融合Graph state LSTM与注意力机制的跨句多元关系抽取

CROSS-SENTENCE N-ARY RELATION EXTRACTION COMBININGGRAPH STATE LSTM AND ATTENTION MECHANISM

已有的跨句多元关系抽取工作将输入文本表示为集成句内和句间依赖关系的复杂文档图,但图中包含的噪声信息会影响关系抽取的效果。针对这种情况,该文利用Graph state LSTM获得上下文信息,再分别利用词级注意力机制或位置感知的注意力机制,自动聚焦在对关系抽取起到决定性作用的关键词上,降低噪声信息的影响。并且比较了两种注意力机制对使用Graph state LSTM进行关系抽取的影响。通过在一个重要的精确医学数据集上进行实验,验证了该文所提出模型的有效性。

In the existing cross sentence n-ary relations extraction works,the input text is represented as a complex document graph integrating various intra-sentential and inter-sentential dependencies.The noise information contained in the graph will affect the effect of relation extraction.In view of this,this paper used graph state LSTM to obtain contextual information,and automatically focused on the keywords that played a decisive role in relation extraction by using the word-level attention mechanism or the position aware attention mechanism respectively,to reduce the influence of noise information.This paper compared the influence of two attention mechanisms on relation extraction using graph state LSTM.The validity of the proposed model was validated by conducting experiments on a significant precision medical dataset.