Scholarlink

This paper focuses on applying reinforcement learning and parameter efficient fine-tuning methods to train NLP model to generate dialogue utters in a style of a specific character. The goal is to fine-tune relatively small model that will not consume a lot of computing resources and can be deployed and used on a single machine. As a result, this resource efficient method can be used for implementation of character-specific chat bots playing role of a specific character, generating fiction stories or new data sets. We propose using PPO algorithm together with LoRA to fine-tune GPT2 model to generate a character style utterances. For the reward function, BERT model was trained to distinguish between the desired style texts and regular ones, BERTScore and Self-BLUE were used to improve the dialogue flow quality. Dataset for training was generated with GPT4-mini.

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

Dialogue Style Transfer with Reinforcement Learning and Parameter Efficient Fine-Tuning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ruslan Pravosud,
  • Oleksii Adamov,
  • Oleksandr Marchenko

摘要

This paper focuses on applying reinforcement learning and parameter efficient fine-tuning methods to train NLP model to generate dialogue utters in a style of a specific character. The goal is to fine-tune relatively small model that will not consume a lot of computing resources and can be deployed and used on a single machine. As a result, this resource efficient method can be used for implementation of character-specific chat bots playing role of a specific character, generating fiction stories or new data sets. We propose using PPO algorithm together with LoRA to fine-tune GPT2 model to generate a character style utterances. For the reward function, BERT model was trained to distinguish between the desired style texts and regular ones, BERTScore and Self-BLUE were used to improve the dialogue flow quality. Dataset for training was generated with GPT4-mini.