基于多任务深度学习的关键词生成方法

Keyword generation method based on multi-task deep learning

针对现有的关键词生成模型往往不能充分利用题目与关键词之间密切的关系预测关键词的问题,提出一种基于序列到序列的多任务注意力联合训练模型(Joint-MT)。将关键词生成任务作为主要任务,题目生成作为辅助任务;在目标函数上,除独立的多任务交叉熵损失,还添加一致性损失,加强多任务注意力机制之间的约束。实验结果表明,Joint-MT无论是在文内关键词预测还是在缺失关键词预测上都优于其它对比模型,说明Joint-MT模型能够增强任务之间的相互关系,提升关键词预测的效果。

To deal with the problem that existing keyphrase generation models often cannot make full use of the close connection between title and keyphrases to predict keyphrases,a sequence-to-sequence-based multi-task attention joint training model(Joint-MT)was proposed.Keyword generation was treated as the main task,and topic generation was treated as an auxiliary task.In the objective function,the multi-task cross-entropy loss was also added,the agreement-based loss was added to strengthen the constraints between the attention mechanisms of multi tasks.Experimental results show that Joint-MT is proven to be superior to other comparative models in terms of present keyword prediction and absent keyphrase prediction.Joint-MT can enhance the relationship between tasks and improve the effect of keyphrase prediction.