Persuasive Product Snippet Generation

With the rapid development of computer t,echnology,people are trying to make the machine more humanized.Therefore,this paper we study the problem of how to make machines intelligently generate a persuasive Natural Language de-scription that both conveys product information and delivers explanations related to user needs.This problem might benefit from current large amount of research work on end-to-end deep neural networks.However,the success of deep neural networks is due to the support of massive training data.We are unable to obtain large-scale and persuasive text descriptions.Therefore,the lack of labelled data and subjective judgments still pose a severe challenges to training such a model.In response to the above problems,we divide the system into two parts.The first part is the weak supervision framework.By analyzing the rhetorical,lexical and grammatical features contained in persuasive t.exts in external data sources,we have written a set of rules with high covera.ge and low accuracy.The weak supervision framework automatically labels unlabeled dat,a based on these rules.The training data with proba.bility labels is genera.ted to obta.in the tra.ining data of the subsequent deep model.It is a data level solution.The second part is the framework of generating model.In order to strengthen the relationship between user consumption scenarios and product attributes,We explore the way of representation based on knowledge graph to integrate the knowl-edge obtained from heterogeneous information sources.This is a knowledge level solution.At the model level,we designed the encoder-decoder framework.At the encoder layer,we proposed the Global-Local module to overcome the weak super-vision problem and the dependency between the scene and the product.At the decoder layer,we added the Copy mechanism to better solve the OOV problem and make the generated text description smoother.The comparison with a variety of comparison methods demonstrates the effec-tiveness and superiority of our system and demonstrates the possibility of making the machine work like a skilled salesperson.