The Q & A system is a kind of system which can answer user’s questions with accurate and natural language. Some improvement measures have been tried for “named entity recognition”. Aiming at the time- and labor-consuming problem of traditional one-way template matching， this paper proposes a lattice bi-directional structure of long short-term memory （Lattice Bi-LSTM）network， which solves the problems of improper sentence processing and dependence on the result of word segmentation in named entity recognition. Compared with the unidirectional structure， the bi-directional structure can make better use of sentence information and make the output more robust， thus capturing semantic information more accurately. To solve the problem of non-linear coupling of similarity between entities in traditional methods， a method is proposed to link “similar” entities to the knowledge base accurately by using periodic kernel function. The two improved methods are verified by experiments， whose results show that they have significant improvement effects compared with the classical method.