Journal of Frontiers of Computer Science and Technology ›› 2023, Vol. 17 ›› Issue (8): 1961-1973.DOI: 10.3778/j.issn.1673-9418.2205064

• Artificial Intelligence·Pattern Recognition • Previous Articles     Next Articles

Low Resource Summarization Model Based on Latent Structural Semantic En-hancement

LIU Yu, LIU Xiaoming, LIU Weiguang, YANG Guan, LIU Jie   

  1. 1. School of Computer Science, Zhongyuan University of Technology, Zhengzhou 450007, China
    2. Henan Key Laboratory on Public Opinion Intelligent Analysis, Zhengzhou 450007, China
    3. Software College, Zhongyuan University of Technology, Zhengzhou 450007, China
    4. School of Information Science, North China University of Technology, Beijing 100144, China
    5. China Language Intelligence Research Center, Beijing 102206, China
  • Online:2023-08-01 Published:2023-08-01

基于潜层结构化语义增强的低资源摘要模型

刘宇,刘小明,刘卫光,杨关,刘杰   

  1. 1. 中原工学院 计算机学院,郑州 450007
    2. 河南省网络舆情监测与智能分析重点实验室,郑州 450007
    3. 中原工学院 软件学院,郑州 450007
    4. 北方工业大学 信息学院,北京 100144
    5. 国家语委中国语言智能研究中心,北京 102206

Abstract: At present, low-resource summary generation tasks are usually processed by data enhancement or pre-training combined with fine-tuning, which cannot make full use of the latent structural semantic information between the source text and the target summary. For this reason, this paper proposes a low resource summary model based on latent structural semantic enhancement, which enhances the utilization of structured information in the way of graph structure alignment. First of all, the model obtains the latent semantic features of the source text and prediction summary through the structural feature representation layer. Then, the obtained semantic features are aligned with the latent structured alignment module for node alignment and edge alignment, which helps the model to capture the structured information in the semantic features, thus enhancing the model??s use of structured knowledge. Finally, the model uses the structured feature alignment distance between the source text and the prediction summary as the regular term of target loss to assist the model in optimization. Experiments are performed on a low-resource dataset across six domains. The model achieves an average improvement of 0.58 in ROUGE-1 scores relative to the baseline model. The results show that the model can effectively improve the ability of generating low-resource summaries by using latent structured semantic knowledge.

Key words: low resources, structured, semantic features, graph structure

摘要: 当前低资源摘要生成任务通常采用数据增强或预训练结合微调的方式进行处理,对于源文本与目标摘要之间的潜层结构化语义信息未能充分利用。为此,提出一种基于潜层结构化语义增强的低资源摘要模型,以图结构对齐的方式增强模型对结构化信息的利用。首先,该模型通过结构特征表示层获取源文本与预测摘要的潜层结构化语义特征。然后,将获得的语义特征利用潜层结构对齐模块进行节点对齐和边对齐,这种对齐有助于模型捕捉语义特征中的结构化信息,从而增强模型对结构化知识的利用。最后,利用源文本与预测摘要之间的结构化特征对齐距离作为目标损失的正则项来辅助模型进行优化。在六个领域的低资源数据集上进行实验,ROUGE-1分值相对于基线模型平均提高了0.58。结果表明利用潜层结构化语义知识可以有效提高低资源摘要生成的能力。

关键词: 低资源, 结构化, 语义特征, 图结构