Using Complex Networksfor Language Processing: The Case ofSummary Evaluation

Theability toaccess embeddedknowledge makes complex networks extremely promising fornatural language processing, whichnormallyrequires deep knowledge representation thatisnotaccessible withfirst-order statistics. In thispaper,we demonstrate thatfeatures ofcomplex networks, whichhavebeenshowntocorrelate withtext quality, canbeusedtoevaluate summaries. Themetrics are theaverage degree, cluster coefficient, andtheextent towhich thedynamics ofnetwork growthdeviates fromastraight line. Theywerefoundtobemuchsmaller forthehigh-quality, manualsummaries, andincreased forautomatic summaries, thuspointing toaloss ofquality, asexpected. We alsodiscuss thecomparative performance ofautomatic summarizers