基于BERT模型的图书表示学习与多标签分类研究  

Representation Learning and Multi-label Classification of Books Based on BERT

作　　者：蒋彦廷 胡韧奋[1] Jiang Yanting;Hu Renfen

机构地区：[1]北京师范大学中文信息处理研究所,北京海淀100875

出　　处：《新世纪图书馆》2020年第9期38-44,共7页New Century Library

基　　金：国家社科基金青年项目“面向汉语国际教育的智能测试技术研究”(项目编号:18CYY029)研究成果之一。

摘　　要：中文图书细粒度多标签分类的自动化,有利于促进图书的检索与学科的沟通。文章充分发挥BERT语言模型的微调特性,提出一种通过21类粗粒度分类微调语言模型,学习到更好的图书表示,进而实现细粒度分类的新策略。结果显示,在单标签的分类任务上,BERT模型的正确率分别较LSTM与Fasttext模型提升约4.9%与2.0%。KNN-ML对257类的细粒度多标签分类证明了前期微调的有效性。最佳情况下,有75.82%的图书细粒度类别恰好全部预测正确,92.10%的图书至少被正确预测了一个细粒度类别。因此可以得出结论,该系统有助于实现图书自动的细粒度归类,并帮助图书标引者补充合理的分类号。The automation of the fine-grained multi-label classification of Chinese books is beneficial to the book index and subject communication.This paper makes full use of fine-tuning of BERT model and puts forward a novel strategy which fine-tunes the model on the coarse-grained classification task to learn a better book representation,and then completes the multi-label classification.The result shows that on the single-label classification,the accuracy of BERT has increased by about 4.9%and 2.0%compared with LSTM and Fasttext.The classification result of KNNML indicates the effectiveness of fine-tuning.Under the best situation,75.82%of books are correctly sorted out,and 92.10%of books are predicted with at least one correct label.It draws a conclusion that this system is of great benefit to automatic fine-grained classification,and can help book annotators replenish the potential missing category code.

关 键 词：中文图书 BERT模型 深度学习 微调策略 多标签分类 

分 类 号：G254.1[文化科学—图书馆学]