Using Conditional Random Fields for Sentence Boundary Detection in Speech

Proceedings of the 43rd Annual Meeting of the ACL, pages 451–458, Ann Arbor, June 2005. c⃝2005 Association for Computational Linguistics Using Conditional Random Fields For Sentence Boundary Detection In Speech Yang Liu ICSI, Berkeley yangl@icsi.berkeley.edu Andreas Stolcke Elizabeth Shriberg SRI and ICSI stolcke,ees@speech.sri.com Mary Harper Purdue University harper@ecn.purdue.edu Abstract Sentence boundary detection in speech is important for enriching speech recogni- tion output, making it easier for humans to read and downstream modules to process. In previous work, we have developed hid- den Markov model (HMM) and maximum entropy (Maxent) classifiers that integrate textual and prosodic knowledge sources for detecting sentence boundaries. In this paper, we evaluate the use of a condi- tional random field (CRF) for this task and relate results with this model to our prior work. We evaluate across two cor- pora (conversational telephone speech and broadcast news speech) on both human transcriptions and speech recognition out- put. In general, our CRF model yields a lower error rate than the HMM and Max- ent models on the NIST sentence bound- ary detection task in speech, although it is interesting to note that the best results are achieved by three-way voting among the classifiers. This probably occurs be- cause each model has different strengths and weaknesses for modeling the knowl- edge sources. 1 Introduction Standard speech recognizers output an unstructured stream of words, in which the important structural features such as sentence boundaries are missing. Sentence segmentation information is crucial and as- sumed in most of the further processing steps that one would want to apply to such output: tagging and parsing, information extraction, summarization, among others. 1.1 Sentence Segmentation Using HMM Most prior work on sentence segmentation (Shriberg et al., 2000; Gotoh and Renals, 2000; Christensen et al., 2001; Kim and Woodland, 2001; NIST- RT03F, 2003) have used an HMM approach, in which the word/tag sequences are modeled by N- gram language models (LMs) (Stolcke and Shriberg, 1996). Additional features (mostly related to speech prosody) are modeled as observation likelihoods at- tached to the N-gram states of the HMM (Shriberg et al., 2000). Figure 1 shows the graphical model representation of the variables involved in the HMM for this task. Note that the words appear in both the states1 and the observations, such that the word stream constrains the possible hidden states to matching words; the ambiguity in the task stems entirely from the choice of events. This architec- ture differs from the one typically used for sequence tagging (e.g., part-of-speech tagging), in which the “hidden” states represent only the events or tags. Empirical investigations have shown that omitting words in the states significantly degrades system performance for sentence boundary detection (Liu, 2004). The observation probabilities in the HMM, implemented using a decision tree classifier, capture the probabilities of generating the prosodic features 1In this sense, the states are only partially “hidden”. 451 P (F i jE i ; W i ).2 An N-gram LM is used to calculate the transition probabilities: P (W i E i jW …

Content truncated to 500 words