Detecting Structural Metadata with Decision Trees and Transformation-Based Learning

Detecting Structural Metadata with Decision Trees and Transformation-Based Learning Joungbum Kim† and Sarah E. Schwarm‡ and Mari Ostendorf† †Dept. of Electrical Engineering ‡Dept. of Computer Science University of Washington Seattle, WA 98195. USA {bummie,sarahs,mo}@ssli.ee.washington.edu Abstract The regular occurrence of disfluencies is a distinguishing characteristic of spontaneous speech. Detecting and removing such disflu- encies can substantially improve the usefulness of spontaneous speech transcripts. This pa- per presents a system that detects various types of disfluencies and other structural information with cues obtained from lexical and prosodic information sources. Specifically, combina- tions of decision trees and language models are used to predict sentence ends and interruption points and, given these events, transformation- based learning is used to detect edit disfluen- cies and conversational fillers. Results are re- ported on human and automatic transcripts of conversational telephone speech. 1 Introduction Automatic speech-to-text (STT) transcripts of sponta- neous speech are often difficult to comprehend even with- out the challenges arising from word recognition errors introduced by imperfect STT systems (Jones et al., 2003). Such transcripts lack punctuation that indicates clausal or sentential boundaries, and they contain a number of dis- fluencies that would not normally occur in written lan- guage. Repeated words, hesitations such as “um” and “uh”, and corrections to a sentence in mid-stream are a normal part of conversational speech. These disflu- encies are handled easily by human listeners (Shriberg, 1994), but their existence makes transcripts of sponta- neous speech ill-suited for most natural language pro- cessing (NLP) systems developed for text, such as parsers or information extraction systems. Similarly, the lack of meaningful segmentation in automatically generated speech transcripts makes them problematic to use in NLP systems, most of which are designed to work at the sen- tence level. Detecting and removing disfluencies and lo- cating sentential unit boundaries in spontaneous speech transcripts can improve their readability and make them more suitable for NLP. Automatically annotating dis- course markers and other conversational fillers is also likely to be useful, since proper handling is needed to fol- low the flow of conversation. Hence, the overall goal of our work is to detect such structural information in con- versational speech using features generated by currently available speech processing systems and statistical ma- chine learning tools. This paper is organized as follows. In Section 2, we describe the types of metadata that this work addresses, followed by a discussion of related prior work in Sec- tion 3. Section 4 describes the system architecture and details the algorithms and features used by our system. Section 5 discusses the experimental paradigm and re- sults. Finally we provide a summary and directions for future work in Section 6. 2 Structural Metadata We consider three main types of structural metadata: sentence-like units, conversational fillers and edit disflu- encies. These structures were chosen primarily because of the availability of annotated conversational speech data from the Linguistic Data Consortium (Strassel, 2003) and standard scoring tools (NIST, 2003). 2.1 Sentence Units Conversational speech lacks the clear sentence bound- aries of written …

Content truncated to 500 words