CHARMinG -CHARacter MINing and Generation

The hero, the villain, the servant, the mentor, and many more ... movie and drama continue to rely on a repertoire of archetypical characters. But what makes a character? The proposed project CHARMinG will develop and apply AI methods from text and sentiment mining, natural language processing and machine learning to identify, from electronic sources of fictional dialogues (movie scripts, transcripts, drama texts), a set of indicators that convey the core of the relational/functional features and personality of characters, thereby leading to the generation of more colourful and engaging virtual characters. The project endeavours to establish the field of character mining for both fictional and non-fictional text domains. Fictional texts are a useful starting point because we can benefit from the theoretical background on characters design and rely on dialogue for their development. This research work will provide insights and methods for the design of artificial actors, assistive, educational, or entertaining virtual characters, and socially assistive robots. A central goal of the project is to make advances in the design of interesting virtual characters. These can be NPCs (non-player characters) for interactive stories, dramas, and games, but also assistive or educational agents, and their supporting cast. Through our research, which links textual surface features of a character to its underlying aspects in a systematic way, we endeavour at first to the modelling and generation of colourful, socially interrelated, and narratively supportive cast members that provide a social environment for the lead figures. The project will also investigate the portability of the newly adapted and developed methods as well as their synergies to non-fictional, primarily dialogic, text genres, e.g. online social communities.

Recent decades have seen a proliferation of virtual characters in interactive stories, games, assistive and educational technologies. However, the design of interesting artificial characters that would converse with users in consistent and coherent manner is still a challenge. The objective behind the project was to provide insights and novel interdisciplinary approaches to address this challenge. The core presupposition behind our investigation was that an important part of what constitutes a character transpires in dialogue. In line with this hypothesis, we developed and applied a range of natural language processing and machine learning methods that, in integration with fine-grained human-driven discourse analysis approaches, were used to identify, from electronic sources of fictional dialogues (movie scripts, transcripts, drama texts), a set of indicators that convey the core of the relational/functional features of characters. In summary, within a course of our project several landmarks were achieved. Primarily, we developed an integrative approach that complemented linguistic analysis with interactive and communication characteristics for the purpose of the automated identification of characters in fiction. Further, we evaluated a range of state-of-the-art universal word and sentence embeddings methods to advance the automated classification of the extensive set of categories: a) Expressivity features: sentiment, interjections, ellipsis; b) Social-relational features (e.g., dialog act types, address, discontinuities etc.) c) specific characters (mentor) strategic talk features (e.g., moves, intents). Given the wider range of associated functional and psycho-social characteristics that are relevant in light of the rapidly developing sectors of online tutoring, educational and assistive robotics, we directed our focus to mentor character and their dialogic interactions with mentee characters. Based on the integrative approach outlined above, a layered model of mentor-mentee interactions was proposed. Further, we have also developed computational tools for the automated analysis of character interactions in dialogues and evaluated a variant of conversational in character (i.e.: mentor) system. Finally, the portability of the methods and tools in other domains and application scenarios such as analysis of users characteristics in social media was explored.