Incremental Disfluency Detection for Spoken Learner English

Dialogue-based computer-assisted language learning (CALL) concerns the application and analysis of automated systems that engage with a language learner through dialogue. Routed in an interactionist perspective of second language acquisition, dialogue-based CALL systems assume the role of a speaking partner, providing learners the opportunity for spontaneous production of their second language. One area of interest for such systems is the implementation of corrective feedback. However, the feedback strategies employed by such systems remain fairly limited. In particular, there are currently no provisions for learners to initiate the correction of their own errors, despite this being the most frequently occurring and most preferred type of error correction in learner speech. To address this gap, this thesis proposes a framework for implementing such functionality, identifying incremental self-initiated self-repair (i.e. disfluency) detection as a key area for research. Taking an interdisciplinary approach to the exploration of this topic, this thesis outlines the steps taken to optimise an incremental disfluency detection model for use with spoken learner English. To begin, a linguistic comparative analysis of native and learner disfluency corpora explored the differences between the disfluency behaviour of native and learner speech, highlighting key features of learner speech not previously explored in disfluency detection model analysis. Following this, in order to identify a suitable baseline model for further experimentation, two state-of-the-art incremental self-repair detection models were trained and tested with a learner speech corpus. An error analysis of the models' outputs found an LSTM model using word embeddings and part-of-speech tags to be the most suitable for learner speech, thanks to its lower number of false positives triggered by learner errors in the corpus. Following this, several adaptations to the model were tested to improve performance. Namely, the inclusion of character embeddings, silence and laughter features, separating edit term detection from disfluency detection, lemmatization and the inclusion of learners' prior proficiency scores led to over an eight percent model improvement over the baseline. Findings from this thesis illustrate how the analysis of language characteristics specific to learner speech can positively inform model adaptation and provide a starting point for further investigation into the implementation of effective corrective feedback strategies in dialogue-based CALL systems.