Discriminate and Reconstruct: Learning from Language Model to Answer Keyword Questions

We consider a new problem of question answering when the questions are in form of keywords, rather than natural language. While searching on machines or interacting with the robots, people usually prefer to raise a query by several keywords rather than a complete sentence. The new task of Keyword Question Answering (KQA) is challenging and significant because small variations to a question may completely change its semantical information, thus yields different answers. In this paper, we propose a simple but strong system for KQA composed of (1) a Keyword Question Discriminator to recognize the keyword questions, (2) a Question Reconstructor that learns from a language model to reconstruct the keyword questions and (3) a question answering model to produce answers. We further finetune the reconstructor via reinforcement learning by the quality of the answers to help generate answerable questions. Moreover, We also develop a semi-supervised learning method to build the keyword question datasets. Empirical results demonstrate the effectiveness of our method in comparison with various baselines, and we also find that the high layers of the language model are helpful in handling grammatical blunder and semantic fuzziness.