Unsupervised induction of latent semantic grammars with application to parsing
This dissertation investigated the utility of latent semantic information for unsupervised grammar induction. Utility was directly measured by the relative performance of induced latent semantic grammars in parsing sentences and by the relative performance of induced grammars on making meaning judgments. Latent semantic grammars did not outperform the state of the art method for unsupervised parsing. However, latent semantic grammars performed as well or better than all other methods, indicating that latent semantic parsing is worthy of future research. Latent semantic grammars had equivalent performance to Latent Semantic Analysis on comparative meaning tasks, suggesting that latent semantic grammars are a patentable alternative to Latent Semantic Analysis. Finally, the non-orthogonal approach to singular value decomposition that was developed in order to create latent semantic grammars is applicable in many scientific fields. By allowing supercomputer-sized matrices to be computed on standard personal computers, this non-orthogonal approach makes large scale singular value decomposition accessible to everyone.