Topic-Aware Networks for Answer Selection

Answer selection is an essential task in the study of natural language processing, which is involved in many applications such as a dialog system, reading comprehension, and so on. It is a task of selecting the correct answer from a set of given candidates for certain questions. One of the challenging problems for this task is that traditional deep learning model for answer selection lacks real-world background knowledge, which is crucial for answering questions in real-world applications. In this paper, we propose a set of deep learning networks to enhance the traditional answer selection models with topic modeling, so that we could use topic models as external knowledge for the baseline models and improve the performance of the model. Our topic-aware networks (TANs) are specially designed for answer selection task. We proposed a novel method to generate topic embedding for both questions and answers separately. We designed two kinds of TAN models and evaluate our models in two commonly used answer selection datasets. The results verify the advantages of TAN in improving the performance of traditional answer selection deep learning models.