Intelligent Question Answering System is one of the artificial intelligent project, which belongs to both nature language process and information retrieval area. There are two different intelligent question answering system separated by focus task, one is open domain, the other is close domain. This thesis applied some research achievements on close domain intelligent question answering system. These work had impored the accuracy of question answering test.The models proposed in this thesis are mixture vector model based on convolutional neural network, and learning to rank model. The mixture vector model mainly used two technologies, one is word embedding and the other is convolutional neural network which target is max margin. The word embedding technology is derived by Google open source project word2vec which proposed in2013. And the convolitional neural network coded by python after algorithm deduced. The mixture vector model contain two sub-models which are question vector generate model and answer vector generate model. Both two models are not specific to any target, but maximise the vectors which belong to different classes. When training the answer vectors, answer vector generate model will also training the question vectors. These training steps will eventually be collaborative optimal.Learning to rank model is the model which use machine learning method to solove rank targets. There are three kinds of learning to rank model, first one is point wise model which treat single document as training object, second one is pair wise model which use document pair’s relationship to training object, the last one is list wise model, this model maintain whole document list as optimize object. We choose the listNet model as learning to rank model to optimize current model’s performance, which based on list wise model. In this thesis, we choose mixture vector model, key words, fuzzy match result as features to liseNet model.By projected and coded these two model, the answer retrieve accuracy of the intelligent question and answering system had a remarkable improvement, and thus we prove these functions work effectively in engineering field. |