| The language model is a basic research topic in many artificial intelligence fields.In recent years,inspired by the quantum theory,a quantum language model has been proposed in information retrieval field.The quantum language model utilizes the density matrix in quantum theory to model term dependencies in the sentences.Compared with the traditional n-gram language model,the quantum language can model the dependencies arrange any length without increasing the scale of the model parameters.However,in the quantum language model,the representation for each word is a one-hot vector,which only encode the local dependencies.The density matrix in the quantum language model is estimated via an iterative process and the quantum language model deals representation and matching separately.Thus the quantum language cannot be jointly optimized,which limits its applicability to the related research areas.In order to broaden the theoretical and practical basis of the quantum language model,this paper presents the neural network based quantum language models and applies them into question answering task.A sentence can be regarded as a quantum system.In the neural network,the density matrix representation of the sentence is construct based word embeddings,and then the joint representation of the QA pair is obtained according to the density matrix representation.The features are extracted by the convolutional neural network from the joint representation to measure the similarity between the question and the answer.Compared with the original quantum language model,these models,which are the end-to-end models,can make full use of the advantages of big data,and have great improvement on the effect.The experimental results of TREC-QA and WIKI-QA datasets verify the validity of the models. |
