Simplicity is Better: Revisiting Single Kernel PPI Extraction

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 206–214, Beijing, August 2010 Simplicity is Better: Revisiting Single Kernel PPI Extraction Sung-Pil Choi Information Technology Laboratory Korea Institute of Science and Technol- ogy Information spchoi@kisti.re.kr Sung-Hyon Myaeng Department of Computer Science Korea Advanced Institute of Science and Technology myaeng@kaist.ac.kr Abstract It has been known that a combination of multiple kernels and addition of various resources are the best options for im- proving effectiveness of kernel-based PPI extraction methods. These supple- ments, however, involve extensive ker- nel adaptation and feature selection processes, which attenuate the original benefits of the kernel methods. This pa- per shows that we are able to achieve the best performance among the state- of-the-art methods by using only a sin- gle kernel, convolution parse tree kernel. In-depth analyses of the kernel reveal that the keys to the improvement are the tree pruning method and consideration of tree kernel decay factors. It is note- worthy that we obtained the perfor- mance without having to use any addi- tional features, kernels or corpora. 1 Introduction Protein-Protein Interaction (PPI) Extraction refers to an automatic extraction of the interac- tions between multiple protein names from nat- ural language sentences using linguistic features such as lexical clues and syntactic structures. A sentence may contain multiple protein names and relations, i.e., multiple PPIs. For example, the sentence in Fig.1 contains a total of six pro- tein names of varying word lengths and three explicit interactions (relations). The interaction type between phosphoprotein and the acronym P in the parentheses is “EQUAL.” A longer pro- tein name phosphoprotein of vesicular stomati- tis virus is related to nucleocapsid protein via “INTERACT” relation. Like the first PPI, nuc- leocapsid protein is equivalent to the abbre- viated term N. It is not straightforward to extract PPIs from a sentence or textual segment. There may be multiple protein names and their relationships, which are intertwined in a sentence. An interac- tion type may be expressed in a number of dif- ferent ways. Figure 1. An example sentence containing mul- tiple PPIs involving different names of varying scopes and relations1 A significant amount of efforts have been devoted to kernel-based approaches to PPI ex- tractions (PPIE) as well as relation extractions2 (Zhang et al., 2006; Pyysalo et al., 2008; Guo- Dong et al., 2007; Zhang et al., 2008; Airola et al., 2008; Miwa et al., 2009). They include word feature kernels, parse tree kernels, and graph kernels. One of the benefits of using a kernel method is that it can keep the original 1 BioInfer, Sentence ID:BioInfer.d10.s0 2 Relation extraction has been studied massively with the help of the ACE (www.nist.gov/tac) competition work- shop and its corpora. The ACE corpora contain valuable information showing the traits of target entities (e.g., ent- ity types, roles) for relation extraction in single sentences. Since all target entities are of the same type, protein name, in PPIE, however, we cannot use relational infor- mation that exists among entity types. This makes …

Content truncated to 500 words