基于注意力机制的农业金融文本关系抽取研究

Extracting Relationship of Agricultural Financial Texts with Attention Mechanism

【目的】研究中文文本中关系自动抽取的方法。【方法】以224家农业上市公司2015年–2017年的678份年报为数据来源,采用基于双重注意力机制的门控循环单元算法,进行中文文本关系自动抽取研究。【结果】最终模型在农业金融文本数据集上的平均准确率达78%,相较循环神经网络算法,该算法平均准确率提高约12%。【局限】仅针对224家农业上市公司的数据进行研究,研究涉农企业对象有待进一步拓展。【结论】该模型能够在农业金融相关文本的关系抽取上取得较好效果。

[Objective] This paper proposes a new method to extract relations from Chinese texts automatically.[Methods] We retrieved annual reports of 224 listed agricultural companies from 2015 to 2017. Then we adopted the Gated Recurrent Unit algorithm based on double attention mechanism to extract the needed data.[Results] The average accuracy of our model on the agricultural financial dataset reached 78%. Compared with the Recurrent Neural Network algorithm, the average accuracy of the new model increased by about 12%.[Limitations] We only studied data from 224 companies, which needs to be expanded.[Conclusions] The proposed model can effectively extract relationship from agricultural financial texts.