Optimizing CRF-Based Model for Proper Name Recognition in Polish Texts

In this paper we present several optimizations introduced to Conditional Random Fields-based model for proper names recognition in Polish running texts. The proposed optimizations refer to word-level segmentation problems, gazetteers incompleteness, problem of unambiguous generalization features, feature construction and selection, and finally recognition of common proper names on the basis of external sources of knowledge. The problem of proper name recognition is limited to recognition of person first names and surnames, names of countries, cities and roads. The evaluation is performed in two ways: a single domain evaluation using 10-fold cross validation on a Corpus of Stock Exchange Reports and a cross-domain evaluation on a Corpus of Economic News. An additional corpus of Wikipedia articles, namely InfiKorp is used in the feature selection. Finally, we evaluate three configurations of proposed modifications. The top configuration improved the final result from 94.53% to 95.65% of F-measure for single domain and from 70.86% to 79.63% for cross-domain evaluation.