Finite-state pre-processing for natural language analysis

Wide-coverage natural language parsers are typically not very efficient. Finite-state techniques are less powerful, but offer the advantage of being very fast, and good at representing language locally. This dissertation constitutes empirical research into the construction and use of a finite-state approximation of a wide-coverage parser to increase parsing performance. The finite-state approximation is in the form of a hidden Markov model, inferred from parser-annotated data. This model is used in a part-of-speech tagger, which is applied in various ways and using several different models to reduce ambiguity in parsing, by setting it up as a filter that removes unlikely options in the first stage of parsing.