Experiments on Sentence Boundary Detection

This paper explores the problem of identifying sentence boundaries in the transcriptions produced by automatic speech recognition systems. An experiment which determines the level of human performance for this task is described as well as a memorybased computational approach to the problem. 1 T h e P r o b l e m This paper addresses the problem of identifying sentence boundaries in the transcriptions produced by automatic speech recognition (ASR) systems. This is unusual in the field of text processing which has generally dealt with well-punctuated text: some of the most commonly used texts in NLP are machine readable versions of highly edited documents such as newspaper articles or novels. However, there are many types of text which are not so-edited and the example which we concentrate on in this paper is the output from ASR systems. These differ from the sort of texts normally used in NLP in a number of ways; the text is generally in single case (usually upper), unpunctuated and may contain transcription errors. 1 Figure 1 compares a short text in the format which would be produced by an ASR system with a fully punctuated version which includes case information. For the remainder of this paper errorfree texts such as newspaper articles or novels shall be referred to as "standard text" and the output from a speech recognition system as "ASR text". There are many possible situations in which an NLP system may be required to process ASR text. The most obvious examples are NLP systems which take speech input (eg. Moore et al. (1997)). Also, dictation software programs do not punctuate or capitalise their output but, if this information could be added to ASR text, the results would be far more usable. One of the most important pieces of inform1 Speech recognition systems are often evaluated in terms of word error rate (WER), the percentage of tokens which are wrongly transcribed. For large vocabulary tasks and speakerindependent systems, WER varies between 7% and 50%, depending upon the quality of the recording being recognised. See, e.g., Cole (1996). G00D EVENING GIANNI VERSACE ONE OF THE WORLDS LEADING FASHION DESIGNERS HAS BEEN MURDERED IN MIAMI POLICE SAY IT WAS A PLANNED KILLING CARRIED OUT LIKE AN EXECUTION SCHOOLS INSPECTIONS ARE GOING TO BE TOUGHER TO FORCE BAD TEACHERS OUT AND THE FOUR THOUSAND COUPLES WH0 SHARED THE QUEENS GOLDEN DAY Good evening. Gi~nni Versace, one of the world's leading fashion designers, has been murdered in Miami. Police say it was a planned killing carried out like an execution. Schools inspections are going to be tougher to force bad teachers out. And the four thousand couples who shared the Queen's golden