Joint Incremental Disfluency Detection and Dependency Parsing

Joint Incremental Disfluency Detection and Dependency Parsing Matthew Honnibal Department of Computing Macquarie University Sydney, Australia matthew.honnibal@mq.edu.edu.au Mark Johnson Department of Computing Macquarie University Sydney, Australia mark.johnson@mq.edu.edu.au Abstract We present an incremental dependency parsing model that jointly performs disflu- ency detection. The model handles speech repairs using a novel non-monotonic tran- sition system, and includes several novel classes of features. For comparison, we evaluated two pipeline systems, us- ing state-of-the-art disfluency detectors. The joint model performed better on both tasks, with a parse accuracy of 90.5% and 84.0% accuracy at disfluency detection. The model runs in expected linear time, and processes over 550 tokens a second. 1 Introduction Most unscripted speech contains filled pauses (ums and uhs), and errors that are usually edited on-the-fly by the speaker. Disfluency detection is the task of detecting these infelicities in spoken language transcripts. The task has some imme- diate value, as disfluencies have been shown to make speech recognition output much more dif- ficult to read (Jones et al., 2003), but has also been motivated as a module in a natural language understanding pipeline, because disfluencies have proven problematic for PCFG parsing models. Instead of a pipeline approach, we build on re- cent work in transition-based dependency parsing, to perform the two tasks jointly. There have been two small studies of dependency parsing on un- scripted speech, both using entirely greedy pars- ing strategies, without a direct comparison against a pipeline architecture (Jorgensen, 2007; Rasooli and Tetreault, 2013). We go substantially beyond these pilot studies, and present a system that com- pares favourably to a pipeline consisting of state- of-the-art components. Our parser largely follows the design of Zhang and Clark (2011). We use a structured averaged perceptron model with beam- search decoding (Collins, 2002). Our feature set is based on Zhang and Clark (2011), and our transition-system is based on the arc-eager system of Nivre (2003). We extend the transition system with a novel non-monotonic transition, Edit. It allows sen- tences like ‘Pass the pepper uh salt’ to be parsed incrementally, without the need to guess early that pepper is disfluent. This is achieved by re- processing the leftward children of the word Edit marks as disfluent. For instance, if the parser at- taches the to pepper, but subsequently marks pep- per as disfluent, the will be returned to the stack. We also exploit the ease with which the model can incorporate arbitrary features, and design a set of features that capture the ‘rough copy’ structure of some speech repairs, which motivated the Johnson and Charniak (2004) noisy channel model. Our main comparison is against two pipeline systems, which use the two current state-of-the- art disfluency detection systems as pre-processors to our parser, minus the custom disfluency fea- tures and transition. The joint model compared favourably to the pipeline parsers at both tasks, with an unlabelled attachment score of 90.5%, and 84.0% accuracy at detecting speech repairs. An ef- ficient implementation is available under an open- source license.1 The future prospects of …

Content truncated to 500 words