Semantic Definition Ranking

Question answering has been a focus of much attention from academia and industry. Search engines have already tried to provide direct answers for question-like queries. Among these queries, “What” is one of the biggest segments. Since results excerpted from Wikipedia often have a coverage problem, some models begin to rank definitions that are extracted from web documents, including Ranking SVM and Maximum Entropy Context Model. But they only adopt syntactic features and cannot understand definitions semantically. In this paper, we propose a language model incorporating knowledge bases to learn the regularities behind good definitions. It combines recurrent neural network based language model with a process of mapping words to context-appropriate concepts. Using the knowledge learnt from neural networks, we define two semantic features to evaluate definitions, one of which is confirmed to be effective by experiments. Results show that our model improves precision a lot. Our approach has been applied in production.