Disfluency Detection Using Multi-step Stacked Learning

Proceedings of NAACL-HLT 2013, pages 820–825, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Disfluency Detection Using Multi-step Stacked Learning Xian Qian and Yang Liu Computer Science Department The University of Texas at Dallas {qx,yangl}@hlt.utdallas.edu Abstract In this paper, we propose a multi-step stacked learning model for disfluency detection. Our method incorporates refined n-gram features step by step from different word sequences. First, we detect filler words. Second, edited words are detected using n-gram features ex- tracted from both the original text and filler fil- tered text. In the third step, additional n-gram features are extracted from edit removed texts together with our newly induced in-between features to improve edited word detection. We use Max-Margin Markov Networks (M3Ns) as the classifier with the weighted hamming loss to balance precision and recall. Experiments on the Switchboard corpus show that the re- fined n-gram features from multiple steps and M3Ns with weighted hamming loss can signif- icantly improve the performance. Our method for disfluency detection achieves the best re- ported F-score 0.841 without the use of addi- tional resources.1 1 Introduction Detecting disfluencies in spontaneous speech can be used to clean up speech transcripts, which help- s improve readability of the transcripts and make it easy for downstream language processing modules. There are two types of disfluencies: filler words in- cluding filled pauses (e.g., ‘uh’, ‘um’) and discourse markers (e.g., ‘I mean’, ‘you know’), and edited words that are repeated, discarded, or corrected by 1Our source code is available at http://code.google.com/p/disfluency-detection/downloads/list the following words. An example is shown below that includes edited words and filler words. I want a flight to Boston � �� edited� �uh I mean �� � filler to Denver Automatic filler word detection is much more ac- curate than edit detection as they are often fixed phrases (e.g., “uh”, “you know”, “I mean”), hence our work focuses on edited word detection. Many models have been evaluated for this task. Liu et al. (2006) used Conditional Random Fields (CRFs) for sentence boundary and edited word de- tection. They showed that CRFs significantly out- performed Maximum Entropy models and HMM- s. Johnson and Charniak (2004) proposed a TAG- based noisy channel model which showed great im- provement over boosting based classifier (Charniak and Johnson, 2001). Zwarts and Johnson (2011) extended this model using minimal expected F-loss oriented n-best reranking. They obtained the best re- ported F-score of 83.8% on the Switchboard corpus. Georgila (2009) presented a post-processing method during testing based on Integer Linear Programming (ILP) to incorporate local and global constraints. From the view of features, in addition to tex- tual information, prosodic features extracted from speech have been incorporated to detect edited words in some previous work (Kahn et al., 2005; Zhang et al., 2006; Liu et al., 2006). Zwarts and Johnson (2011) trained an extra language model on additional corpora, and used output log probabili- ties of language models as features in the reranking stage. They reported that the language model gained about absolute 3% …

Content truncated to 500 words