Regression-Based Summarization of Email Conversations

In this paper we present a regression-based machine learning approach to email thread summarization. The regression model is able to take advantage of multi- ple gold-standard annotations for training purposes, in contrast to most work with binary classiers. We also investigate the usefulness of novel features such as speech acts. This paper also introduces a newly created and publicly available email corpus for summarization research. We show that regression-based classiers per- form better than binary classiers because they pre- serve more information about annotator judgements. In our comparison between dierent regression-based classiers, we found that Bagging and Gaussian Pro- cesses have the highest weighted recall.