Generating Fluent Translations from Disfluent Text Without Access to Fluent References: IIT Bombay@IWSLT2020

Proceedings of the 17th International Conference on Spoken Language Translation (IWSLT), pages 178–186 July 9-10, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 178 Generating Fluent Translations from Disfluent Text Without Access to Fluent References: IIT Bombay@IWSLT2020 Nikhil Saini, Jyotsana Khatri, Preethi Jyothi, Pushpak Bhattacharyya Department of Computer Science and Engineering Indian Institute of Technology Bombay Mumbai, India {nikhilra, jyotsanak, pjyothi, pb}@cse.iitb.ac.in Abstract Machine translation systems perform reason- ably well when the input is well-formed speech or text. Conversational speech is spon- taneous and inherently consists of many dis- fluencies. Producing fluent translations of dis- fluent source text would typically require par- allel disfluent to fluent training data. How- ever, fluent translations of spontaneous speech are an additional resource that is tedious to obtain. This work describes the submission of IIT Bombay to the Conversational Speech Translation challenge at IWSLT 2020. We specifically tackle the problem of disfluency removal in disfluent-to-fluent text-to-text trans- lation assuming no access to fluent references during training. Common patterns of disflu- ency are extracted from disfluent references and a noise induction model is used to sim- ulate them starting from a clean monolingual corpus. This synthetically constructed dataset is then considered as a proxy for labeled data during training. We also make use of addi- tional fluent text in the target language to help generate fluent translations. This work uses no fluent references during training and beats a baseline model by a margin of 4.21 and 3.11 BLEU points where the baseline uses disfluent and fluent references, respectively. Index Terms- disfluency removal, machine translation, noise induction, leveraging mono- lingual data, denoising for disfluency removal. 1 Introduction and Related Work Spoken language translation often suffers due to the presence of disfluencies. In conversational speech, speakers often use disfluencies such as filler words, repetitions of fillers, repetitions of fluent phrases, false starts, and corrections which do not occur in the text. Standard machine translation and spoken translation systems perform competitively when the input is well-formed text or rehearsed speech as in TED talks or broadcast news (Cho et al., 2014; Wang et al., 2010; Honal and Schultz, 2005; Za- yats et al., 2016). With the increasing popularity of end-to-end speech translation systems (Weiss et al., 2017; Bansal et al., 2018), one may not want disfluency removal to be treated as an intermediate step between ASR and MT. It might be more desir- able for disfluency removal to be handled within the model itself, or as a separate post-processing step. To produce fluent translations from disfluent text, one would typically require access to disfluent speech (or text) and its corresponding fluent trans- lations during training. While some corpora with labeled disfluencies exist (Cho et al., 2014; Burger et al., 2002), only subsets have been translated and/or released. (Salesky et al., 2018) introduced a set of fluent references for the Fisher Spanish- English conversational speech corpus (David Graff and Cieri.). This has enabled a new task of end- to-end training and evaluation on fluent references. (Salesky et al., 2019) reports …

Content truncated to 500 words