Markov Retrieval Model Based On Transferring Learning

In traditional information retrieval, the model is always trained on static dataset, which is lack of a mechanism that can adapt its structure and parameters to new data automatically. However, the web information is real-time updating in practice, so the model trained in one time may perform worse when applied to the new data sometime later. To solve the problem, a new model should be trained in the new data, but it may cost so much time and would also be a waste to throw away all the old data and the old model that has been trained for a long time. Thus, it's a critical transfer-learning problem how to make good use of old data to retrieval in the new data instead of retraining.In this paper, we propose a Markov Retrieval Model based on transfer learning (or TLMR model for short), which extends the traditional retrieval model based on Markov Network. It first constructs a retrieval model based on Markov Network on old data, because Markov Network can be considered as learning and updating mechanism, we then employ new data to update old Markov Network. This learning process corresponds to transferring knowledge learned from the old data to new situations. In this paper, we estimate the correlation of different datasets by measuring the KL-divergence between the two Markov networks, in order to decide the transfer ratio.Moreover, the web information is dynamically-changing rather than change only once, so it's worth exploring whether the model having been transferred once can still perform well when we transfer it to another data again. In our work, we do multi-step transferring with our TLMR model among several datasets, so as to investigate the model's adaptability to dynamic data.The one-step and multi-step transferring experiments on TREC datasets prove that the model can transfer well among several datasets. The multi-step transferring experiment also indicates that the learning ability of model has nothing to do with the learning sequence. New points in this thesis:1. When all the researches on transfer learning are about classification, we apply transfer learning theory in information retrieval. Knowledge transferring and supervised transfer learning are applied successfully in our retrieval model based on Markov Network, we employ the Markov Networks on both datasets to transfer knowledge.2. The TLMR model is proposed to transfer among several datasets based on transfer learning, in order to explore whether the old knowledge after transferred several times can still help retrieval in new task in condition that the data is represented by Markov Network. So as to investigate the the model's adaptability to dynamic data.3. The one-step and multi-step transferring experiments on TREC datasets prove that the model can transfer well among several datasets. The multi-step transferring experiment also indicates that the learning ability of model has nothing to do with the learning sequence, which complies with the rules of information processing in cognition..