1	Understanding Documents via Concept Links
2	Outline DUC 2005 System Task Targeted Sentences Our Approach System Overview Concept Link Sentence Similarity Sentence Ranker: A modified MMR Evaluation Conclusions
3	DUC 2005 System Task Task Definition in [Amigo et al, 04] … topic-oriented, informative multi-document summarization, … compressed version of a set of documents … Topic Creation Instructions to formulate a topic out of interesting aspects “At least 25 documents must each contribute some material to the answer” of a quest of the topic Our view of the task A general, and topic-oriented summary.
4	Targeted Sentences Good DUC 2005 summary: an extract consists of sentences that highly representative highly relevant to the topic General Specific: named entities are favored with minimal redundancy
5	System Overview
6	Concept Detection
7	Concept Link There exists a Concept Link between each pair of similar concepts Concept Similarity: maximal sense overlapping (Banerjee et al, 2003) Consider all senses of each concept Extended sense Sx: Synset + Gloss + hypernymy + meronymy set(1 level)
8	Concept Link Detection 1) A year ago Mr Douglas Hurd foreign secretary became the first UK cabinet minister to visit Argentina since the 1982 Falkland islands conflict. 2) Today Argentina gets out the red carpet for the UK Duke of York the first official royal visitor since the end of the Anglo Argentine Falklands war in 1982.
9	Concept Links between sentences
10	Sentence Similarity Sum of “strength” of concept links
11	Sentence ranker Original Weight: Representative Power
12	Sentence ranker MMR modified
13	Evaluation: ROUGE
14	Evaluation: Pyramid
15
16	Experiments:
17	Conclusions: A simple system features Concept Link: new way to calculate sentence similarity; no chunker/parser involved concept differs from NPs in Lexical Chain Considering sentence similarity/relatedness via Concept Link: Alleviate the influence of expression variations; (but might involve inaccurate sense guess) Outperforms Word co-occurrence approach Minimizing Redundancy via Modified MMR; No extra heuristics involved.
18	Future work Error analysis; How to automatically set parameters; Comparison with alternative Similarity Measures; How about more knowledge (syntactic, semantic parsers …)? …