摘要： 在信息抽取过程中，无法被判别的回指易造成信息抽取不完整的情况，这种指代关系可通过分析当前语境下的指代部分、被指代部分、周围的信息及原文内容生成的唯一判别信息进行判断。为此，构建一个多层注意力机制模型，在不同层次上对上述信息进行基于注意力机制的概率计算，利用最终结果判别回指关系是否成立。在指代部分与被指代部分向量化后，通过2个注意力层上的4次概率计算，使每一个训练结果在判别之前都具有唯一性。在OntoNotes 5.0数据集上的实验结果表明，该模型F值在显性指代和零指代均存在的条件下为70.1%，在存在零指代的条件下为60.7%，高于尹庆宇等人提出的模型。

关键词: 指代关系, 注意力机制, 显性指代, 零指代, 多层注意力机制模型

Abstract: In the information extraction process,the nondeterministic anaphora can cause incomplete information extraction.By analyzing the only discriminate information generated by the anaphoric part,the referenced part,the surrounding information,the referenced surrounding information and the original content in the current context,the anaphora relations are judged and a multilayer attention mechanism model is constructed.The probability calculation based on attention mechanism is performed on these five parts at different levels,and the final results are used to determine whether the anaphora relations can be proved or not.With the vectorization of the anaphoric part and the referenced part,the four probability calculations on two attention layers make every training result unique before judgment.Experimental results on OntoNotes 5.0 dataset show that the F value of the proposed model is 70.1% when both overt anaphora and zero anaphora are presented.When only zero anaphora are presented,the F value is 60.7%,which is higher than the model proposed by YIN Qingyu et al.

Key words: anaphora relations, attention mechanism, overt anaphora, zero anaphora, multilayer attention mechanism model

中图分类号: 

• TP18