Retrieval of broadcast news documents with the THISL system

Abstract

This paper describes a spoken document retrieval system, combining the ABBOT large vocabulary continuous speech recognition (LVCSR) system developed by Cambridge University, Sheffield University and Softsound, and the PRISE information retrieval engine developed by NIST. The system was constructed to enable us to participate in the TREC 6 Spoken Document Retrieval experimental evaluation. Our key aims in this work were to produce a complete system for the SDR task, to investigate the effect of a word error rate of 30-50% on retrieval performance and to investigate the integration of LVCSR and word spotting in a retrieval task.

Advances in Information Retrieval

“…A form of pseudo relevance feedback (known as local context analysis) was used to expand topic texts with additional terms taken from the recognized transcript collection. 6. shef-s1, University of Sheffield with collaborators at Cambridge University (Abberley et al, 1998). Recognition was performed using the Abbot recognizer system with a vocabulary of 65,532 words producing a transcript with a 35.9% WER.…”

The CLEF 2003 Cross-Language Spoken Document Retrieval Track

“…This formulation required that sites perform some automatic segmentation of the full broadcasts into smaller units suitable for retrieval. Using an approach inspired by [6], we performed story segmentation as follows. First we created 30 second segments based on the word recognition time stamps using a 10 second step to create overlapping segment windows.…”

Speech and Hand Transcribed Retrieval

“…Although a great deal of research has been conducted in the retrieval of corrupted data, be it scanned text [Jones01], speech ([Garofolo99b], [Abberley98]), or translated foreign language documents [Franz98], relatively little work has investigated the notion of varying levels of corruption in the collection(s) being retrieved. This is perhaps surprising, as it is quite reasonable to expect such variation.…”

“…Spoken document retrieval (SDR) is being used: internally within corporations ([Abberley98], [Renals99], [THISL01]), and as a Web search engine ([SpeechB00A], [SpeechB00B], [SpeechB01]). Word error rates (WER) on the audio documents being retrieved are still relatively high, however, because relevant documents generally contain each query term with a high term frequency (tf), as long as a few of the term occurrences are recognised, relevant documents will be retrieved.…”