Sentence Ordering in Multidocument Summarization

Sentence Ordering in Multidocument Summarization Regina Barzilay Computer Science Department 1214 Amsterdam Ave New York, 10027, NY, USA regina@cs.columbia.edu Noemie Elhadad Computer Science Department 1214 Amsterdam Ave New York, 10027, NY, USA noemie@cs.columbia.edu Kathleen R. McKeown Computer Science Department 1214 Amsterdam Ave New York, 10027, NY, USA kathy@cs.columbia.edu ABSTRACT The problem of organizing information for m ultido cumen t summarization so that the generated summary is coheren t has receiv ed relativ ely little atten tion. In this pap er, w e describ e t w o naiv e ordering tec hniques and sho w that they do not p erform w ell. W e presen t an in tegrated strategy for ordering information, com bining constrain ts from c hronolog- ical order of ev en ts and cohesion. This strategy w as deriv ed from empirical observ ations based on exp erimen ts asking h u- mans to order information. Ev aluation of our augmen ted algorithm sho ws a signi�can t impro v emen t of the ordering o v er the t w o naiv e tec hniques w e used as baseline. 1. INTRODUCTION Multido cumen t summarization p oses a n um b er of new c hallenges o v er single do cumen t summarization. Researc hers ha v e already in v estigated issues suc h as iden tifying rep eti- tions or con tradictions across input do cumen ts and deter- mining whic h information is salien t enough to include in the summary [1, 3, 6, 11, 15, 19]. One issue that has receiv ed little atten tion is ho w to organize the selected information so that the output summary is coheren t. Once all the rel- ev an t pieces of information ha v e b een selected across the input do cumen ts, the summarizer has to decide in whic h order to presen t them so that the whole text mak es sense. In single do cumen t summarization, one p ossible ordering of the extracted information is pro vided b y the input do cu- men t itself. Ho w ev er, [10] observ ed that, in single do cumen t summaries written b y professional summarizers, extracted sen tences do not retain their precedence orders in the sum- mary . Moreo v er, in the case of m ultiple input do cumen ts, this do es not pro vide a useful solution: information ma y b e dra wn from di�eren t do cumen ts and therefore, no one do cumen t can pro vide an ordering. F urthermore, the order b et w een t w o pieces of information can c hange signi�can tly from one do cumen t to another. W e in v estigate constrain ts on ordering in the con text of m ultido cumen t summarization. W e �rst describ e t w o naiv e ordering algorithms, used in sev eral systems and sho w that they do …

Content truncated to 500 words