Existing multi-document summarization methods ignore the fact that there often exist many relevant documents that provide surrounding (topical related) background knowledge, which can help generate a salient and discriminative summary for a given set of documents.

To run sumdocs, you need to write a config files like test_config_news and test_config_NLP. The most important parameters are $input_file, $background_file, $output_file and $summ_method.

python sumdocs.py $config_file

Test the Multi-News dataset

python sumdocs.py test_config_news

Test the Scientific-NLP dataset

python sumdocs.py test_config_NLP

Set summ_method in config files

Multi-facet summarization annotation protocol, details TBD.