An End-to-End Opinion Mining Model With Weak Supervision

User Generate Reviews(UGR)widely exist in Internet platform,such as social media and e-commerce websites.These reviews contain tremendous information and potential application value.For example,users can get consumer guidance through the opinions of other users;enterprises or businesses can get guidance of business decisions by analyzing user feedback.With the development of Natural Language Processing(NLP)technology,researchers have proposed many methods to automatically summarize user opinions from large-scale review data.A complete opinion in an opinionated review includes two parts: product features which are entities or attributes of entities,and corresponding sentiment polarities.The extraction of these two parts are covered by two NLP tasks,which are Aspect Extraction and Aspect-level Sentiment Analysis.Most existing approaches have two disadvantages in common.Firstly,they deal these two subtasks separately.Some approaches only focus on single subtask and yield incomplete opinions,which may not provide sufficient information.Some approaches jointly model two tasks.However,they usually take pipeline manner,which ignore the relevance between two subtasks.Secondly,aspect-level annotated data require much manual effort,which makes many supervised approaches suffer from domain adaption.In address to these problems,this paper proposed two models for each of two subtasks.Then we combine two models and proposed an end-to-end opinion mining model with weak supervision.To summarize,our effort is in three folds:(1)Research on unsupervised Aspect Extraction.Topic Model is the popular approach of this task,inspired by which we proposed to deal this problem with VAE-style neural topic model instead of LDA-style conventional topic model.Considering the specialty of this task,we proposed Multi-Space Neural Topic Model(MS-NTM),which model multiple latent variables in distinct probabilistic spaces so that it effectively alleviate the noisy semantic in topic modelling.(2)Research on Aspect-level Sentiment Analysis.Firstly,we analyze the advantages and disadvantages of two popular mechanisms of this task,which are attention mechanism and gated mechanism.Then,we proposed GatedAttentive LSTM(AG-LSTM)which combine attention and gated mechanisms so that it capture key features and filter noisy features simultaneously.Besides,our proposed approach has strong expansibility and compatibility.(3)Construction of an end-to-end opinion mining model.After the superiority of MS-NTM and Ag-LSTM has been proven,we combines the two models and propose an end-to-end opinion mining model(OMNET)with weak supervision.The experimental results have demonstrated that our proposed model significantly outperforms strong baseline methods under weak supervised training,and is close to or even better than the strong baseline methods under supervised training.Besides,due to the characteristics of weak supervision and end-to-end,the proposed model has good domain migration and practicability.