Narrative Dialogue Dataset: Speaker and Emotion Annotated Conversational Corpus.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Artificial Intelligence

Background:

• Dialogue datasets are crucial for training conversational AI.
• Existing datasets often lack nuanced emotional and speaker annotations.
• Narrative fiction offers rich, context-embedded dialogue for dataset creation.

Purpose of the Study:

• To develop a novel, automatically annotated dialogue dataset from narrative short stories.
• To enable keyword-conditioned narrative generation and emotionally expressive AI.
• To provide a resource for fine-tuning generative language models.

Main Methods:

• Utilized a Hugging Face story corpus for data sourcing.
• Employed FastCoref and BookNLP for speaker attribution and entity resolution.
• Developed a Bi-GRU emotion classifier and Mistral-7B for keyword extraction.

Main Results:

• Created a dataset of 2,588 stories with 27,074 annotated dialogue utterances.
• Each dialogue is tagged with speaker, emotion (6 categories), and story keywords.
• Demonstrated dataset utility through keyword-conditioned narrative generation and emotional Text-to-Speech (TTS).

Conclusions:

• The novel dataset significantly enhances capabilities for training emotionally aware dialogue systems.
• This resource facilitates advancements in AI-driven storytelling and expressive speech synthesis.
• The methodology provides a scalable approach for creating similar annotated dialogue corpora.