lives in” than the relation “is made of”. Again, LSA was able to detect the underlying relationship. Experiment 3 was a comparison of word pairs and an analysis of the amount of information that was not redundant with the relationship labels. Thus, when the relation “lives in” was added to the word pairs of bird - nest and bear - cave, there should be a minimal amount of new information added than when the relation “is made of” was added. LSA was not able to pick up this difference. Two of the three experiments demonstrated that LSA was capable of recognizing the underlying relationship between sets of word pairs that share a relation and the relationship between a word pair and a relationship label. These results should provide further evidence that LSA does capture the “latent” structure of word usage.