- Abstract: Classifying question sentences into their corresponding categories is an important task with wide applications, for example in many websites' FAQ sections. However, traditional question classification techniques do not fully utilize the well-prepared answer data which has great potential for improving question sentence representations which could lead to better classification performance. In order to encode answer information into question representation, we first introduce novel group sparse autoencoders which could utilize the group information in the answer set to refine question representation. We then propose a new group sparse convolutional neural network which could naturally learn the question representation with respect to their corresponding answers by implanting the group sparse autoencoders into the traditional convolutional neural network. The proposed model show significant improvements over strong baselines on four datasets.
- Conflicts: oregonstate.edu, us.ibm.com