Learning from numerous untailored summaries

We present an attempt to use a large amount of summaries contained in the New York Times Annotated Corpus (NYTAC). We introduce five methods inspired by domain adaptation techniques in other research areas to train our supervised summarization system and evaluate them on three test sets. Among the five methods, the one that is trained on the NYTAC followed by fine-tuning on the target data (i.e. the three test sets; DUC2002, RSTDTBlong and RSTDTBshort) performs the best for all the test sets. We also propose an instance selection method according to the faithfulness of the extractive oracle summary to the reference summary and empirically show that it improves summarization performance.