Transfer-based Enrichment of a Hungarian Named Entity Dataset

Proceedings of Recent Advances in Natural Language Processing, pages 1060–1067 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_119 1060 Transfer-based Enrichment of a Hungarian Named Entity Dataset Attila Nov´ak and Borb´ala Nov´ak MTA-PPKE Hungarian Language Technology Research Group, P´azm´any P´eter Catholic University Faculty of Information Technology and Bionics Pr´ater u. 50/a, 1083 Budapest, Hungary {surname.firstname}@itk.ppke.hu Abstract In this paper, we present a major update to the first Hungarian named entity dataset, the Szeged NER corpus. We used zero-shot cross- lingual transfer to initialize the enrichment of entity types annotated in the corpus us- ing three neural NER models: two of them based on the English OntoNotes corpus and one based on the Czech Named Entity Corpus fine-tuned from multilingual neural language models. The output of the models was auto- matically merged with the original NER an- notation, and automatically and manually cor- rected and further enriched with additional an- notation, like qualifiers for various entity types. We present the evaluation of the zero-shot per- formance of the two OntoNotes-based mod- els and a transformer-based new NER model trained on the training part of the final corpus. We release the corpus and the trained model. 1 Introduction 1.1 Resources Named entity recognition is a fundamental NLP task that has played an important role in tasks like information extraction, document deidentification, conversational models, etc. Following the annota- tion scheme used in the CoNLL 2002/2003 NER annotation tasks, legacy named entity corpora usu- ally contain annotation of four entity types: orga- nizations (ORG), persons (PER), locations (LOC) and general entity category covering all the rest (MISC). This is the case for all named entity cor- pora available for Hungarian, the Szeged NER cor- pus (Szarvas et al., 2006), the silver-standard Hun- garian hunNERwiki corpus (Simon and Nemeskey, 2012) automatically derived from Wikipedia, and the recently published NerKor corpus.1 The En- glish OntoNotes 5 corpus (Weischedel et al., 2013), 1https://github.com/nytud/NYTK-NerKor on the other hand, contains a richer set of enti- ties. Geopolitical entities (GPE: countries, settle- ments, etc.) and facilities (FAC: buildings, roads, airports etc.) are differentiated from geographi- cal locations like continents or bodies of waters. Within the MISC category, products (PROD), laws and other norms (LAW), events (EVENT) and titles of works of art (WORK OF ART) are differenti- ated. In addition, the OntoNotes NER tagset also encompasses time and numerical expressions dis- tinguishing dates and times, cardinal and ordinal numbers, quantities, percentages and amounts of money. In addition, other categories covering non- entities like languages (LANGUAGE) and national- ities, religions and political affiliations (NORP ‘na- tionality/other/religion/political’) are covered, pre- sumably just because English orthography happens to prescribe capitalization for words (adjectives in the case of NORP) belonging to this category. Some resources in languages other than English also use NER tagsets richer than the basic four- class tagset. Although the NoSta-D resource used in the GermEval2014 shared task targeting German NER (Benikova et al., 2014) maintains a four-class distinction, words (especially adjectives) derived from names as well as compounds containing them are marked as such. …

Content truncated to 500 words