Research On Question Answering Over Large Scale Knowledge Base With Hierarchical Memory Network

As more and more knowledge is stored in computers by people in a structured way,a large number of large-scale knowledge bases are formed.It has become an important and challenging task to fully utilize these knowledge bases to answer natural language problems.Although most existing knowledge base question answering systems can find the internal connection between the question and the relevant knowledge by modeling the problem and the knowledge base,they rely too much on semantic knowledge and feature engineering,which make it hard to migrate to large-scale knowledge bases.In this thesis,a memory network model with hieraraical structure is proposed by combining vector modeling and memory access technology.Firstly,we train the TransE model to get feature vectors of all entities and relationships,then initialize the memory module in each memory sub-network with these vectors and the generalization module.After the memory module is initialized,the problem input module can obtain the distribution probability of the memory in each layer by processing the problem.Then the policy module queries the memory module according to these distribution probability to find the most relevant supporting memory and send it to the answer module to get the final answer.By stratifying memory modules and vector mapping tasks,this thesis not only increases the capacity of memory model,but also improves the accuracy of question-answer model,which is is ideal for storing and querying large-scale knowledge bases.We performed our experiments on on two data sets,WebQuestions and SimpleQuestions.Then we compared with some other knowledge base question answer models in terms of F1-score and accuracy respectively.The results demonstrate that the proposed hierarchical memory network model can perform better than those models in the absence of semantic knowledge and feature engineering,both on a small scale and on a large scale.