Stylistic Dialogue Generation Based on Character Personality in Narrative Films.

Abstract

Traditional narrative systems consist of two steps of process, story generation and discourse generation. However, many interactive systems make more effort on story generation rather than discourse generation. For discourse generation, dialogue is an important way used to unfold story and reveal characters in stories, and it is reasonable to expand the capability of narrative system by exploring the potential of dialogue generation in narratives. Also, Recent research in conditional dialogue generation is mostly focusing on the context of natural conversation generation with speakers’ profile information. While incorporating the styles that relevant to narratives is yet to be widely investigated. According to the research made, in this document, we propose an approach using a pre-trained language model, in order to explore the potential of generating dialogues with embedded narrative-related features within the context of narrative films. In this approach, three different embedding methods are leveraged to incorporate Big-Five personalities of characters into transformer-based neural networks, training on a new corpus, which is created and well-parsed from screenplays. We conduct experiments using both automatic metrics and human evaluation to measure the quality of the generated dialogue and personality identification accuracy. All the dialogues for evaluation and analysis are generated with settings of the perspectives of embedding method, personality trait, personality level, and film genre, which is to explore the impact of different setting on dialogue generation with additional narrative-related styles. According to the automatic experimental results, we demonstrate that our approach is able to generate dialogues with increased variety. Also overall, the generated dialogues are able to correctly reflect the given target personality. We also conduct three user studies for evaluate dialogues with human judgements. In the first and the second user study, we evaluate the dialogues generated with film- level personality using CTE (Combined Textual Embedding) embedding method. The results show that human participants are inclined to perceive one extreme end of each personality trait. In the third user study, we evaluate generated dialogues with all setting combinations synthetically. Overall, the results show that target personalities can be identified with various degrees of accuracy. Also, a negative correlation between personality identification accuracy and dialogue quality is observed. In this thesis, we propose a new approach for stylistic dialogue generation and demonstrate its effectiveness. We believe the observations and discoveries could be a start and a tryout to apply deep learning technique and big data to boost narrative dialogue generation. And we also believe that our research can be applied in plenty of potential scenarios, such as helping the authors creating huge amount of conversations between different characters by popping utterance options corresponding to the character settings.