【摘要】 【目的/意义】从大量非结构化文本中抽取出结构化的实体及其关系,是优化搜索引擎、建立知识图谱、开发智能问答系统的基础工作。【方法/过程】介绍了深度学习框架下不同神经网络模型实现实体关系抽取的方法,比较了各种模型的优劣势,结合远程监督和注意力机制进一步提高关系抽取性能,最后指出了深度学习模型的不足及未来发展方向。【结果/结论】实验发现,卷积神经网络擅长捕获句子局部关键信息,循环神经网络擅长捕获句子的上下文信息,能反映句子多个实体之间的高阶关系,递归神经网络适合短文本的关系抽取。如果模型能结合自然语言的先验知识,实体关系抽取将会取得更好的效果。更多还原

【Abstract】 【Purpose/significance】Extracting structured entities and their relation from a large number of unstructured textsis the basic work of optimizing search engine, building knowledge graph and developing intelligent question answering sys-tem.【Method/process】This paper introduced the methods of different neural network to implement entity relation extrac-tion, compared the advantages and disadvantages of various models, further improved the performance of relation extractionwith distant supervision and attention mechanism, paper finally pointed out the shortcomings of the deep learning and thefuture development.【Result/conclusion】The experiments show that convolutional neural network is good at capturing thelocal key information of sentences, recurrent neural network is good at capturing the contextual information of sentences, re-flecting the higher order relations between multiple entities, and the recursive neural network is suitable for the relation ex-traction of short texts. If model can combine the prior knowledge of natural language, the entity relation extraction willachieve better results.更多还原