Research On Deceptive Spam Review Detection Algorithm Based On Structural Attention Enhancement

With the continuous development of commenting and shopping platforms,consumers can publish and share real consumption experiences after shopping,which has also become one of the key reference factors for subsequent consumers to purchase the product or service.In order to expand their market share,some unscrupulous merchants have formed a large number of fake reviews by forging the content of the reviews.On the one hand,it misleads consumers in their decision-making,affecting their vital interests and increasing the risk of return and exchange? on the other hand,it creates a vicious competition environment for business platforms,which is not favourable to their long-term development.In order to effectively model and identify fake reviews,this paper uses the review text to build a HSEAN(Hierarchical Structure Attention Enpower Network)model.The model not only extracts rich semantic features,but also induces the generation of text structural dependency trees and traversal features for extracting tree structures.The research content is summarized as follows:(1)A network model HSEAN is proposed to embed the structural attention enhancement mechanism.Based on the hierarchical network of 'word-sentence-segment',the hierarchical embedding structure attention enhancement mechanism is used for the dependency learning of text units and the generation of post-order structural dependency trees.It can not only supplement the semantic complement of free word order units in the context of GRU,but also detect the consistency of the context of text units.In this way,it can be judged whether the reviewer maliciously edited the review by browsing the real comment when writing the text.The experimental results show that compared with other traditional machine learning classification algorithms based on feature mining and common neural networks models based on word embedding,the recognition model proposed in this paper can achieve the best performance.(2)A structure-dependent tree-induced generation algorithm for end-to-end training is constructed.For comment texts without third-party text analysis tools and additional domain annotation datasets,the text dependency matrix and root probability vector learned by the model are used to combine beam search and greedy search.Structural dependency trees that induce text in the data.(3)Four types of traversal features based on structural dependency tree are designed.Defines a set of traversal properties that describe the tree structure.Significant differences between fake review text and real text are extracted and analyzed.It further shows that the contextual consistency of fake reviews is poor,and for us to face datasets on different platforms,we only model the review text,and point out new feature extraction directions and methods.