Integrating sentence- and word-level error identification for disfluency correction

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 765–774, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Integrating sentence- and word-level error identification for disfluency correction Erin Fitzgerald Johns Hopkins University Baltimore, MD, USA erinf@jhu.edu Frederick Jelinek Johns Hopkins University Baltimore, MD, USA jelinek@jhu.edu Keith Hall Google Inc. Z¨urich, Switzerland kbhall@google.com Abstract While speaking spontaneously, speakers often make errors such as self-correction or false starts which interfere with the successful application of natural language processing techniques like summarization and machine translation to this data. There is active work on reconstructing this error- ful data into a clean and fluent transcript by identifying and removing these simple errors. Previous research has approximated the potential benefit of conducting word-level reconstruction of simple errors only on those sentences known to have errors. In this work, we explore new approaches for automatically identifying speaker con- struction errors on the utterance level, and quantify the impact that this initial step has on word- and sentence-level reconstruc- tion accuracy. 1 Introduction A system would accomplish reconstruction of its spontaneous speech input if its output were to rep- resent, in flawless, fluent, and content-preserving text, the message that the speaker intended to con- vey. While full speech reconstruction would likely require a range of string transformations and po- tentially deep syntactic and semantic analysis of the errorful text (Fitzgerald, 2009), in this work we will attempt only to resolve less complex errors, correctable by deletion alone, in a given manually- transcribed utterance. The benefit of conducting word-level recon- struction of simple errors only on those sen- tences known to have errors was approximated in (Fitzgerald et al., 2009). In the current work, we explore approaches for automatically identifying speaker-generated errors on the utterance level, and calculate the gain in accuracy that this initial step has on word- and sentence-level accuracy. 1.1 Error classes in spontaneous speech Common simple disfluencies in sentence-like ut- terances (SUs) include filler words (i.e., “um”, “ah”, and discourse markers like “you know”), as well as speaker edits consisting of a reparandum, an inter- ruption point (IP), an optional interregnum (like “I mean”), and a repair region (Shriberg, 1994), as seen in Figure 1. [that′s] � �� � reparandum IP ���� + {uh} ���� interregnum that′s � �� � repair a relief Figure 1: Typical edit region structure. These reparanda, or edit regions, can be classified into three main groups: 1. In a repetition (above), the repair phrase is approximately identical to the reparandum. 2. In a revision, the repair phrase alters reparan- dum words to correct the previously stated thought. EX1: but [when he] + {i mean} when she put it that way EX2: it helps people [that are going to quit] + that would be quitting anyway 3. In a restart fragment an utterance is aborted and then restarted with a new train of thought. EX3: and [i think he’s] + he tells me he’s glad he has one of those EX4: [amazon was incorporated by] {uh} …

Content truncated to 500 words