Extractive Summarisation via Sentence Removal: Condensing Relevant Sentences into a Short Summary Marco Bonzanini, Miguel Martinez-Alvarez, and Thomas Roelleke Queen Mary University of London SIGIR '13
Introduction u The main contribution of this paper is the definition of an algorithm for sentence removal, developed to maximise the score between the summary and the original document. u Instead of ranking the sentences and selecting the most important ones, the algorithm iteratively removes unimportant sentences until a desired compression rate is reached. 2
Extractive Summarisation (1/4) u The task of extractive summarisation is to select the subset of sentences and to combine them into a summary which better represents the topic. u In order to form the summary, a length limit has to be considered, based on the number of sentences or the number of words. 3
Extractive Summarisation (2/4) 4
Extractive Summarisation (3/4) 5
Extractive Summarisation (4/4) 6
Sentence Selection (1/3) 7
Sentence Selection (2/3) 8
Sentence Selection (3/3) 9
Sentence Removal 10
Sentence Removal (Cont’d) 11
SR Algorithm 12
Opinosis Dataset u The Opinosis dataset is a collection of opinion- oriented data, divided into 51 different topics. u Each topic includes a number of sentences (min. 50, max. 575, avg. 139), taken from different reviews from popular review web sites. u For each topic, 4 or 5 golden standard (human- written) summaries are provided. u The golden standard summaries hence present the pivot opinion for each topic, in a concise way (approx. 2 sentences each). u For this reason, the maximum length of the system generated summaries is fixed to two sentences. 13
ROUGE Framework u The ROUGE framework is used to provide a quantitative assessment between the candidate summaries and the golden standards. u Specifically, the results for ROUGE-1, ROUGE-2, and ROUGE-SU4 are reported. u This study also reports the results for MEAD, a state-of-the-art extractive summariser based on cluster centroids. u The best overall results are shown in bold, and the best results within the same scoring function are shown in italic. u Best results labelled with a † show that second-best results are outside their 95% confidence interval. 14
Results for Recall ROUGE-1ROUGE-2ROUGE-SU4 MEAD49.32 † †
Results for Precision ROUGE-1ROUGE-2ROUGE-SU4 MEAD
ROUGE-1ROUGE-2ROUGE-SU4 MEAD
ROUGE-1ROUGE-2ROUGE-SU4 MEAD †07.54 †08.28 †
Conclusions 19
THANKS SIGIR’13, July 28–August 1, 2013, Dublin, Ireland. Copyright 2013 ACM /13/07...$