Liner2-A Customizable Framework for Proper Names Recognition for Polish

摘要

In the paper we present a customizable and open-source framework for proper names recognition called Liner2. The framework consists of several universal methods for sequence chunking which in clude: dictionary look-up, pattern matching and statistical processing. The statistical processing is performed using Conditional Random Fields and a rich set of features including morphological, lexical and semantic information. We present an application of the framework to the task of recognition proper names in Polish texts (5 common categories of proper names, i.e. first names, surnames, city names, road names and country names). The Liner2 framework was also used to train an extended model to recognize 56 categories of proper names which was used to bootstrap the manual annotation of KPWr corpus. We also present the CRF-based model integrated with a heterogeneous named entity similarity function. We show that the similarity function added to the best configuration improved the final result for cross-domain evaluation. The last section presents NER-WS-a web service for proper names recognition in Pol ish texts utilizing the Liner2 framework and the model for 56 categories of proper names.