Research Of Key Technologies For Context-Based Information Retrieval

Context-based information retrieval has attracted much attention in both academia and industry since it came into being.Regarding to the amount of information,it takes advantage of various contextual information,which effectively alleviates the problem caused by short queries and provides a better understanding of users' information needs.As to the ranking tenchnology,it provides more diverse matching methods.Specifically,in addition to the traditional keyword match,the match between the involved contextual information is also incorporated into retrieval.There are many types of contextual information,such as the device context,spatiotemporal context,user context,task context,and document context.Among all these contexts,the document context plays an important role in context-based information retrieval due to its more convenient access and richer content.However,there are still many deficiencies in the research of document context based information retrieval,such as the inaccurate semantic understanding and coarse-grained context modeling.Furthermore,most studies focus on incorporating the contextual information into traditional ranking models.Regarding to the neural networks which have been proved to be very effective for information retrieval in recent years,the integration with the contextual information has not been well studied due to the internal complex structures.Therefore,we focus on the document context in this paper.To alleviate the deficiencies mentioned above,we conduct further studies on the common dimensions of contextual information,namely topics,time,position,and alignment,and achieve significant improvements in both accuracy and granularity.In addition to adapting the traditional context-based retrieval models,we explore how to incorporate the contextual information into the neural network based ranking models,which significantly boost the retrival performance with respect to all evaluation measures.To be specific,the main contributions of this paper are as follows:1.A ranking method based on the topics of contextual snippets is proposed.This paper investigates the effectiveness of modeling queries with contextual snippets rather than individual words in the pseudo relevance feedback(PRF)documents.First,three context relevance metrics are explored to identify the high-quality contextual snippets that are relevant to the query.Then,a context-aware topic model is proposed.This model infers the topic distributions of the snippets from the corresponding PRF documents rather than the entire corpus,which makes the topic modeling more accurate.Finally,the topic distributions of the obtained snippets are integrated into the retrieval models,yielding a great improvement in the retrieval accuracy.2.A ranking framework based on the fine-grained time of the context is proposed.In addition to the traditional document-level temporal predictor,this framework also includes a word-level temporal predictor proposed in this paper,which can help capture more fine-grained temporal information.Compared with the existing time-based retrieval models,the models adapted within our ranking framework are more accurate and robust.3.A ranking model based on neural networks with context positional attention is proposed.It is the first attempt to incorporate the context positional information into the attention mechanism of recurrent neural networks,which is then applied for the task of answer selection in information retrieval.By modeling the context positional information of the question words in the high-dimensional hidden space,the bias caused by the accumulated semantics in the traditional attention mechanism can be alleviated,and the ranking performance of the candidate answers is significantly boosted.4.A ranking model based on context-aligned neural networks is proposed.This paper investigates the effectiveness of context-aligned information for similar sentence retrieval under the neural network framework.To be specific,a contextaligned recurrent neural network is proposed for ranking.When generating the hidden states for the current sentence,the model can automatically absorb the contextual information of the aligned words in the other sentence via the internal alignment gating.In this way,the deep interactions between the contexts in two sentences are implemented.At the same time,the irrelevant contexts will be reduced,which effectively boosts the accuracy of sentence similary ranking.