Joint Learning from Labeled and Unlabeled Data for Information Retrieval

Proceedings of the 27th International Conference on Computational Linguistics, pages 293–302 Santa Fe, New Mexico, USA, August 20-26, 2018. 293 Joint Learning from Labeled and Unlabeled Data for Information Retrieval Bo Li, Ping Cheng, Le Jia School of Computer Science Central China Normal University Wuhan, China libo@mail.ccnu.edu.cn Abstract Recently, a significant number of studies have focused on neural information retrieval (IR) mod- els. One category of works use unlabeled data to train general word embeddings based on term proximity. The general embeddings can be integrated into traditional IR models. The other category employs labeled data (e.g. click-through data) to train end-to-end neural IR models consisting of layers for target-specific representation learning. The latter idea accounts better for the IR task and is favored by recent research works, which is the one we will follow in this paper. We hypothesize that general semantics learned from unlabeled data can complement task-specific representation learned from labeled data of limited quality, and that a combination of the two is favorable. To this end, we propose a learning framework which can benefit from both labeled and more abundant unlabeled data for representation learning in the context of IR. Through a joint learning fashion in a single neural framework, the learned representation is optimized to min- imize both the supervised loss on query-document matching and the unsupervised loss on text reconstruction. Standard retrieval experiments on TREC collections indicate that the joint learn- ing methodology leads to significant better performance of retrieval over several strong baselines for IR. 1 Introduction In recent years, the research community has noticed the great success of neural networks in computer vision (Krizhevsky et al., 2012), speech recognition (Hinton et al., 2012) and natural language processing (Mikolov et al., 2013) tasks. However, the potential of neural networks has not been fully investigated in the IR field. Although a significant number of studies (e.g. (Huang et al., 2013; Ganguly et al., 2015; Zheng and Callan, 2015; Guo et al., 2016; Zamani and Croft, 2016; Dehghani et al., 2017; Mitra et al., 2017)) try to apply neural networks in IR, there have been few studies reporting the performance that is comparable to state-of-the-art IR models. These approaches rely on the general idea that neural network can provide a low-dimensional and semantics-rich representation for both queries and documents. Such a representation can bridge lexical and semantic gaps in traditional IR models. Depending on if the embeddings are trained with discriminative information for IR tasks, existing works can be broadly divided into two categories (Zhang et al., 2016; Mitra and Craswell, 2017). The first category of approaches extend traditional IR models to incorporate word embeddings that are trained on huge and unlabeled corpora with existing models such as Word2vec (Mikolov et al., 2013) and GloVe (Pennington et al., 2014) in an unsupervised manner. These approaches (e.g. (Zheng and Callan, 2015; Nalisnick et al., 2016)) leverage semantic information captured by word embeddings in order to enhance traditional IR models. We note that such models trained without references …

Content truncated to 500 words