A Quantum Interference Inspired Neural Matching Model For Ad-hoc Retrieval

Ad-hoc retrieval task is mainly to specify the user's information needs through a query,and then the information retrieval system will search for documents that may be related to the user's needs and return them to the user.With the development of deep learning,researchers have proposed many neural network matching models and achieved significant performance improvement in ad-hoc retrieval task.Existing neural matching models regard a word or n-gram fragment vector embedding as a matching unit,first locally interact each query matching unit with all matching units in the document for calculating the relevance features(for example,cosine similarity)into a matching feature matrix,then apply the neural network to extract the effective matching features and predict the document relevance.We find that this matching idea makes an independent assumption for the text matching unit,that is,the final relevance prediction made by the neural network matching model is obtained by first calculating the matching degree of the document under each query matching unit separately and then accumulating them.However,this assumption does not always hold in a real retrieval situations because of the dependencies between text matching units.Meanwhile,through formal analysis and comparison based on probability theory,we find that the matching idea of the current neural network matching models obeys classical probability,and has no ability to model the dependency matching features generated by the dependencies of different matching units between a query and document,which will cause a deviation of document relevance prediction and affect retrieval performance.In this work,we propose a Quantum Interference inspired Neural Matching model(QINM)under the framework of quantum probability,which can apply the interference phenomena naturally contained in quantum probability theory to effectively construct dependency matching features generated by different matching units between a query and document.Specifically,we firstly construct a query-document composite system,and then encodes the probability distribution of a document into the reduced density operator by observing the composite system.Then through a n-gram Window Convolution Network and Query Attention mechanism,we select the effective matching features in the operator.Finally,the scoring layer composed of Multi-Layer Perceptron(MLP)is then utilized to generate the final ranking score,and whole neural network is updated with the Learningto-rank algorithm.Experimental results on two benchmark collections demonstrate that our approach outperforms the retrieval models based on quantum probability theory,and some well-known neural retrieval models in ad-hoc retrieval task.