Latent semantic language modeling for speech recognition

Statistical language models used in large vocabulary speech recognition must properly capture the various constraints, both local and global, present in the language. While n-gram modeling readily accounts for the former, it has been more difficult to handle the latter, and in particular long-term semantic dependencies, within a suitable data-driven formalism. This paper focuses on the use of latent semantic analysis (LSA) for this purpose. The LSA paradigm automatically uncovers meaningful associations in the language based on word-document co-occurrences in a given corpus. The resulting semantic knowledge is encapsulated in a (continuous) vector space of comparatively low dimension, where are mapped all (discrete) words and documents considered. Comparison in this space is done through a simple similarity measure, so familiar clustering techniques can be applied. This leads to a powerful framework for both automatic semantic classification and semantic language modeling. In the latter case, the large-span nature of LSA models makes them particularly well suited to complement conventional n-grams. This synergy can be harnessed through an integrative formulation, in which latent semantic knowledge is exploited to judiciously adjust the usual n-gram probability. The paper concludes with a discussion of intrinsic trade-offs, such as the influence of training data selection on the resulting performance enhancement.