A Generative Adversarial Net Assisted Method for User Intention Recognition on Imbalanced Dataset

Abstract

Class imbalance is a classic problem in machine learning and deep learning. This problem bothers recognizing user intention in knowledge graph question answering domain. Some intentions would not be recognized well when the training set is unbalanced. Sampling and weighted loss methods are widely used to address it. However, these methods pay little attention to the form of question texts in feature space. We propose a novel method based on generative adversarial net to deal with imbalanced class problem, which firstly learns how a user question text of a certain class is represented in feature space and then generates samples. Our method designs a model that contains three parts: text encoder, generator, and discriminator. The text encoder transforms user question text into words vectors. The generator generates vectors that would converge to expected classes and the discriminator is responsible for recognizing classes. We split the training period into two stages. In the first stage, the discriminator is going to distinguish between samples from encoder and generator. In the second stage, the generator learns data distribution and generates related samples to enhance the ability of encoder. Experimental results show that our method can balance the model performance on each class and significantly outperforms traditional methods on unbalanced datasets for user intention recognition.