Sentiment Analysis Based on Bi-Dual Inference

In recent years, machine learning has achieved rapid development in solving artificial intelligence tasks in many fields such as translation and image. Dual learning as a new learning paradigm, the core idea is to use primitive tasks (mapping from domain X to domain Y ) and dual tasks (from domain Y to domain X). Mapping to improve the performance of the two tasks. Such as translation from one language to another and reverse translation, emotion recognition and sentence generation, etc. The importance of this duality is shown by these studies that it can promote the learning of two tasks in the dual form. The dual inference is proposed on the basis of dual learning to study how to use structural duality in the reasoning stage of artificial intelligence tasks to improve the performance of the model. In this article, we propose a general Bi-dual inference framework, through the use of different initialization and data access order random seed independent training to obtain different agent models ( f, g ), so that multiple f or g Perform reasoning on a single task. We apply it to five-element sentiment classification, finally, the effectiveness of the model is verified on the data sets NLPCC2014 and NLPCC2018, and good results are obtained.