Beyond Utterance Extraction: Summary Recombination for Speech Summarization

This paper describes a template filling approach for creating conversation summaries. The templates are generated from generalized summary fragments from a training corpus. Necessary pieces of information for filling them are extracted automatically from the conversation transcripts given linguistic features, and drive the fragment selection process. The approach obtains ROUGE-2 scores of 0.08471 on the RATP-DECODA corpus, which represents a significant improvement over extractive baselines and hand-written templates.