基于知识增强的深度新闻推荐网络

在新闻推荐场景下，传统的基于文本特征的新闻推荐模型只考虑了词的共现关系，无法捕获词语的隐含词义和关联知识；而基于深度学习的推荐模型在融合知识图谱信息中仅仅考虑实体的信息，忽略了远距离实体之间的联系，造成实体之间的关联信息和深层次语义联系的缺失.针对该问题提出了一种基于知识增强的深度新闻推荐网络（deep knowledge-enhanced network，DKEN），利用长短期记忆网络提取知识图谱中的实体路径特征，补充到注意力网络中，然后针对不同的候选新闻动态地构建用户的特征.实验表明该实体路径信息能提高模型的效果，在F₁指标上提升大约1%. 

Deep Knowledge-Enhanced Network for News Recommendation

In the news recommendation scenario, the traditional text-based feature recommendation model only considers the co-occurrence relationship of words, and cannot capture the implicit meaning and associated knowledge of words.The recommendation model based on deep learning only considers the information of the entity in the process of merging the knowledge graph information, ignoring the connection between the distant entities, resulting in the lack of related information and deep semantic relations between entities.A model named deep knowledge-enhanced network (DKEN) was proposed to solve the problem.Firstly, a long-short-term memory network was used to extract the entity path features from the knowledge graph.And then, path features were added to the attention network and the user feature was built dynamically based on the candidate news.Finally, some experiments were carried out.The results show that the entity path features can improve the model''s effect and increase by about 1% on the F₁ indicator.