A system for named entity recognition based on local grammars

The existence of large-scale lexical resources for Serbian, e-dictionaries in particular, coupled with local grammars in the form of finite-state transducers, enabled the development of a complex system for named entity recognition and tagging. The system is not general in nature, but targets some specific types of name, temporal and numerical expressions. In order to improve the precision of recognition we used local grammars to describe the context of named entities. In the case of personal names the widest context was used to include the recognition of nominal phrases describing a person's position. The evaluation of our system was performed twice on a corpus of 3,000 short agency news. Results obtained by the system were manually evaluated, all omissions and incorrect recognitions precisely identified, and most of them corrected before the second evaluation. The overall recall R = 0.88 for types and R = 0.94 for tokens, and overall precision P = 0.96 for types and P = 0.98 for tokens indicated that our system gives priority to precision. The evaluation of recognition of surnames only, with and without positions, and also names of distinguished persons such as royalty and church dignitaries confirmed this fact, albeit with less satisfactory results for both precision and recall.