A Comparative Study of Methods for Topic Modeling in Spoken Document Retrieval

Computational Linguistics and Chinese Language Processing Vol. 17, No. 1, March 2012, pp. 65-86 65 © The Association for Computational Linguistics and Chinese Language Processing A Comparative Study of Methods for Topic Modeling in Spoken Document Retrieval Shih-Hsiang Lin+,∗ and Berlin Chen#,∗ Abstract Topic modeling for information retrieval (IR) has attracted significant attention and demonstrated good performance in a wide variety of tasks over the years. In this paper, we first present a comprehensive comparison of various topic modeling approaches, including the so-called document topic models (DTM) and word topic models (WTM), for Chinese spoken document retrieval (SDR). Moreover, different granularities of index features, including words, subword units, and their combinations, are also exploited to work in conjunction with various extensions of topic modeling presented in this paper, so as to alleviate SDR performance degradation caused by speech recognition errors. All of the experiments were performed on the TDT Chinese collection. Keywords: Information Retrieval, Document Topic Models, Word Topic Models, Spoken Document Retrieval. 1. Introduction Due to the advances in computer technology and the proliferation of Internet activity, huge volumes of multimedia data, such as text files, broadcast radio and television programs, lectures, and digital archives, are continuously growing and filling networks. Development of intelligent and efficient information retrieval techniques to provide people with easy access to all kinds of information is now becoming more and more emphasized. Meanwhile, with the rapid evolution of speech recognition technology, substantial efforts and very encouraging results on spoken document retrieval (SDR) also have been demonstrated in the recent past. Although most retrieval systems participating in the TREC-SDR evaluations claimed that speech recognition errors do not seem to cause much adverse effect on SDR performance + Voice Division Research Center, Delta Electronics E-mail: shlin@csie.ntnu.edu.tw # Department of Computer Science & Information Engineering, National Taiwan Normal University E-mail: berlin@csie.ntnu.edu.tw ∗ Corresponding authors. 66 Shih-Hsiang Lin and Berlin Chen when merely using imperfect recognition transcripts derived from one-best recognition results from a speech recognizer (Garofolo et al., 2000; Chelba et al., 2008), this is probably attributed to the fact that the TREC-style test queries tend to be quite long and contain different words describing similar concepts that can help the queries match their relevant spoken documents. Furthermore, a query word (or phrase) may occur repeatedly (more than once) within a relevant spoken document, and it is not always the case that all of the occurrences of the word would be misrecognized totally as other words. We, however, believe that SDR would still present a challenge in situations where the queries are relatively short and there exists severe deviation in word usage between the queries and spoken documents. Among several promising information retrieval approaches, statistical language modeling (LM) (Ponte & Croft, 1998), aiming to capture the regularity in human natural language and quantify the acceptability of a given word sequence, has continuously been a focus of active research in the last decade (Miller et al., 1999; Hofmann, 2001). The basic idea is that each individual document in …

Content truncated to 500 words