Case Study of Model Adaptation: Transfer Learning and Online Learning

Many NLP tools are released as programs that include statistical models. Unfortunately, the models do not always match the documents that the tool user is interested in, which forces the user to update the models. In this paper, we investigate model adaptation under the condition that users cannot access the data used in creating the original model. Transfer learning and online learning are investigated as adaptation strategies. We test them on the category classification of Japanese newspaper articles. Experiments show that both transfer and online learning can appropriately adapt the original model if the dataset for adaptation contains all data, not just the data that cannot be well handled by the original model. In contrast, we confirmed that the adaptation fails if the dataset contains only erroneous data as indicated by the original model.