融合语义联想和BERT的图情领域SAO短文本分类研究

目的/意义] 针对SAO结构短文本分类时面临的语义特征短缺和领域知识不足问题，提出一种融合语义联想和BERT的SAO分类方法，以期提高短文本分类效果。方法/过程] 以图情领域SAO短文本为数据源，首先设计了一种包含"扩展-重构-降噪"三环节的语义联想方案，即通过语义扩展和SAO重构延展SAO语义信息，通过语义降噪解决扩展后的噪声干扰问题；然后利用BERT模型对语义联想后的SAO短文本进行训练；最后在分类部分实现自动分类。结果/结论] 在分别对比了不同联想值、学习率和分类器后，实验结果表明当联想值为10、学习率为4e-5时SAO短文本分类效果达到最优，平均F1值为0.852 2，与SVM、LSTM和单纯的BERT相比，F1值分别提高了0.103 1、0.153 8和0.140 5。

Research on SAO Short Text Classification in LIS Based on Semantic Association and BERT

Purpose/significance] Aiming at the shortage of semantic features and insufficient domain knowledge in the classification of SAO structure short texts, this paper proposes a SAO classification method combining semantic association and BERT in order to improve the classification effect.Method/process] Taking the SAO short text in the library and information science field as the data source, firstly, a semantic association scheme including the three links of "Expansion-Reconstruction-NoiseReduction" was designed. The semantic information of SAO was extended through semantic expansion and SAO reconstruction, and the extended noise interference problem was solved by semantic noise reduction; then used the BERT model to train the SAO short text after semantic association; finally realized automatic classification in the classification part.Result/conclusion] After comparing different association values, learning rates and classifiers, the experimental results show that when the association value is 10 and the learning rate is 4e-5, the SAO short text classification effect is optimal, and the average F1 value is 0.852 2, which is comparable to SVM and LSTM compared with pure BERT, the F1 value is increased by 0.103 1, 0.153 8 and 0.140 5 respectively.