Combining Tree Structures, Flat Features and Patterns for Biomedical Relation Extraction

Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, pages 420–429, Avignon, France, April 23 - 27 2012. c⃝2012 Association for Computational Linguistics Combining Tree Structures, Flat Features and Patterns for Biomedical Relation Extraction Md. Faisal Mahbub Chowdhury † ‡ and Alberto Lavelli ‡ ‡ Fondazione Bruno Kessler (FBK-irst), Italy † University of Trento, Italy {chowdhury,lavelli}@fbk.eu Abstract Kernel based methods dominate the current trend for various relation extraction tasks including protein-protein interaction (PPI) extraction. PPI information is critical in un- derstanding biological processes. Despite considerable efforts, previously reported PPI extraction results show that none of the approaches already known in the literature is consistently better than other approaches when evaluated on different benchmark PPI corpora. In this paper, we propose a novel hybrid kernel that combines (auto- matically collected) dependency patterns, trigger words, negative cues, walk fea- tures and regular expression patterns along with tree kernel and shallow linguistic ker- nel. The proposed kernel outperforms the exiting state-of-the-art approaches on the BioInfer corpus, the largest PPI benchmark corpus available. On the other four smaller benchmark corpora, it performs either bet- ter or almost as good as the existing ap- proaches. Moreover, empirical results show that the proposed hybrid kernel attains con- siderably higher precision than the existing approaches, which indicates its capability of learning more accurate models. This also demonstrates that the different types of in- formation that we use are able to comple- ment each other for relation extraction. 1 Introduction Kernel methods are considered the most effective techniques for various relation extraction (RE) tasks on both general (e.g. newspaper text) and specialized (e.g. biomedical text) domains. In particular, as the importance of syntactic struc- tures for deriving the relationships between en- tities in text has been growing, several graph and tree kernels have been designed and experi- mented. Early RE approaches more or less fall in one of the following categories: (i) exploitation of statis- tics about co-occurrences of entities, (ii) usage of patterns and rules, and (iii) usage of flat features to train machine learning (ML) classifiers. These approaches have been studied for a long period and have their own pros and cons. Exploitation of co-occurrence statistics results in high recall but low precision, while rule or pattern based ap- proaches can increase precision but suffer from low recall. Flat feature based ML approaches em- ploy various kinds of linguistic, syntactic or con- textual information and integrate them into the feature space. They obtain relatively good results but are hindered by drawbacks of limited feature space and excessive feature engineering. Kernel based approaches have become an attractive alter- native solution, as they can exploit huge amount of features without an explicit representation. In this paper, we propose a new hybrid kernel for RE. We apply the kernel to Protein–protein interaction (PPI) extraction, the most widely re- searched topic in biomedical relation extraction. PPI1 information is very critical in understanding biological processes. Considerable progress has been made for this task. Nevertheless, empirical …

Content truncated to 500 words