LARD: Large-scale Artificial Disfluency Generation

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 2327–2336 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 2327 LARD: Large-scale Artificial Disfluency Generation Tatiana Passali∗†, Thanassis Mavropoulos∗, Grigorios Tsoumakas†, Georgios Meditskos∗†, Stefanos Vrochidis∗ ∗Centre for Research and Technology Hellas, Thessaloniki, Greece {tpassali, mavrathan, gmeditsk, stefanos}@iti.gr †School of Informatics, Aristotle University of Thessaloniki, Thessaloniki, Greece {scpassali, greg, gmeditsk}@csd.auth.gr Abstract Disfluency detection is a critical task in real-time dialogue systems. However, despite its importance, it remains a relatively unexplored field, mainly due to the lack of appropriate datasets. At the same time, existing datasets suffer from various issues, including class imbalance issues, which can significantly affect the performance of the model on rare classes, as it is demonstrated in this paper. To this end, we propose LARD, a method for generating complex and realistic artificial disfluencies with little effort. The proposed method can handle three of the most common types of disfluencies: repetitions, replacements, and restarts. In addition, we release a new large-scale dataset with disfluencies that can be used on four different tasks: disfluency detection, classification, extraction, and correction. Experimental results on the LARD dataset demonstrate that the data produced by the proposed method can be effectively used for detecting and removing disfluencies, while also addressing limitations of existing datasets. Keywords: disfluency detection, synthetic dataset, data augmentation 1. Introduction Virtual assistants and spoken dialogue systems are in- creasingly used in many applications, such as auto- mated customer service (Gnewuch et al., 2017; Shee- han et al., 2020), education (Clarizia et al., 2018; Heryandi, 2020) and healthcare (Tian et al., 2019; Sez- gin et al., 2020). Recent systems, which are powered by powerful Deep Learning (DL) models (Golovanov et al., 2019; Wu et al., 2019; Majumder et al., 2020; Zhong et al., 2020; Zhao et al., 2020), led to signifi- cantly more accurate systems compared to previous ap- proaches. However, despite the considerable progress in this area, many difficult problems still remain unre- solved. One of them is the fluidity of spoken language, which often includes various errors and self-corrections made on the fly, rendering the deployment of such sys- tems in the wild especially challenging. Indeed, when humans recognize that they have made an error in their speech, they automatically attempt to cor- rect it by editing, reformulating, or completely restart- ing themselves. This is part of the natural and intuitive process that happens subconsciously during a sponta- neous human dialogue flow. These interruptions and self-corrections of a speaker’s utterances are called dis- fluencies (Shriberg, 1994) and are a common trait of spontaneous speech. There is a large number of stud- ies about the structure of disfluencies and their influ- ence on the conversation between humans in theoreti- cal level (Sparks, 1994; Shriberg, 1994; Plejert, 2004; Colman and Healey, 2011; Emrani and Hooshmand, 2019) which led to the development of various practical applications and models for disfluency detection (Zay- ats et al., 2016; Wang et al., 2016; Wang et al., 2020; …

Content truncated to 500 words