Fine-Grained Named Entity Recognition in Question Answering with DBpedia

Named Entity Recognition in Question Answer tasks can help the language understanding. Traditional NER task usually solved by a supervised model, which need a large number of annotated corpora and extract context information as feature for training, however questions are usually short sentences, with little context information, little public annotated corpora. On the other hand coarse-grained named entities' types cannot offer enough information for question understanding. In this paper, we propose a novel model to address both problems, using a distant supervised method. Firstly, we use the web search to obtain more relevant information. Secondly, we present a greedy n-grams algorithm to extract the entity mentions. Finally, we use the kNN to classification and get fine-gained entity types combining with the entity mentions which can link with DBpedia. Experimental results show that our model outperforms various state-of-art systems in public dataset--TREC.