Automated speech-unit delimitation in spoken learner English

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 782–793, Osaka, Japan, December 11-17 2016. Automated speech-unit delimitation in spoken learner English Russell Moore1 Andrew Caines1 Calbert Graham1 Paula Buttery2 Automated Language Teaching & Assessment Institute 1Department of Theoretical & Applied Linguistics 2Computer Laboratory University of Cambridge, Cambridge, U.K. rjm49,apc38,crg29,pjb48@cam.ac.uk Abstract In order to apply computational linguistic analyses and pass information to downstream appli- cations, transcriptions of speech obtained via automatic speech recognition (ASR) need to be divided into smaller meaningful units, in a task we refer to as ‘speech-unit (SU) delimitation’. We closely recreate the automatic delimitation system described by Lee and Glass (2012), ‘Sen- tence detection using multiple annotations’, Proceedings of INTERSPEECH, which combines a prosodic model, language model and speech-unit length model in log-linear fashion. Since state- of-the-art natural language processing (NLP) tools have been developed to deal with written text and its characteristic sentence-like units, SU delimitation helps bridge the gap between ASR and NLP, by normalising spoken data into a more canonical format. Previous work has focused on native speaker recordings; we test the system of Lee and Glass (2012) on non-native speaker (or ‘learner’) data, achieving performance above the state-of-the-art. We also consider alternative evaluation metrics which move away from the idea of a single ‘truth’ in SU delimitation, and frame this work in the context of downstream NLP applications. 1 Introduction By convention, texts written using the Latin alphabet are normally subdivided into smaller units – sen- tences – by capitalised initial characters and closing full-stops (periods), question marks and exclamation marks. The sentence has in turn become an orthodox unit of analysis for much linguistic research, from natural language processing to syntactic theory. Speech, meanwhile, is hardly ever so neatly portioned up. Pauses and turn-taking (in conversation) may at first appear to correspond to sentence-marking ortho- graphic devices, and often they do delimit sensible language chunks, but not always. Speakers pause in strange places, make false starts, leave thoughts unfinished, and interrupt or overlap each other (Sacks et al., 1974; Dingemanse and Floyd, 2014; Carter and McCarthy, in press). These characteristic features of spontaneous speech pose a problem for researchers investigating monologues or dialogues in naturalistic scenarios: what is a sentence-like unit of spoken language? This question has been addressed by conversation analysts, acquisition researchers assessing perfor- mance accuracy, and compilers of large corpora, among others. Common practice is to transform speech recordings into written transcripts in order to work further with the data, whether with manual or auto- mated means. Now comes the dilemma of how to subdivide those transcripts, where appropriate, into chunks one can work with. For conversational data the first division is made by speakers’ turns into what are known as ‘utterances’. But then how should those utterances be further subdivided (if at all) into something akin to sentences? The consensus, theoretically-speaking, has been to identify smaller, sentence-like units of speech on syntactic and/or semantic grounds (with more emphasis on the former), often …

Content truncated to 500 words