WCCL Relation — a Toolset for Rule-based Recognition of Semantic Relations Between Named Entities

Proceedings of Recent Advances in Natural Language Processing, pages 436–442, Hissar, Bulgaria, 7-13 September 2013. WCCL Relation — a Toolset for Rule-based Recognition of Semantic Relations Between Named Entities Michał Marci´nczuk Wrocław University of Technology michal.marcinczuk@pwr.wroc.pl Abstract In this paper we cover the problem of recognition of semantic relations between proper names (PNs) in running text. We focus on the manual rule creation ap- proach and discuss to what extent the ex- isting tools can be used for this task. As a result of our initial research we developed a rule-based toolset for recognition of re- lations between PNs called WCCL Rela- tion. The toolset is built on the top of WCCL Match — a language for text anno- tation, which is a part of a WCCL frame- work (an open source, released under the GNU LGPL 3.0). The WCCL Relation toolset is language independent and can be used for almost any natural language and language tagset. We present several use cases and sample rules for recognition of semantic relations in Polish texts. 1 Introduction Recognition of semantic relations between named entities is one of the information extraction ma- jor tasks. Its goal is to identify pairs of named entities (text fragments) connected other by a se- mantic relation on the basis of their context. In the majority of approaches the named entities are recognised beforehand and the task is limited to discovering and categorising connections between those entities. The list of possible relation cat- egories is unbounded and it depends on the de- sired application, the scope of the named enti- ties and the available resources. For example Marci´nczuk and Ptak (2012) defined 8 coarse- grained categories of semantic relations (location, origin, nationality, affiliation, neighbourhood, cre- ator, composition and alias). In turn Linguistic Data Consortium (2008) defined a set of 8 gen- eral relations (i.e., physical, part-whole, personal- social, organization-affiliation, agent-artifact and general-affiliation) with several subcategories. In the bioinformatic domain there are two common categories of relations between genes, proteins and associated entities — protein-component and subunit-complex (Pyysalo et al., 2011). There are two main approaches to relation recognition — construction of human-readable rules and construction of statistical models (ma- chine learning). According to Jiang (2012) the most common approach is the one based on the statistical models. There are also several rule-based approaches, like manual rule creation (Marciniak and Mykowiecka, 2007; Santos et al., 2010; Abacha and Zweigenbaum, 2011) and rule induction (Feldman et al., 2006; Brun and Hagège, 2009). The low interest in developing rule-based systems might be caused by a lack of robust and accessible tools for rule construction and execu- tion. For example, the well-known general frame- work GATE (Cunningham et al., 2011) does not support relation recognition within its rule formal- ism JAPE (Cunningham et al., 2000). Despite the manual rule creation is less pop- ular than the statistical approaches in the task of relation recognition, the rule-based approaches have several advantages over statistical-based ap- proaches. The first one is the traceability and full …

Content truncated to 500 words