基于层次型文本交互的表示学习研究

Research on Representation Learning Based on Hierarchical Text Interaction

传统的推荐系统面临着诸如数据稀疏性、无法解释的推荐等几个挑战.为了解决这些问题,许多研究通过挖掘评论文本语义信息来提高推荐性能.然而,这些方法在文本特征建模和文本交互方面存在问题.在文本建模方面,它们简单地将用户/物品的所有评论拼接成一个单一的评论.然而,单词/短语级别的语义信息可能与评论文本的整体语义信息相悖.在文本交互方面,它们将交互推迟到预测层,无法捕捉用户和物品之间复杂的相关性.为了解决这些问题,我们提出了一种新颖的基于层次型文本交互的表示学习方法.在该方法中,我们以层级方式对低级单词语义和高级评论文本进行建模,以便在不同粒度上挖掘文本信息.为了进一步捕捉复杂的用户-物品的交互关系,我们提出在不同层次上挖掘用户-物品之间的语义关联.在单词级别上,我们提出了一种针对每对用户-物品个性化的注意力机制,来捕捉表示每个评论的重要单词.在文本级别上,我们在用户和物品之间相互传播文本语义信息,并捕捉针对目标任务有用的评论文本.最后,我们通过协同过滤框架,将该方法应用于评分预测应用场景,并通过在公开数据集上的对比实验,证明该方法在评分预测方面的性能优于现有方法.

Review-based recommendations have the limitations of monotonous textual representation and deferred user-item interactions.While previous research has exploited the textual semantics to alleviate the challenges,certain limitations still exist.Specifically,for textual modeling,they simply concatenate all the reviews of a user/item into a single review.However,feature extraction at word/phrase level can violate the meaning of the original reviews.As for textual interactions,they defer the interactions to the prediction layer,causing the failure to capture complex correlations between users and items.To address these limitations,an innovative Hierarchical Interaction(HTI)model is presented for the representation learning tasks.Within the HTI framework,a hierarchical representation of both low-level word semantics and high-level review content is presented.This hierarchical approach enables the exploitation of textual attributes across various levels of granularity.Furthermore,to more comprehensively capture complex user-item interactions,harnessing semantic correlations across different hierarchy levels for each user-item pair is made.At the word level,a dedicated attention mechanism tailored to individual user-item pairs is presented,facilitating the identification of pivotal words essential for accurately representing each review.At the review level,the mutual propagation of textual features between users and items is made,effectively capturing informative review instances.The learned representations are seamlessly integrated into a collaborative filtering framework for rating prediction task.Experiments on public datasets demonstrate the superiority of the proposed model.