基于ChatGPT增强和监督对比学习的政策工具归类研究

doi:10.3778/j.issn.1002-8331.2308-0354

摘要/Abstract

摘要： 政策工具的归类是政策文本量化分析和研究的重要维度之一。由于训练数据的缺乏，政策文本相似度高，模型难以学习到足够丰富的特征表示，对它的预测结果缺少置信度，有较高的错误分类风险。为此提出了基于ChatGPT增强和监督对比学习的政策工具分类方法，该方法分为预训练语言模型微调和ChatGPT决策增强两个阶段，第一阶段使用ChatGPT大语言模型增强政策文本以增加训练数据数量，结合监督对比学习微调RoBERTa模型，使模型学习到更丰富的政策文本表示；第二阶段使用ChatGPT辅助决策预训练语言模型置信度较低的文本，降低对相似文本的错误分类风险。在数字产业政策工具分类数据集和Tnews数据集上的实验表明，所提方法优于主流的研究方法，能够有效提升基模型的性能，且在训练样本较少时对基线模型的提升更显著。

关键词: 文本分类, ChatGPT, 数据增强, 监督对比学习, 政策工具

Abstract: The classification of policy tools is an important dimension in the quantification and analysis of policy texts. Due to the scarcity of training data, models are prone to overfitting, resulting in reduced prediction confidence and an increased risk of misclassification. Therefore, policy tool classification method based on ChatGPT augmentation and supervised contrastive learning is proposed. The method consists of two stages：pre-training language model fine-tuning and ChatGPT decision augmentation. In the first stage, ChatGPT, a large language model, augments policy texts to increase the training dataset, while supervised contrastive learning fine-tune the RoBERTa model to improve classification performance. In the second stage, ChatGPT assists in the decision-making process for low confidence texts through the pre-trained language model and reduces the risk of misclassifying similar texts. Experiments on the digital industry policy tools classification dataset and the Tnews dataset show that the proposed method surpasses mainstream research approaches and can effectively improve the performance of the base model, with a more significant improvement observed when the training samples are limited.

Key words: text classification, ChatGPT, data augment, supervised contrastive learning, policy tools

胡志强, 李朋骏, 王金龙, 熊晓芸. 基于ChatGPT增强和监督对比学习的政策工具归类研究[J]. 计算机工程与应用, 2024, 60(7): 292-305.

HU Zhiqiang, LI Pengjun, WANG Jinlong, XIONG Xiaoyun. Research on Policy Tools Classification Based on ChatGPT Augmentation and Supervised Contrastive Learning[J]. Computer Engineering and Applications, 2024, 60(7): 292-305.

参考文献

[1] 李娜, 姜恩波, 朱一真, 等. 政策工具自动识别方法与实证研究[J]. 图书情报工作, 2021, 65(7): 115-122.
LI N, JIANG E B, ZHU Y Z, et al. Policy tool identification method and empirical research based on deep learning[J]. Library and Information Service, 2021, 65(7): 115-122.
[2] 霍朝光, 霍帆帆, 王婉如, 等. 基于WordBERT和BiLSTM的政策工具自动分类方法研究[J]. 图书情报知识, 2023, 40(3): 129-138.
HUO C G, HUO F F, WANG W R, et al. Automatic classificatin method of policy tools based on WordBERT and BiLSTM[J]. Documentation, Information & Knowledge, 2023, 40(3): 129-138.
[3] 马雨萌, 黄金霞, 王昉, 等. 融合BERT与多尺度CNN的科技政策内容多标签分类研究[J]. 情报杂志, 2022, 41(11): 157-163.
MA Y M, HUANG J X, WANG F, et al. Research on multi-label classification of S&T policy content combing BERT and multi-sale CNN[J]. Journal of Intelligence, 2022, 41(11): 157-163.
[4] 朱娜娜, 王航, 张家乐, 等. 基于预训练语言模型的政策识别研究[J]. 中文信息学报, 2022, 36(2): 104-110.
ZHU N N, WANG H, ZHANG J L, et al. Policy identification based on pretrained language model[J]. Journal of Chinese Information Processing, 2022, 36(2): 104-110.
[5] GUO C, PLEISS G, SUN Y, et al. On calibration of modern neural networks[C]//International Conference on Machine Learning, 2017: 1321-1330.
[6] WEI J, TAY Y, BOMMASANI R, et al. Emergent abilities of large language models[J]. arXiv:2206.07682, 2022.
[7] 张华平, 李林翰, 李春锦. ChatGPT中文性能测评与风险应对[J]. 数据分析与知识发现, 2023, 7(3): 16-25.
ZHANG H P, LI L H, LI C J. ChatGPT performance evaluation on Chinese language and risk measures[J]. Data Analysis and Knowledge Discovery, 2023, 7(3): 16-25.
[8] ZHOU J, KE P, QIU X, et al. ChatGPT: potential, prospects, and limitations[J]. Frontiers of Information Technology & Electronic Engineering, 2023: 1-6.
[9] LIANG D, YI B. Two-stage three-way enhanced technique for ensemble learning in inclusive policy text classification[J]. Information Sciences, 2021, 547: 271-288.
[10] 胡吉明, 付文麟, 钱玮, 等. 融合主题模型和注意力机制的政策文本分类模型[J]. 情报理论与实践, 2021, 44 (7): 159-165.
HU J M, FU W L, QIAN W, et al. Research on policy text classification method based on topic model and attention mechanism[J]. Information Studies: Theory & Application, 2021, 44(7): 159-165.
[11] 石金泽. 基于图神经网络的政策文本多标签分类[D]. 济南: 齐鲁工业大学, 2023.
SHI J Z. Multi-label classification of policy text based on graph neural network[D]. Jinan: Qilu University of Technology, 2023.
[12] DEVLIN J, CHANG M W, LEE K, et al. BERT: pre-training of deep bidirectional transformers for language understanding[C]//Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Stroudsburg, PA: Association for Computational Linguistics, 2019: 4171-4186.
[13] YE J, CHEN X, XU N, et al. A comprehensive capability analysis of GPT-3 and GPT-3.5 series models[J]. arXiv:2303.10420, 2023.
[14] SUN X, LI X, LI J, et al. Text classification via large language models[J]. arXiv:2305.08377, 2023.
[15] 张重生, 陈杰, 李岐龙, 等. 深度对比学习综述[J]. 自动化学报, 2023, 49(1): 15-39.
ZHANG C S, CHEN J, LI Q L, et al. Deep contrastive learning: a survey[J]. Acta Automatica Sinica, 2023, 49(1): 15-39.
[16] KHOSLA P, TETERWAKT P, WANG C, et al. Supervised contrastive learning[C]//Proceedings of the 34th International Conference on Neural Information Processing Systems, 2020: 18661-18673.
[17] GUNEL B, DU J, CONNEAU A, et al. Supervised contrastive learning for pre-trained language model fine-tuning[C]//International Conference on Learning Representations, 2021.
[18] GAO T, YAO X, CHEN D. SimCSE: simple contrastive learning of sentence embeddings[C]//2021 Conference on Empirical Methods in Natural Language Processing, 2021: 6894-6910.
[19] SEDGHAMIZ H, RAVAL S, SANTUS E, et al. SupCL-Seq: supervised contrastive learning for downstream optimized sequence representations[C]//Findings of the Association for Computational Linguistics, 2021: 3398-3403.
[20] LI S, HU X, LIN L, et al. Pair-level supervised contrastive learning for natural language inference[C]//2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022: 8237-8241.
[21] 高怡, 纪焘, 吴苑斌, 等. 基于标签增强和对比学习的鲁棒小样本事件检测[J]. 中文信息学报, 2023, 37(4): 98-108.
GAO Y, JI T, WU Y B, et al. Robust few shot event detection based on label augmentation and contrastive learning[J]. Journal of Chinese Information Processing, 2023, 37(4): 98-108.
[22] LIU Y, OTT M, GOYAL N, et al. RoBERTa: a robustly optimized BERT pretraining approach[J]. arXiv:1907.11692, 2019.
[23] LI B, HOU Y, CHE W. Data augmentation approaches in natural language processing: a survey[J]. AI Open, 2022, 3: 71-90.
[24] DAI H, LIU Z, LIAO W, et al. AugGPT: leveraging ChatGPT for text data augmentation[J]. arXiv:2302.13007, 2023.
[25] 严豫, 杨笛, 尹德春. 融合大语言模型知识的对比提示情感分析方法[J]. 情报杂志, 2023, 42(11): 126-134.
YAN Y, YANG D, YIN D C. Contrastive-based prompt-tuning sentiment analysis method incorporating large language model knowledge[J]. Journal of Intelligence , 2023, 42(11): 126-134.
[26] SUN X, DONG L, LI X, et al. Pushing the limits of ChatGPT on NLP Tasks[J]. arXiv:2306.09719, 2023.
[27] BANG Y, CAHYAWIJAVA S, LEE N, et al. A multitask, multilingual, multimodal evaluation of ChatGPT on reasoning, hallucination, and interactivity[J]. arXiv:2302.04023, 2023.
[28] ZHANG Z, ZHANG A, LI M, et al. Automatic chain of thought prompting in large language models[J]. arXiv:2210.
03493, 2022.
[29] ROTHWELL R. Reindustrialization and technology: towards a national policy framework[J]. Science and Public Policy, 1985, 12(3): 113-130.
[30] XU L, LU X, YUAN C, et al. FewCLUE: a Chinese few-shot learning evaluation benchmark[J]. arXiv:2107.07498, 2021.
[31] SUN Y, WANG S, FENG S, et al. Ernie 3.0: large-scale knowledge enhanced pre-training for language understanding and generation[J]. arXiv:2107.02137, 2021.
[32] SUN Z, LI X, SUN X, et al. ChineseBERT: Chinese pretraining enhanced by glyph and pinyin information[C]//Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 2021: 2065-2075.
[33] LIU X, ZHENG Y, DU Z, et al. GPT understands, too[J]. arXiv:2103.10385, 2021.
[34] LIU X, JI K, FU Y, et al. P-tuning v2: prompt tuning can be comparable to fine-tuning universally across scales and tasks[J]. arXiv:2110.07602, 2021.