Off-Topic Text Detection Based on Neural Networks Combined with Text Features

Abstract

Off-topic text detection refers to detecting whether the topic of a text deviates from the required topic. Our work addresses the problem of predicting whether a text is off-topic or not for a given prompt. Prior studies use classical content vector to represent the text and predict it with machine learning based methods. To our knowledge, there are few studies investigated this task using neural networks. In this paper, we propose to use a combination of neural features and surface text features as a representation of a text. Then we input the hidden representation of the text into a softmax layer to get the probability of the prediction. We do several experiments on four datasets. The experimental results show that our method achieves better performance compared with the baseline method.