Dependency-directed Tree Kernel-based Protein-Protein Interaction Extraction from Biomedical Literature

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 10–19, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Dependency-directed Tree Kernel-based Protein-Protein Interaction Extraction from Biomedical Literature Longhua Qian Guodong Zhou Natural Language Processing Lab School of Computer Science and Technology, Soochow University 1 Shizi Street, Suzhou, China 215006 {qianlonghua,gdzhou}@suda.edu.cn Abstract Structured information plays a critical role in many NLP tasks, such as semantic relation ex- traction between named entities and semantic role labeling. This paper proposes a principled way to automatically generate constituent structure representation for tree kernel-based protein-protein interaction (PPI) extraction. The main idea behind our approach is that the critical portion in a constituent parse tree for PPI extraction can be automatically deter- mined by the shortest dependency path be- tween the two involved proteins, while other portion can be regarded as noise and ignored safely. Evaluation on multiple PPI corpora shows that our dependency-directed tree ker- nel-based method achieves promising results. This justifies the effectiveness of tree kernel- based methods for PPI extraction, in particular the advantage of dependency-directed con- stituent structure representation. 1 Introduction Since determining protein interaction partners is crucial to understand both the functional role of individual proteins and the organization of the entire biological process, there is a significant interest in protein-protein interaction (PPI) ex- traction. However, manual collection of relevant PPI information from thousands of biomedical research papers published every day (e.g. MED- LINE) is so time-consuming and labor- demanding that automatic extraction approaches with the help of NLP techniques become neces- sary. In principle, PPI extraction is much like the semantic relation extraction subtask (so called Relation Detection and Classification, RDC) defined by the ACE project (ACE, 2002-2007) in the newswire domain. Therefore, various kinds of machine learning methods have been borrowed from the newswire domain to the biomedical domain: feature-based methods (Mitsumori et al., 2006; Giuliano et al., 2006; Sætre et al., 2007; Liu et al., 2010) and kernel- based methods (Bunescu et al., 2005a; Erkan et al., 2007; Airola et al., 2008; Kim et al., 2010). Early studies on PPI extraction employ feature-based methods. However, the feature- based methods often fail to effectively capture the structured information, which is essential to identify the relationship between two proteins in a constituent or dependency-based syntactic representation. With the wide adoption of kernel-based methods to many NLP tasks, particularly for semantic relation extraction and semantic role labeling, various kernels such as subsequence kernels (Bunescu et al., 2005a) and tree kernels (Li et al., 2008) have been applied to PPI extraction. On one hand, dependency-based kernels, such as edit distance kernels (Erkan et al. 2007), graph kernels (Airola et al., 2008) and subsequence kernels (Kim et al., 2010), show some promising results for PPI extraction. This suggests that dependency information plays a critical role in PPI extraction, much like semantic relation extraction in the newswire narratives (Culotta and Sorensen, 2004; Bunescu et al., 2005b). On the other hand, while tree kernels based on constituent parse trees achieve …

Content truncated to 500 words