Conversation analysis for computational modelling of task-oriented dialogue

Duran, Nathan

Abstract

Current methods of dialogue modelling for Conversational AI (CAI) bear little resemblance to the manner in which humans organise conversational interactions. The way utterances are represented, interpreted, and generated are determined by the necessities of the chosen technique and do not resemble those used during natural conversation. In this research we propose a new method of representing task-oriented dialogue, for the purpose of computational modelling, which draws inspiration from the study of human conversational structures, Conversation Analysis (CA). Our approach unifies two well established, yet disparate, methods of dialogue representation: Dialogue Acts (DA), which provide valuable semantic and intentional information, and the Adjacency Pair (AP), which are the predominant method by which structure is defined within CA. This computationally compatible approach subsequently benefits from the strengths, whilst overcoming the weaknesses, of its components.
To evaluate this thesis we first develop and evaluate a novel CA Modelling Schema (CAMS), which combines concepts of DA’s and AP’s to form AP-type labels. Thus creating a single annotation scheme that is able to capture the semantic and syntactic structure of dialogue. We additionally annotate a task-oriented corpus with our schema to create CAMS-KVRET, a first-of-its-kind DA and AP labelled dataset. Next, we conduct detailed investigations of input representation and architectural considerations in order to develop and refine several ML models capable of automatically labelling dialogue with CAMS labels. Finally, we evaluate our proposed method of dialogue representation, and accompanying models, against several dialogue modelling tasks, including next label prediction, response generation, and structure representation.
With our evaluation of CAMS we show that it is both reproducible, and inherently learnable, even for novice annotators. And further, that it is most intuitively applied to task-oriented dialogues. During development of our ML classifiers we determined that, in most cases, input and architectural choices are equally applicable to DA and AP classification. We evaluated our classification models against CAMS-KVRET, and achieved high test set classification accuracy for all label components of the corpus. Additionally, we were able to show that, not only is our model capable of learning the semantic and structural aspects of both the DA and AP components, but also that AP are more predictive of future utterance labels, and thus representative of the overall dialogue structure. These finding were further supported by the results of our next-label prediction and response generation experiments. Moreover, we found AP were able to reduce the perplexity of the generative model. Finally, by using χ2 analysis to create dialogue structure graphs, we demonstrate that AP produce a more generalised and efficient method of dialogue representation. Thus, our research has shown that integrating DA with AP, into AP-type labels, captures the semantic and syntactic structure of an interaction, in a format that is independent of the domain or topic, and which benefits the computational modelling of task-oriented dialogues.

• Thesis Type: Thesis
• Deposit Date: Oct 6, 2022
• Publicly Available Date: Feb 27, 2023
• Keywords: Conversation Analysis, Adjacency Pairs, Dialogue Acts, Machine Learning, Artificial Intelligence, Natural Language Processing
• Public URL: https://uwe-repository.worktribe.com/output/10021806
• Award Date: Feb 27, 2023