Scholarlink

This scientific article describes a modification of the Retrieval Augmented Generation (RAG) method aimed at increasing the emotional content of model responses without supervised fine-tuning. In this model, RAG is not used to expand text generation beyond language knowledge but serves as a key tool for forming prompts based on several examples during a dialogue. The retrieval block of the model finds texts with suitable emotional contexts, which are later used to create prompts. This approach allows the model to generate responses rich in emotions that correspond to the context without the need for strict tuning. An investigation of this kind is being conducted for the first time for Russian-speaking dialogue agents. The developed approach enables enhancing emotional expressiveness and the quality of responses. Applying this method has shown a significant improvement in SSA values by 44% compared to baseline models trained on a dataset with emotional annotations, as well as a 25% enhancement in automatic evaluation based on a classifier.

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

RAG and Few-Shot Prompting in Emotional Text Generation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Elizaveta Vologina,
  • Anastasiia Matveeva,
  • Olesia Makhnytkina,
  • Yuri Matveev,
  • Nursaule Burambayeva

摘要

This scientific article describes a modification of the Retrieval Augmented Generation (RAG) method aimed at increasing the emotional content of model responses without supervised fine-tuning. In this model, RAG is not used to expand text generation beyond language knowledge but serves as a key tool for forming prompts based on several examples during a dialogue. The retrieval block of the model finds texts with suitable emotional contexts, which are later used to create prompts. This approach allows the model to generate responses rich in emotions that correspond to the context without the need for strict tuning. An investigation of this kind is being conducted for the first time for Russian-speaking dialogue agents. The developed approach enables enhancing emotional expressiveness and the quality of responses. Applying this method has shown a significant improvement in SSA values by 44% compared to baseline models trained on a dataset with emotional annotations, as well as a 25% enhancement in automatic evaluation based on a classifier.