Data-driven Repair Models for Text Chat with Language Learners

This research analyses participants' orientation to linguistic identities in chat and introduces data-driven computational models for communicative Intelligent Computer-Assisted Language Learning (communicative ICALL). Based on non-pedagogical chat conversations between native speakers and non-native speakers, computational models of the following types are presented: exposed and embedded corrections, explanations of unknown words following learner's request. Conversation Analysis helped to obtain patterns from a corpus of dyadic chat conversations in a longitudinal setting, bringing together German native speakers and advanced learners of German as a foreign language. More specifically, this work states a bottom-up, data-driven research design which takes “conversation” from its genuine personalised dyadic environment to a model of a conversational agent. It allows for an informal functional specification of such an agent to which a technical specification for two specific repair types is provided.

Starting with the open research objective to create a machine that behaves like a language expert in an informal conversation, this research shows that various forms of orientation to linguistic identities are on participants' disposal in chat. In addition it shows that dealing with computational complexity can be approached by a separation between local models of specific practices and a high-level regulatory mechanism to activate them. More specifically, this work shows that learners' repair initiations may be analysed as turn formats containing resources for signalling trouble and referencing trouble source. Based on this finding, this work shows how computational models for recognition of the repair initiations and trouble source extraction can be formalised and implemented in a chatbot. Further, this work makes clear which level of description of error corrections is required to satisfy computational needs, and how these descriptions may be transformed to patterns for various error correction formats and which technological requirements they imply. Finally, this research shows which factors in interaction influence the decision to correct and how the creation of a high-level decision model for error correction in a Conversation-for-Learning can be approached.

In sum, this research enriches the landscape of various communication setups between language learners and communicative ICALL systems explicitly covering Conversations-for-Learning. It strengthens multidisciplinary connections by showing how the multidisciplinary research field of ICALL benefits from including Conversation Analysis into the research paradigm. It highlights the impact of the micro-analytic understanding of actions accomplished by utterances in talk within a specific speech exchange system on computational modelling on the example of chat with language learners.