부산대학교 도서관

This research studies the problem of email overload and proposes a system that automatically detects whether the email is "to read" or "to do". The goal of our research is to test if a Language Model trained on the entire Enron email data (around 500,000 emails) then retrained on a subset of intent labeled emails (i.e. transfer learning) would outperform an LSTM model trained on the intent email dataset only. We conduct several experiments using transfer learning in three different scenarios: 1) using pre-trained Language Model as features extractor for SVC. 2) retrain all layers of the pre-trained Language Model on the intent email dataset. 3) gradually retraining the pre-trained Language Model layer by layer, starting from the last layer. We evaluate the results on a subset of Enron Email Dataset. Comparing with a baseline model consists of two LSTM layers, we were able to increase the accuracy from 85.42% to 88.43%.