Personalized Abstractive Opinion Tagging

With the rapid development of e-commerce platforms,large volumes of reviews of products and services are published every day,which brings about an increasingly serious problem of information overload.In order to help users quickly understand the key information of products to facilitate their choices,many e-commerce platforms utilize abstractive summarization generation techniques to provide product description and tips to reflect the key characteristics of product reviews.However,the content contained in these text forms cannot take into account the integrity of extracted information and the order of key points.Therefore,the researchers proposed the task of abstractive opinion tagging.An opinion tag is a sequence of words on a specific aspect of a product or service,and opinion tags is a ranked list of tags that reflect key characteristics of product reviews.However,current models for opinion tagging can only generate opinion tags that reflects the popular preference,and ignores the personalized characteristics,even though personalization is an essential ingredient of engaging user interactions in e-commerce.To solve this problem,this thesis proposes a personalized abstractive opinion tagging task and designs a corresponding model.The task of personalized abstractive opinion tagging is mainly based on various data in ecommerce platforms,extracts the key information of the product to generate opinion tags,tracking user preferences to determine the order of opinion tags,so as to ensure that the key information of the product can be ranked according to user preferences.There are two main challenges when developing models for the end-to-end generation of personalized opinion tags:sparseness of reviews and difficulty to integrate multitype signals,i.e.,explicit review signals and implicit behavioral signals extracted from various data.To address these challenges,this thesis proposes a personalized opinion tagging framework,that consists of two user preference trackers and a opinion tags generator.In order to verify the effectiveness of the proposed model,this thesis evaluates it based on a real-world dataset collected from e-commerce platforms.The experimental results demonstrate that it significantly outperforms strong baselines in terms of generation metrics and ranking metrics.