Named entity recognition in Slovene text

This paper presents an approach and an implementation of a named entity extractor for Slovene language, based on a machine learning approach. It is designed as a supervised algorithm based on Conditional Random Fields and is trained on the ssj500k annotated corpus of Slovene. The corpus, which is available under a Creative Commons CC-BY-NC-SA licence, is annotated with morphosyntactic tags, as well as named entities for people, locations, organisations, and miscellaneous names. The paper discusses the influence of morphosyntactic tags, lexicons and conjunctions of features of neighbouring words. An important contribution of this investigation is that morphosyntactic tags benefit named entity extraction. Using all the best-performing features the recognizer reaches a precision of 74% and a recall of 72%, having stronger performance on personal and geographical named entities, followed by organizations, but performs poorly on the miscellaneous entities, since this class is very diverse and consequently difficult to predict. A major contribution of the paper is also showing the benefits of splitting the class of miscellaneous entities into organizations and other entities, which in turn improves performance even on personal and organizational names. The software, developed in this research is freely available under the Apache 2.0 licence at http://ailab.ijs.si/~tadej/slner.zip, while development versions are available at https://github.com/tadejs/slner