Reconstructing NER Corpora: a Case Study on Bulgarian

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 4647–4652 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 4647 Reconstructing NER Corpora: a Case Study on Bulgarian Iva Marinova, Laska Laskova, Petya Osenova, Kiril Simov, Alexander Popov LMaKP, IICT-BAS Sofia, Buglaria iva.marinova@identrics.net {laska, petya, kivs, alex.popov}@bultreebank.org Abstract The paper reports on the usage of deep learning methods for improving a Named Entity Recognition (NER) training corpus and for predicting and annotating new types in a test corpus. We show how the annotations in a type-based corpus of named entities (NE) were populated as occurrences within it, thus ensuring density of the training information. A deep learning model was adopted for discovering inconsistencies in the initial annotation and for learning new NE types. The evaluation results get improved after data curation, randomization and deduplication. Keywords: NER corpora, Bulgarian, deep learning 1. Introduction Named Entity Recognition (NER) and Named Entity Link- ing (NEL) have been active research areas for many years now, encompassing a wide range of topics: from rule-based systems to neural networks; from coarse-grained classifica- tion tasks (person, location, organization tagsets) to more fine-grained ones (including products, events, etc.); from local and global approaches to combined ones, etc. While these two related tasks are considered to be well covered in NLP for Germanic, Romance and other language groups, they are under-resourced for the Slavic languages, espe- cially from a multilingual perspective. Recently, a shared task was launched for Slavic languages on NER and linking to the same real-world entity in a cross- lingual context. The second edition in 2019 (Piskorski et al., 2019) included the following languages: Bulgarian, Czech, Polish, Russian. It focused on 5 types (or sorts and categories) of Named Entities (NEs): Person (PER), Loca- tion (LOC), Organization (ORG), Event (EVT), and Prod- uct (PRO). In the work reported here we exploit this newly constructed Bulgarian dataset for NER. The second dataset was ex- tracted from the Bulgarian HPSG-based treebank — Bul- TreeBank (Simov et al., 2002). The named entities in Bul- TreeBank were syntactically annotated and the category for each NE was assigned to the highest node in the syntactic annotation. For the aims of this work, the categories were inherited to the token level. In BulTreeBank, only form categories were annotated: Person (PER), Location (LOC), Organization (ORG), and Other (OTH). The category Other covers a wide range of entity types like Event (EVT), Prod- uct (PRO), and some of their sub-types — different kinds of products, for example. The preparation of the Bulgarian corpus within a highly multilingual guidance environment was very useful for im- provement on two fronts: the quality of the gold data, as well as the prediction power of the deep learning NER mod- els trained on it. The former was achieved through the NE occurrences population and through discovering errors made by the deep model. The latter was achieved by the ap- plication of a deep learning model over the corpus extracted from …

Content truncated to 500 words