Scholarlink

<p>This study presents a complete end-to-end pipeline for generating emotionally annotated dialogues in the format <i>Speaker (Emotion): Dialogue</i>, conditioned on a given set of keywords. The system begins with a Hugging Face short story dataset used to build the training corpus. To accurately identify speakers and gender, FastCoref was applied with dialogue masking, followed by BookNLP for improved entity and gender resolution. Emotional context was assigned to each dialogue using a Bi-GRU–based emotion classification model, while representative keywords were extracted using Mistral-7B through few-shot prompt engineering. The resulting dataset of keyword–story pairs was then used to fine-tune three lightweight generative models: TinyLlama, Phi-1.5, and GPTNeo for conditional story generation. Model outputs were evaluated through both automated and human-centered assessments. Automated metrics such as perplexity and Distinct-1/Distinct-2 measured fluency and lexical diversity, while an LLM-based reviewer (DeepSeek-chat-v3-0324) rated coherence, creativity, and linguistic quality. In parallel, a human evaluation was conducted using Google Forms, where five participants rated 20 generated stories on coherence, emotional engagement, and overall quality. The averaged human ratings were compared with LLM-based scores to validate consistency in subjective and model-based evaluation. Overall, GPTNeo achieved the highest combined performance, with superior lexical diversity and an average rating of 4.0/5, followed by Phi-1.5 (3.6) and TinyLlama (2.8).</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Script generation from synthetic stories using lightweight LLM and emotion-aware TTS pipelines

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月5日

查看全文
  • Uzma Patil,
  • Aniket K. Shahade

摘要

This study presents a complete end-to-end pipeline for generating emotionally annotated dialogues in the format Speaker (Emotion): Dialogue, conditioned on a given set of keywords. The system begins with a Hugging Face short story dataset used to build the training corpus. To accurately identify speakers and gender, FastCoref was applied with dialogue masking, followed by BookNLP for improved entity and gender resolution. Emotional context was assigned to each dialogue using a Bi-GRU–based emotion classification model, while representative keywords were extracted using Mistral-7B through few-shot prompt engineering. The resulting dataset of keyword–story pairs was then used to fine-tune three lightweight generative models: TinyLlama, Phi-1.5, and GPTNeo for conditional story generation. Model outputs were evaluated through both automated and human-centered assessments. Automated metrics such as perplexity and Distinct-1/Distinct-2 measured fluency and lexical diversity, while an LLM-based reviewer (DeepSeek-chat-v3-0324) rated coherence, creativity, and linguistic quality. In parallel, a human evaluation was conducted using Google Forms, where five participants rated 20 generated stories on coherence, emotional engagement, and overall quality. The averaged human ratings were compared with LLM-based scores to validate consistency in subjective and model-based evaluation. Overall, GPTNeo achieved the highest combined performance, with superior lexical diversity and an average rating of 4.0/5, followed by Phi-1.5 (3.6) and TinyLlama (2.8).