Affinity-Preserving Random Walk for Multi-Document Summarization Kexiang Wang, Tianyu Liu, Zhifang Sui and Baobao Chang Key Laboratory of Computational Linguistics, Ministry of Education School of Electronics Engineering and Computer Science, Peking University Collaborative Innovation Center for Language Ability, Xuzhou 221009 China {wke, tianyud421, sz£, chbb}@pku. edu Abstract Multi-docurnent summarization provides users with & short text that summarizes the information in a set of related doc- ‘uments. ‘This paper introduces: alinity- preserving random walk to the sumira- riration unk, which preserves the aflin= ity relations of sentences hy an absorh- ing random walk model. Meanwhile, we pul forwant adjustable attaity-preserving random walk to eaforce the diversity eon- strain of summarization in the random walk proces, The ROUGE evaluations on BUC 2003 topie-focused summarization ask and DUC 200% genes summarizat- tion tsk show the good pertirmance of cour method, which has the best ROUIGE- 2 recall among the graph-based_manking methods, 1 Introduction Mulidtocument stimmatization provides: users with summary that reflects the main information in a set af given documents. ‘The documents are ‘ofien related ane talk about more than one top- cs, Generic multiosument summarization and topic-focused multi-tocument summarization site ment cluster. ‘The sentences with litle informa- tion about the document cluster should not be im- ‘eluded in the suramary. The second one is te versity, The information overlap between sum mury yentcnces should be ax minimal a possic ‘ble duc te the length limit of summary. In other ‘words, the Information coverage of summary is a determinant, which requires that the summary seniences should cover diverse aspects of infor mation. Besides the two goals, thers is another goal for the topic-focused summarization and that is relevaney. [i requires thal the summary sen- tenees be relevant to the topic description, A se- ries of conferences and workshops or automtie Jext summarization (eg. NTCIR, DUC), special opie sexiiony ink ACL, EMNELP and SIGIR have achvunced the technicues to achieve these gonls net smuany approaches have been proposed 0 fr In this paper, we focus on the extractive summia~ ization methods, which extract the summary sem- fences from the input document cluster. ‘We pro- pose alfnity-preserving random walle for mulki- document suramarization. The metho! iy x graphs faved ranking methed, which tke into account the global infewnnation collectively computed from the ebtire seriicnce affinity graph. Dilferent from the previous eroph-based ranking methods, our method adopts “global normalization” to trans-