Serbian NER&Beyond: The Archaic and the Modern Intertwinned

Proceedings of Recent Advances in Natural Language Processing, pages 1252–1260 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_141 1252 Serbian NER&Beyond: The Archaic and the Modern Intertwinned Branislava Šandrih Todorović University of Belgrade Faculty of Philology branislava.sandrih@fil.bg.ac.rs Ranka Stanković University of Belgrade Faculty of Mining and Geology ranka.stankovic@rgf.bg.ac.rs Cvetana Krstev University of Belgrade Faculty of Philology cvetana@matf.bg.ac.rs Milica Ikonić Nešić University of Belgrade Faculty of Philology milica.ikonic.nesic@fil.bg.ac.rs Abstract In this work, we present a Serbian litera- ry corpus that is being developed under the umbrella of the “Distant Reading for European Literary History” COST Acti- on CA16204. Using this corpus of novels written more than a century ago, we ha- ve developed and made publicly available a Named Entity Recognizer (NER) trai- ned to recognize 7 different named enti- ty types, with a Convolutional Neural Ne- twork (CNN) architecture, having F1 score of ≈91% on the test dataset. This model has been further assessed on a separate eva- luation dataset. We wrap up with compa- rison of the developed model with the exi- sting one, followed by a discussion of pros and cons of the both models. 1 Introduction The “Distant Reading for European Literary History”1 (COST Action CA16204) has started in 2017 with the purpose of using computa- tional methods to analyse large collections of literary texts (Stanković et al., 2019; Frontini et al., 2020). The main goal of this ongoing action is to compile a multilingual open-source collection, named European Literary Text Col- lection (ELTeC), containing linguistically anno- tated sub-collections of 100 novels per language written more than 100 years ago. In this paper, we present a collection of Ser- bian texts in this corpus, named SrpELTeC. Alongside, we describe our efforts in developing its Named Entity (NE) layer, defined previou- sly as one of the main action’s deliverables. For this purpose, we adjusted and used the existing rule-based NE recognizer for Serbian, 1Distant Reading, https://www.distant-reading.net dubbed SrpNER, that we will describe in Sec- tion 2 together with some approaches to NE recognition in literary texts. This SrpNER model was applied to the raw version of the selected texts from SrpELTeC collection, pre- sented in Section 3. Based on the specifically tailored guidelines, different evaluators perfor- med careful checks and corrections, yielding a gold standard (SrpELTeC-gold). This ena- bled us to train a CNN-based NE recognizer, named SrpCNNER, presented in Section 4. Having the gold dataset, prepared as described in Subsection 4.1, we trained (Subsection 4.2) and evaluated the model in two different set- tings: first, we discussed our model’s perfor- mance on the SrpELTeC-gold test subset, as shown in Subsection 4.3. Afterwards we car- ried out a detailed evaluation on a collection of novels that were not present in the gold standard, named SrpELTeC-eval, with the findings and a thorough discussion given in Section 5. Finally, conclusions and plans for the future work were stated in Section 6. 2 Related Work The existence of large-scale lexical resources for Serbian, e-dictionaries in particular (Kr- stev, 2008), coupled with local grammars in the …

Content truncated to 500 words