Toshiba BRIDJE at NTCIR-6 CLIR: The Head/Lead Method and Graded Relevance Feedback

摘要: At NTCIR-6 CLIR, Toshiba participated in the Monolingual and Bilingual IR tasks covering three topic languages (Japanese, English Chinese) one document language (Japanese). For Stage 1 (which is usual ad hoc task using new NTCIR6 topics), we submitted two DESCRIPTION runs TITLE for each language. Our first search strategy Selective Sampling with Memory Resetting, our second Head/Lead method, which uses run as of components data fusion. According to Relaxed Rigid Mean Average Precision statistics released by organisers, are top performer all six subtasks. 2 reused NTCIR-3, 4 5 test collections), repeated strategies order enable analysis across four collections. Moreover, conducted some unofficial true relevance feedback experiments exploiting graded provided automatic results show that method slightly but consistently improves performance, while “interactive” suggest graded-relevance metrics favour favours binary feedback. In addition, significance tests Japanese collection “harder” than previous