Punctuation Prediction with Transition-based Parsing

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 752–760, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Punctuation Prediction with Transition-based Parsing Dongdong Zhang1, Shuangzhi Wu2, Nan Yang3, Mu Li1 1Microsoft Research Asia, Beijing, China 2Harbin Institute of Technology, Harbin, China 3University of Science and Technology of China, Hefei, China {dozhang,v-shuawu,v-nayang,muli}@microsoft.com Abstract Punctuations are not available in automatic speech recognition outputs, which could cre- ate barriers to many subsequent text pro- cessing tasks. This paper proposes a novel method to predict punctuation symbols for the stream of words in transcribed speech texts. Our method jointly performs parsing and punctuation prediction by integrating a rich set of syntactic features when processing words from left to right. It can exploit a global view to capture long-range dependencies for punc- tuation prediction with linear complexity. The experimental results on the test data sets of IWSLT and TDT4 show that our method can achieve high-level performance in punctuation prediction over the stream of words in tran- scribed speech text. 1 Introduction Standard automatic speech recognizers output un- structured streams of words. They neither perform a proper segmentation of the output into sentences, nor predict punctuation symbols. The unavailable punctuations and sentence boundaries in tran- scribed speech texts create barriers to many sub- sequent processing tasks, such as summarization, information extraction, question answering and machine translation. Thus, the segmentation of long texts is necessary in many real applications. For example, in speech-to-speech translation, continuously transcribed speech texts need to be segmented before being fed into subsequent ma- chine translation systems (Takezawa et al., 1998; Nakamura, 2009). This is because current ma- chine translation (MT) systems perform the trans- lation at the sentence level, where various models used in MT are trained over segmented sentences and many algorithms inside MT have an exponen- tial complexity with regard to the length of inputs. The punctuation prediction problem has at- tracted research interest in both the speech pro- cessing community and the natural language pro- cessing community. Most previous work primar- ily exploits local features in their statistical mod- els such as lexicons, prosodic cues and hidden event language model (HELM) (Liu et al., 2005; Matusov et al., 2006; Huang and Zweig, 2002; Stolcke and Shriberg, 1996). The word-level mod- els integrating local features have narrow views about the input and could not achieve satisfied performance due to the limited context infor- mation access (Favre et al., 2008). Naturally, global contexts are required to model the punctu- ation prediction, especially for long-range de- pendencies. For instance, in English question sen- tences, the ending question mark is long-range de- pendent on the initial phrases (Lu and Ng, 2010), such as “could you” in Figure 1. There has been some work trying to incorporate syntactic features to broaden the view of hypotheses in the punctua- tion prediction models (Roark et al., 2006; Favre et al., 2008). In their methods, the punctuation prediction is treated as a separated post-procedure of parsing, which may suffer …

Content truncated to 500 words