Semantic Parsing of Disfluent Speech

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 1748–1753 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 1748 Semantic Parsing of Disfluent Speech Priyanka Sen Amazon Alexa AI Cambridge, UK sepriyan@amazon.com Isabel Groves Amazon Alexa AI Cambridge, UK isabeg@amazon.com Abstract Speech disfluencies are prevalent in sponta- neous speech. The rising popularity of voice assistants presents a growing need to han- dle naturally occurring disfluencies. Semantic parsing is a key component for understanding user utterances in voice assistants, yet most semantic parsing research to date focuses on written text. In this paper, we investigate se- mantic parsing of disfluent speech with the ATIS dataset. We find that a state-of-the-art se- mantic parser does not seamlessly handle dis- fluencies. We experiment with adding real and synthetic disfluencies at training time and find that adding synthetic disfluencies not only im- proves model performance by up to 39% but can also outperform adding real disfluencies in the ATIS dataset. 1 Introduction Spoken language differs from written language. Unlike written texts, spontaneous speech frequently contains disfluencies such as filled pauses, rep- etitions, repairs, and false starts. These affect around 6% of words (Kasl and Mahl, 1965; Tree, 1995; Bortfeld et al., 2001) and occur in both human-human and human-computer interactions (Oviatt, 1995). While considerable research has been done on speech disfluencies in syntactic pars- ing (Johnson and Charniak, 2004; Wang et al., 2018; Jamshid Lou et al., 2019), semantic parsing of disfluent speech has received less attention. Semantic parsing is the task of mapping written text to a representation of its meaning. In voice as- sistants, such as Amazon Alexa, Google Assistant, or Siri, an automated speech recognition (ASR) component is often followed by a semantic parsing task that identifies the intent and slots of the tran- scribed utterance. In voice-based contexts, speech disfluencies are to be expected, and so a voice as- sistant must accurately support disfluent speech to support more natural user interaction. Despite the growing popularity of voice assistants, semantic parsing research to date has focused on written lan- guage. Most popular semantic parsing datasets only include written forms that were generated from grammars (Coucke et al., 2018) or crowd- sourced (Gupta et al., 2018) and do not contain speech phenomena such as disfluencies. One approach to parsing disfluent speech is to detect and remove transcribed disfluencies in a post-processing step after ASR with a disfluency detector (Zayats et al., 2016; Jamshid Lou et al., 2018). Alternatively, some ASR models attempt to directly produce fluent transcriptions from dis- fluent inputs (Jamshid Lou and Johnson, 2020) but have lower performance. We investigate seman- tic parsing of disfluent utterances directly, without ASR post-processing, by training a single semantic parsing model that handles both fluent and disfluent utterances. Our approach has the advantage of be- ing simple and easier to implement than approaches that require multiple models in a pipeline. We investigate semantic parsing of disfluent speech using the ATIS semantic parsing dataset (Price, 1990), …

Content truncated to 500 words