Detecting Speech Repairs Incrementally Using a Noisy Channel Approach

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1371–1378, Beijing, August 2010 Detecting Speech Repairs Incrementally Using a Noisy Channel Approach Simon Zwarts, Mark Johnson and Robert Dale Centre for Language Technology Macquarie University {simon.zwarts|mark.johnson|robert.dale}@mq.edu.au Abstract Unrehearsed spoken language often contains disfluencies. In order to cor- rectly interpret a spoken utterance, any such disfluencies must be identi- fied and removed or otherwise dealt with. Operating on transcripts of speech which contain disfluencies, our particular focus here is the identifica- tion and correction of speech repairs using a noisy channel model. Our aim is to develop a high-accuracy mecha- nism that can identify speech repairs in an incremental fashion, as the ut- terance is processed word-by-word. We also address the issue of the evalu- ation of such incremental systems. We propose a novel approach to evalua- tion, which evaluates performance in detecting and correcting disfluencies incrementally, rather than only assess- ing performance once the processing of an utterance is complete. This demon- strates some shortcomings in our ba- sic incremental model, and so we then demonstrate a technique that improves performance on the detection of disflu- encies as they happen. 1 Introduction One of the most obvious differences between written language and spoken language is the fact that the latter presents itself incremen- tally over some time period. Most natural lan- guage processing applications operate on com- plete sentences; but for real time spontaneous speech, there are potential benefits to incre- mentally processing the input so that a system can stay responsive and interact directly be- fore a speaker’s utterance is complete. Work in psycholinguistics supports the view that the human parsing mechanism works incremen- tally, with partial semantic interpretations be- ing produced before the complete utterance has been heard (Marslen-Wilson, 1973). Our interest is in developing similarly incremental processing techniques for natural language in- terpretation, so that, for example, a speech recognizer might be able to interject during a long utterance to object, cut the speaker short, or correct a mistaken assumption; such a mechanism is even required for the appro- priate timing of backchannel signals. Addi- tionally the incremental nature of the model allows potential application of this model in speech recognition models. Another feature of unrehearsed spoken lan- guage that has no obvious correlate in written language is the presence of disfluencies.1 Dis- fluencies are of different types, ranging from simple filled pauses (such as um and uh) to more complicated structures where the se- quence of words that make up the utterance is ‘repaired’ while it is being produced. Whereas simpler disfluencies may be handled by sim- ply deleting them from the sequence of words under consideration, the editing terms in a speech repair are part of the utterance, and therefore require more sophisticated process- ing. There are three innovations in the present paper. First, we demonstrate that a noisy channel model of speech repairs can work ac- curately in an incremental fashion. Second, we provide an approach to the evaluation of 1Although …

Content truncated to 500 words