Labeling consecutive search query pairs using siamese networks

As internet users interact with search engines to meet their information needs, a huge amount of search queries are stored. Proper analysis of such query data enhances prediction and understanding of user tasks. User tasks can be used to increase the performance of search engines and recommendations. Query segmentation is an initial step that is commonly performed while analyzing user queries. It determines whether two consecutive query expressions belong to the same sub-Task. Any deficits in query segmentation process is likely to affect all other advanced query based steps and activities like task identification and query suggestion. Recently, some researchers focused on application of algorithms including Recurrent Neural Networks (RNN) to seek for the semantics of queries, and attention based neural networks, but such methodologies are not task-specific. In this paper, we propose a Siamese Convolutional Neural Network (CNN) that models input queries into a more task-specific embedding and a decider network that does the labelling. The proposed method is compared with Context Attention Based Long Short Term Memory (CA-LSTM) and Bi-RNN Gated Retified Unit (GRU) models on Webis Search Mission Corpus 2012 (WSMC12) and Cross-Session Task Extraction (CSTE) datasets. The proposed model performs 95%, implying a 1% improvement over the already existing models and an accuracy of 81% on CSTE dataset implying an improvement classification accuracy of 6% over the previous best results.