Punctuation Prediction for Unsegmented Transcript Based on Word Vector

Punctuation Prediction for Unsegmented Transcript Based on Word Vector Xiaoyin Che, Cheng Wang, Haojin Yang, Christoph Meinel Hasso Plattner Institute Prof.-Dr.-Helmert-Str. 2-3, 14482 Potsdam, Germany xiaoyin.che@hpi.de, cheng.wang@hpi.de, haojin.yang@hpi.de, christoph.meinel@hpi.de Abstract In this paper we propose an approach to predict punctuation marks for unsegmented speech transcript. The approach is purely lexical, with pre-trained Word Vectors as the only input. A training model of Deep Neural Network (DNN) or Convolutional Neural Network (CNN) is applied to classify whether a punctuation mark should be inserted after the third word of a 5-words sequence and which kind of punctuation mark the inserted one should be. TED talks within IWSLT dataset are used in both training and evaluation phases. The proposed approach shows its effectiveness by achieving better result than the state-of-the-art lexical solution which works with same type of data, especially when predicting puncuation position only. Keywords: Punctuation Prediction, Word Vector, Deep Learning, Convolutional Neural Network 1. Introduction Generally the text transcript generated by automatic speech recognition (ASR) systems is unpunctuated and unseg- mented. However, the readability of the transcript can be largely improved by the existence of punctuation marks, and the segmentation of the transcript based on punctuation positions will also increase the efficiency of many down- stream natural languages processing (NLP) tasks, such as semantic parsing, question answering or machine transla- tion(Matusov et al., 2007; Wang et al., 2010). It may also act as effective pre-processing to facilitate some other tasks like subtitling(Tiedemann, 2007). Therefore many efforts, some of which may focus on sim- iliar topic “sentence boundary detection”, have been made to restore or predict the punctuation marks. Most of the re- cent works can be divided into three categories: prosodic or acoustic feature based (Levy et al., 2012; Xie et al., 2012; Sinclair et al., 2014), lexical feature based (Gravano et al., 2009; Lu and Ng, 2010; Ueffing et al., 2013) and the hybrid of previous two (Wang et al., 2012; Xu et al., 2014; Hasan et al., 2014). In this paper, we focus on the lexical aspect only. A pure lexical model is understandably less powerful than a hybrid model, or even a pure acoustic model. But it is able to cope with the situation either with or without the audio files. Another fact is that the training data of a hy- brid model, which take both lexical and acoustic features as input, must be the standardized ASR transcripts, but a pure lexical model can take any kind of textual material into consideration, which is almost endless and extremely easy to get. However, the posterior probabilities of the in- dependent lexical model can be freely used in late fusion with acoustic features(Tilk and Alum¨ae, 2015; Khomitse- vich et al., 2015). Considering all these facts, we believe a good lexical model is widely applicable. In previous research efforts, frequently used lexical features include the language model (LM) score, token, part-of- speech (POS) tag, chunk tag and so on. However, we would like attempt another possibility by using the Word Vector. …

Content truncated to 500 words