基于深度学习的文本中细粒度知识元抽取方法研究

Extracting Fine-grained Knowledge Units from Texts with Deep Learning

【目的】改进Bootstrapping方法,建立深度学习模型从文本中抽取多类型细粒度的知识元。【方法】利用搜索引擎和Elsevier关键词构建知识元词库;基于Bootstrapping技术自动构建大规模的标注语料库,利用知识元评分模型和模式评分模型控制标注的质量;基于已标注多类型知识元的语料库训练LSTM-CRF模型,从文本中抽取新的知识元。【结果】基于17 756篇ACL论文摘要抽取"研究范畴"、"研究方法"、"实验数据"、"评价指标及取值"这4种知识元,其人工评价平均正确率为91%。【局限】模型参数的预设与调整需要人工参与,未对不同领域文本进行适用性验证。【结论】引入知识元与模式的评分模型,能够有效缓解"语义漂移"问题;基于深度学习模型抽取知识元实现快速且正确率高,为情报大数据智能分析提供了一种高效可靠的数据获取手段。

[Objective] This paper tries to extract fine-grained knowledge units from texts with a deep learning model based on the modified bootstrapping method. [Methods] First, we built the lexicon for each type of knowledge unit with the help of search engine and keywords from Elsevier. Second, we created a large annotated corpus based on the bootstrapping method. Third, we controlled the quality of annotation with the estimation models of patterns and knowledge units. Finally, we trained the proposed LSTM-CRF model with the annotated corpus, and extracted new knowledge units from texts. [Results] We retrieved four types of knowledge units(study scope, research method, experimental data, as well as evaluation criteria and their values) from 17,756 ACL papers. The average precision was 91%, which was calculated manually. [Limitations] The parameters of models were pre-defined and modified by human. More research is needed to evaluate the performance of this method with texts from other domains. [Conclusions] The proposed model effectively addresses the issue of semantic drifting. It could extract knowledge units precisely, which is an effective solution for the big data acquisition process of intelligence analysis.