A Model of Speech Repairs

Most dialog systems ignore the problem of speech repairs and editing terms um uh etc or use preprocessing techniques to eliminate them from the input These systems also typically enforce a strict turn taking protocol that does not allow speakers to interrupt each other This paper de scribes a parser that can process input containing editing terms speech repairs and second speaker interruptions and include these phenomena in its output Such a parser allows a dialog system to reason about why editing terms were uttered maybe the speaker was uncertain embarrassed reluctant to commit etc The reparandum cor rected material in a speech repair also plays an important role as it may be referenced later take the oranges to Elmira uh I mean take them to Corning Reparanda may also give insight into the speaker s intentions pick up tankers in uh how many cars can an engine pull Second speaker interruptions can provide evidence that the in terrupter is listening if they utter a backchannel such as uh huh or that neither speaker is hearing the other both speakers are talking at the same time This type of evidence is crucial for appli cations such as business meeting summarization Dialog systems are in their infancy Systems use mod ules such as speech recognizers parsers reasoning sys tems text generators and speech synthesizers to inter act with users Researchers are just starting to exper iment with better interfaces between these modules to improve performance For example some speech rec ognizers give parsers the n best word sequences they nd instead of just the sequence they assign the high est probability Another area of development involves going beyond the typical command response interface of a dialog system and allowing users to interrupt the system and speak more than one utterance per turn Clearly humans exhibit both of these behaviors when talking to each other It is also uncontroversial that humans have a high degree of communication between processes in the brain that decode words recognize syn tactic structure and perform general reasoning Lower level information such as word stress is something that we can reason about and pragmatic expectations of what someone is likely to say can help word recogni tion Our work has focused on creating a parser that can process a stream of words with editing terms um I mean speech repairs and second speaker interrup tions This parser provides a syntactic representation to higher level reasoning modules such as a dialog man ager that includes this information This parser is novel in that parsers typically make the simplifying as sumption that any editing terms or speech repairs will be removed from the input Another aspect neglected by current parsers is that people interrupt each other in conversation Our parser allows second speaker inter ruptions and continuations Thus it can handle third party human human conversations as well as allowing users to interrupt the system and vice versa In the rst section of the paper we describe why edit ing terms speech repairs and second speaker interrup tions are important pieces of information that a parser needs to accommodate The second section details how our parser handles these phenomena described in more detail in Core Schubert and the third section investigates assumptions made by the parser The corpus of data used in this work is the TRAINS dialogs Heeman Allen a collection of human human problem solving dialogs in a railway transportation domain In these two speaker dialogs one speaker is given a set of delivery goals to achieve the other speaker acts as a problem solving assistant responsible for carrying out the plan Examples from the TRAINS domain will be used throughout the paper