摘要：

方面情感三元组抽取(ASTE)任务是方面级情感分析的重要子任务之一, 传统的监督学习方法在该任务上取得了SOTA或接近SOTA的效果。然而, 随着深度神经网络的发展, 生成式大型语言模型(LLM)为该任务带来了更多的可能性。目前大多数工作都是直接对LLM进行微调, 但是忽略了LLM的幻觉现象, 导致性能下降。提出一种融合思维链技术和LLM低秩自适应(LoRA)微调LFC方法, 实现生成式的ASTE新范式, 以提升任务性能。在LFC中, 首先基于思维链技术, 通过人工构造少量推理样本, 并利用LLM生成具有推理结构的增强数据集。将增强数据集用于微调ChatGLM3-6B模型的学习。在微调过程中, 采用LoRA微调技术提高在低资源环境下适配ASTE任务的效果。实验结果表明, LFC方法相比于最优的基线模型在Res14、Lap14、Res15和Res16 4个数据集上的F1值分别提升8.37、12.31、11.07和8.43个百分点, 该方法不仅能够准确地识别三元组, 而且在一定程度上优化了LLM的幻觉现象。

关键词: 方面情感三元组抽取, 大型语言模型, 低秩自适应微调, 思维链, 提示学习

Abstract:

The Aspect Sentiment Triplet Extraction (ASTE) task is an important subtask of aspect-level sentiment analysis. Conventional supervised learning methods achieve SOTA or near-SOTA results in this task. However, in developing deep neural networks, generative Large Language Models (LLM) offer additional possibilities for this task. Currently, most studies directly fine-tune the LLM but overlook its hallucinations, leading to performance degradation. To improve task performance, this paper proposes a LFC method for implementing a new generative ASTE paradigm. This method combines the Chain-Of-Thought (COT) technique and a fine-tuning approach based on LLM Low-Rank Adaption (LoRA). In LFC, based on COT technology, a few inference samples are manually constructed, and an enhanced dataset with an inference structure is generated using LLM. It uses an enhanced dataset to fine-tune the learning of the ChatGLM3-6B model. During the fine-tuning process, LoRA fine-tuning technology improves the effectiveness of adapting to ASTE tasks in low-resource environments. Experimental results show that compared with the optimal baseline model, the LFC method improves the F1 values by 8.37, 12.31, 11.07, and 8.43 percentage points on the Res14, Lap14, Res15, and Res16 datasets, respectively. This method accurately identifies triples and optimizes the hallucinations of the LLM to a certain extent.

Key words: Aspect Sentiment Triplet Extraction(ASTE), Large Language Model(LLM), Low-Rank Adaptation(LoRA) fine-tuning, Chain-Of-Thought(COT), prompt learning