Zero-shot Disfluency Detection for Indian Languages

Proceedings of the 29th International Conference on Computational Linguistics, pages 4442–4454 October 12–17, 2022. 4442 Zero-shot Disfluency Detection for Indian Languages Rohit Kundu, Preethi Jyothi and Pushpak Bhattacharyya Department of Computer Science and Engineering Indian Institute of Technology Bombay {rkundu, pjyothi, pb}@cse.iitb.ac.in Abstract Disfluencies that appear in the transcriptions from automatic speech recognition systems tend to impair the performance of downstream NLP tasks. Disfluency correction models can help alleviate this problem. However, the un- availability of labeled data in low-resource lan- guages impairs progress. We propose using a pretrained multilingual model, finetuned only on English disfluencies, for zero-shot disflu- ency detection in Indian languages. We present a detailed pipeline to synthetically generate dis- fluent text and create evaluation datasets for four Indian languages: Bengali, Hindi, Malay- alam, and Marathi. Even in the zero-shot set- ting, we obtain F1 scores of 75 and higher on five disfluency types across all four languages. We also show the utility of synthetically gener- ated disfluencies by evaluating on real disflu- ent text in Bengali, Hindi, and Marathi. Fine- tuning the multilingual model on additional synthetic Hindi disfluent text nearly doubles the number of exact matches and yields a 20- point boost in F1 scores when evaluated on real Hindi disfluent text, compared to training with only English disfluent text. 1 Introduction Disfluencies (e.g., filled pauses, repetitions, dis- course markers) are artefacts that are inherent to spontaneous or conversational speech. Disfluen- cies typically obey the following surface structure comprising: a reparandum, an interruption point (+) that marks the end of the reparandum, an in- terregnum, and finally the repair (Shriberg, 1994). The reparandum consists of one or more words that are not intended by the speaker and will be re- placed or ignored. The interregnum consists of an editing term indicating that the reparandum will be edited, or it may be empty, or it can contain fillers, discourse markers, etc. The repair section reflects the fluent part of the utterance. Words from the reparandum are repeated or corrected in the repair section, or a new chain of thought is started in case of a false start. Consider the following example that illustrates two disfluency types: {well} [i think + {you know} i think] the idea will work The words highlighted in red and green refer to dis- course marker and repetition disfluency types, re- spectively. The part in blue is the fluent version of the original sentence. The example also follows the standard annotation scheme: [reparandum + {interregnum} repair] Disfluencies in automatically transcribed text pose a major challenge for downstream NLP tasks such as machine translation, summarization, etc. (Rao et al., 2007; Wang et al., 2010). Disfluency de- tection/correction is often used as a preprocessing step for NLP, where the goal is to identify/remove the disfluent words (Shriberg et al., 1992). While disfluency correction has been extensively studied for English (Honal and Schultz, 2003; Zayats et al., 2014), it has received far less attention in other lan- guages. This is largely due to the lack …

Content truncated to 500 words