Speech/Text Information Retrieval with Speech Queries for Mandarin Chinese

• Main Authors: Bai, Bo-Ren, 白博仁
• Other Authors: Lin-Shan Lee
• Format: Others
• Language: en_US
• Published: 1998
• Online Access: http://ndltd.ncl.edu.tw/handle/77177297346059554563

博士 === 國立臺灣大學 === 電機工程學系 === 86 === Intelligent and efficient information retrieval techniques allowing easy access to huge amount and various types of information become highly desired and have been extensively studied in recent years. This dissertation deals with the problem of Chinese text and Mandarin speech information retrieval with Mandarin speech queries. To achieve this purpose, key technologies including text keyword extraction and speech keyword spotting are first developed, then approaches integrating the syllable and keyword information for text and speech information retrieval using speech queries are proposed considering relevant technologies such as text processing, speech recognition, information retrieval, and so on. Finally, a prototype system with an interface supporting some user-friendly functions is successfully developed to demonstrate the feasibility of the technologies developed in this dissertation. An efficient text keyword extraction approach is first developed based on the PAT tree data structure to handle dynamic text collection. In the approach, all patterns in a text collection are first taken as possible keyword candidates, then for each keyword candidate, both the association relationships of sub-patterns inside the keyword candidate and the contextual dependency of the keyword candidate with its adjacent patterns on both sides will be examined, thus very good performance can be achieved. On the other hand, a syllable-based speech keyword spotting approach is developed to handle keyword sets with large and flexible vocabularies. The approach is implemented by a multi-phase framework with score normalization and speedup techniques, thus it is very accuracy and fast. With the text keyword extraction and speech keyword spotting techniques developed successfully, by adopting conventional continuous speech recognition techniques for syllable recognition, approaches integrating the syllable and word information for text and speech information retrieval using speech queries are then developed. It has been found that by properly integrating different information considering relevant speech recognition and information retrieval technologies, better retrieving performance can be achieved. On the other hand, techniques such as relevance feedback are also applied to support interactive retrieving. All the above technologies are finally integrated into a prototype system for text and speech information retrieval using speech queries. The system includes an interface supporting some user-friendly functions to play the voice records, display and read the text records, extract keywords from retrieved text records for term suggestion, etc. Also, relevance feedback functions are supported, thus the user can select either voice records, text records, and/or their keywords and feed them back to the system for further retrieving both text and voice records. The prototype system not only verifies the feasibility of the technologies developed in this dissertation, but also provides a very good environment for further investigation on a variety of research topics on information retrieval.