Joint Transition-based Dependency Parsing and Disfluency Detection for Automatic Speech Recognition Texts

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1036–1041, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Joint Transition-based Dependency Parsing and Disfluency Detection for Automatic Speech Recognition Texts Masashi Yoshikawa and Hiroyuki Shindo and Yuji Matsumoto Graduate School of Information and Science Nara Institute of Science and Technology 8916-5, Takayama, Ikoma, Nara, 630-0192, Japan { masashi.yoshikawa.yh8, shindo, matsu }@is.naist.jp Abstract Joint dependency parsing with disfluency de- tection is an important task in speech lan- guage processing. Recent methods show high performance for this task, although most au- thors make the unrealistic assumption that in- put texts are transcribed by human annota- tors. In real-world applications, the input text is typically the output of an automatic speech recognition (ASR) system, which implies that the text contains not only disfluency noises but also recognition errors from the ASR system. In this work, we propose a parsing method that handles both disfluency and ASR errors us- ing an incremental shift-reduce algorithm with several novel features suited to ASR output texts. Because the gold dependency informa- tion is usually annotated only on transcribed texts, we also introduce an alignment-based method for transferring the gold dependency annotation to the ASR output texts to con- struct training data for our parser. We con- ducted an experiment on the Switchboard cor- pus and show that our method outperforms conventional methods in terms of dependency parsing and disfluency detection. 1 Introduction Spontaneous speech is different from written text in many ways, one of which is that it contains disfluen- cies, that is, parts of the utterance that are corrected by the speaker during the utterance. NLP system performance is reported to deteriorate when there are disfluencies, for example, with SMT (Cho et al., 2014). Therefore, it is desirable to preprocess the speech before passing it to other NLP tasks. There are a number of studies that address the problem of detecting disfluencies. Some of these studies include dependency parsing (Honnibal and Johnson, 2014; Wu et al., 2015; Rasooli and Tetreault, 2014), whereas others are dedicated sys- tems (Qian and Liu, 2013; Ferguson et al., 2015; Hough and Purver, 2014; Hough and Schlangen, 2015; Liu et al., 2003). Among these studies, Hon- nibal (2014) and Wu (2015) address this problem by adding a new action to transition-based dependency parsing that removes the disfluent parts of the in- put sentence from the stack. Using this approach, they achieved high performance in terms of both de- pendency parsing and disfluency detection on the Switchboard corpus. However, the authors assume that the input texts to parse are transcribed by human annotators, which, in practice, is unrealistic. In real-world applications, in addition to disfluencies, the input texts contain ASR errors; these issues might degrade the parsing performance. For example, proper nouns that are not contained in the ASR system vocabulary may break up into smaller pieces, yielding a difficult problem for the parsing unit (Cheng et al., 2015): REF: what can we get at Litanfeeth …

Content truncated to 500 words