Unsupervised Approaches for Automatic Keyword Extraction Using Meeting Transcripts

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 620–628, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Unsupervised Approaches for Automatic Keyword Extraction Using Meeting Transcripts Feifan Liu, Deana Pennell, Fei Liu and Yang Liu Computer Science Department The University of Texas at Dallas Richardson, TX 75080, USA {ffliu,deana,feiliu,yangl}@hlt.utdallas.edu Abstract This paper explores several unsupervised ap- proaches to automatic keyword extraction using meeting transcripts. In the TFIDF (term frequency, inverse document frequency) weighting framework, we incorporated part- of-speech (POS) information, word clustering, and sentence salience score. We also evalu- ated a graph-based approach that measures the importance of a word based on its connection with other sentences or words. The system performance is evaluated in different ways, in- cluding comparison to human annotated key- words using F-measure and a weighted score relative to the oracle system performance, as well as a novel alternative human evaluation. Our results have shown that the simple un- supervised TFIDF approach performs reason- ably well, and the additional information from POS and sentence score helps keyword ex- traction. However, the graph method is less effective for this domain. Experiments were also performed using speech recognition out- put and we observed degradation and different patterns compared to human transcripts. 1 Introduction Keywords in a document provide important infor- mation about the content of the document. They can help users search through information more effi- ciently or decide whether to read a document. They can also be used for a variety of language process- ing tasks such as text categorization and informa- tion retrieval. However, most documents do not provide keywords. This is especially true for spo- ken documents. Current speech recognition system performance has improved significantly, but there is no rich structural information such as topics and keywords in the transcriptions. Therefore, there is a need to automatically generate keywords for the large amount of written or spoken documents avail- able now. There have been many efforts toward keyword ex- traction for text domain. In contrast, there is less work on speech transcripts. In this paper we fo- cus on one speech genre — the multiparty meeting domain. Meeting speech is significantly different from written text and most other speech data. For example, there are typically multiple participants in a meeting, the discussion is not well organized, and the speech is spontaneous and contains disflu- encies and ill-formed sentences. It is thus ques- tionable whether we can adopt approaches that have been shown before to perform well in written text for automatic keyword extraction in meeting tran- scripts. In this paper, we evaluate several differ- ent keyword extraction algorithms using the tran- scripts of the ICSI meeting corpus. Starting from the simple TFIDF baseline, we introduce knowl- edge sources based on POS filtering, word cluster- ing, and sentence salience score. In addition, we also investigate a graph-based algorithm in order to leverage more global information and reinforcement from summary sentences. We used different per- formance measurements: comparing …

Content truncated to 500 words