HCP with PSMA: A Robust Spoken Language Parser

" Spoken language " is a field of natural language processing, which deals with transcribed speech utterances. The processing of spoken language is much more complex and complicated than processing standard, grammatically correct natural language , and requires special treatment of typical speech phenomena called " disfluencies " , like corrections , interjections and repetitions of words or phrases. We present a parsing technique that utilizes a Hybrid Connectionist Parser (HCP) extended with a Phrase Structure Matching Algorithm (PSMA) for the syntactic processing (parsing) of spoken language. The HCP is a hybrid of a connec-tionist and a symbolic approach, which builds a neural network dynamically based on a given context free grammar and an input sentence. It has an advantage over traditional parsers due to the use of graded activation and activation passing in the network. These features were exploited in techniques to detect and repair disfluencies, in combination with the Phrase Structure Matching Algorithm , which operates on the graphical structure of the network. This approach to spoken language parsing is purely syntactical and – in contrast to other work – does not require a modified grammar for spoken language, or information derived from the speech input, nor any pre-marking of disfluen-cies. We implemented the combined HCP and PSMA parser and tested it on a standard corpus for spoken language, the HCRC Map Task Corpus. The experiments showed very good results, especially for a purely syntactic method, which detects as well as corrects disfluencies.