«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

期数:

2025年05期

页码:

106-110

栏目:

人工智能

出版日期:

2025-05-10

文章信息/Info

Title:

Research on Aspect-based Sentiment Analysis Based on ChatGLM3-6B

文章编号:

1673-629X(2025)05-0106-05

作者:

孙晓晔1; 2; 成彬1; 2*; 王程1; 2

1. 河北省科学院应用数学研究所,河北 石家庄 050081;

2. 河北省信息安全认证技术创新中心,河北 石家庄 050081

Author(s):

SUN Xiao-ye1; 2; CHENG Bin1; 2*; WANG Cheng1; 2

1. Institute of Applied Mathematics,Hebei Academy of Sciences,Shijiazhuang 050081,China;

2. Information Security Authentication Technology Innovation Center of Hebei Province,Shijiazhuang 050081,China

关键词:

自然语言处理; 方面级情感分析; 大语言模型; 数据筛选; 指令微调

Keywords:

natural language processing; aspect-based sentiment analysis; large language model; data screening; instruction tuning

分类号:

TP391.1

DOI:

10.20165/j.cnki.ISSN1673-629X.2024.0391

摘要:

为了利用大语言模型不依赖于标注数据、上下文理解和迁移能力强的优势,该文将 LLM 用于公开数据集匮乏的统一方面级情感分析(Unified Aspect-Based Sentiment Analysis,UABSA)任务,提出了一种结合数据筛选和 LLM 指令微调的方法 SLIT,并在多个领域的中英文数据集上进行了实验。 首先,通过对比训练集中样本之间的相似度,筛除相似度高的部分样本数据,得到精简数据集;然后,设计提示模板,用精简数据集构造用于微调的指令数据集;最后,对 ChatGLM3-6B 大模型进行 P-Tuning v2 微调,在测试集上验证效果。 实验结果表明,与使用未经过数据筛选的 ChatGLM3-6B+微调方法相比,SLIT 虽然使用的数据更少,但在 Rest14、Laptop14、扫地机器人 3 个数据集上的 F1 值分别提高了 3. 95 百分点、2. 01 百分点、4. 62 百分点,与其他基准方法相比,在 Rest14 数据集上的效果达到了最佳。

Abstract:

To leverage the advantages of large language models,including the independence from labeled data,the excellent capabilities of contextual understanding and transfer learning,we apply LLM to the Unified Aspect-Based Sentiment Analysis (UABSA) task,which lacks labeled datasets. We propose a method called SLIT that combines data screening and LLM instruction tuning and have conducted experiments on Chinese and English datasets across various domains. Firstly,by comparing the similarity between samples in the training set,we filter out some highly similar samples to get a refined dataset. Secondly,we design a prompt template and construct an instruction dataset using the refined dataset. Finally,we apply P-Tuning v2 fine-tuning on the ChatGLM3-6B model and verify the results on the test set. The experiment shows that compared to the ChatGLM3-6B+fine-tuning method without data screening,SLIT improves the F1 score by 3. 95 percentage points,2. 01 percentage points,and 4. 62 percentage points on the Rest14,Laptop14,and robotic vacuum cleaner datasets,respectively. Compared to other baseline methods,SLIT achieves the best performance on the Rest14 dataset.

常用功能

更新日期/Last Update: 2025-05-10