High precision rule based PPI extraction and per-pair basis performance evaluation

Abstract

Virtually all current PPI extraction studies focus on improving F-score, aiming to balance the performance on both precision and recall. However, in many realistic scenarios involving large corpora, one can benefit more from an extremely high precision PPI extraction tool than a high-recall counterpart. We also argue that the current "per-instance" basis performance evaluation method should be revisited. In order to address these problems, we introduce a new rule-based PPI extraction method equipped with a set of ultra-high precision extraction rules. We also propose a new "per-pair" basis performance metric, which is more pragmatic in practice. The proposed PPI extraction method achieves 95-96% per-pair and 94-97% per-instance precisions on the AIMed benchmark corpus.