Domain Adaptation to Summarize Human Conversations

We are interested in improving the summarization of conversations by using domain adaptation. Since very few email corpora have been annotated for summarization purposes, we attempt to leverage the labeled data available in the multiparty meetings domain for the summarization of email threads. In this paper, we compare several approaches to supervised domain adaptation using out-ofdomain labeled data, and also try to use unlabeled data in the target domain through semi-supervised domain adaptation. From the results of our experiments, we conclude that with some in-domain labeled data, training in-domain with no adaptation is most effective, but that when there is no labeled in-domain data, domain adaptation algorithms such as structural correspondence learning can improve summarization.