Proper and Efficient Treatment of Anaphora and Long-Distance Dependency in Context-Free Grammar: An Experiment with Medical Text

Wailok Tam, Kôiti Hasida, Yusuke Matsubara, Eiji Aramaki, Mai Miyabe, Motoyuki Takaai, Hirosi Uozaki

Abstract: These characteristics put the text in pathological reports under the category of controlled natural language, making it a better object text for semantic analysis and knowledge representation. Readers unfamiliar with controlled natural language are recommended to check the survey by (Schwitter2010). The purpose of this paper is to present how to combine a CFG (Context-Free Grammar) with an ontology to account for both syntactic structures and semantic structures of sentences (and discourses) containing longdistance dependencies and anaphora found in pathological reports. The syntactic and semantic framework outlined in this paper are developed on the foundation of the Global Document Annotation (GDA) guidelines proposed by (Hasida2010). When constructing our grammar, we have an application in mind. This application is auto-completion and hence speed matters. We want to do a bit more than bigrams can achieve with auto-completion such that the effect of an antecedent or a relative clause on user input can be captured. It is true that an elaborated feature structurebased grammars with hundreds of features would have little problem with anaphora and long distance dependencies. But speed is a problem for such a grammar. This leaves us with CFGs but typical CFGs can handle neither of the phenomena we are interested in. So we make CFGs do the job.