News spoken document retrieval by considering out-of-vocabulary keywords

摘要

This paper describes a Japanese spoken document retrieval system that is robust for documents including Out-of-Vocabulary (OOV) words. One approach to spoken document retrieval is to automatically transcibe spoken documents into word sequences, which can be directory matched against queries. In this approach, however, a serious problem arises when both the queries and the documents include OOV keywords, where matching through OOV keywords always fails, because the OOV keyword can not be transcribed as a word. To avoid this problem, we use both word based indexing for in-vocabulary keywords and syllable based indexing for OOV keywords. Syllable based indexes are created from a transcript using a recognizer with OOV detection processing. Our approach of combining word based and syllable based indexing has advantages of both indexes. When a keyword included in a query is in-vocabulary words, this is matched directly against the word based index, where neither retrieval speed nor accuracy is harmed. When a keyword is OOV, a syllable form keyword is approximately matched against the syllable form index, or concatenated HMMs, which correspond to the syllable form keyword, are matched against spoken documents by a word spotting technique. Experiment results clearly claimed that the our approach is quite effective and robust in retrieving spoken documents by queries including OOV keywords.