A Study of Text Summarization Based on Large Models

This study explores a large-scale language model (Large Language Models,LLM)-based text summarization generation method using LoRA (Low-Rank Adaptation) fine-tuning technique. In recent years, large-scale pre-trained language models have shown excellent performance in natural language processing tasks, especially in the field of text summarization. However, due to the large model parameters, the fine-tuning and application process often faces challenges in terms of computational and storage resources. To address this problem, the LoRA method is proposed to efficiently fine-tune pre-trained models by introducing low-rank matrix adapters, which significantly reduces computational cost and improves training efficiency. In this thesis, the principle of LoRA fine-tuning and its implementation are first introduced, followed by experimental validation in several text summarization tasks. The experimental results show that the LoRA-tuned model not only maintains comparable summarization quality with the traditional full-parameter-tuned model, but also achieves significant advantages in training speed and computational resource consumption.