Machine Learning and Deep Neural Network-Based Lemmatization and Morphosyntactic Tagging for Serbian

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 3954–3962 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 3954 Machine Learning and Deep Neural Network-Based Lemmatization and Morphosyntactic Tagging for Serbian Ranka Stankovi´c, Branislava ˇSandrih, Cvetana Krstev, Miloˇs Utvi´c, Mihailo ˇSkori´c {Faculty of Mining and Geology, Faculty of Philology} University of Belgrade {Djuˇsina 7, Studentski trg 3} Belgrade, Serbia {ranka.stankovic, mihailo.skoric}@rgf.bg.ac.rs, cvetana@matf.bg.ac.rs, {branislava.sandrih, misko}@fil.bg.ac.rs Abstract The training of new tagger models for Serbian is primarily motivated by the enhancement of the existing tagset with the grammatical category of a gender. The harmonization of resources that were manually annotated within different projects over a long period of time was an important task, enabled by the development of tools that support partial automation. The supporting tools take into account different taggers and tagsets. This paper focuses on TreeTagger and spaCy taggers, and the annotation schema alignment between Serbian morphological dictionaries, MULTEXT-East and Universal Part-of-Speech tagset. The trained models will be used to publish the new version of the Corpus of Contemporary Serbian as well as the Serbian literary corpus. The performance of developed taggers were compared and the impact of training set size was investigated, which resulted in around 98% PoS-tagging precision per token for both new models. The SR BASIC annotated dataset will also be published. Keywords: Part-of-Speech tagging, lemmatization, corpus, evaluation, Serbian, morphological dictionary 1. Introduction The task of assigning to each token its Part-of-Speech cat- egory (noun, verb, adjective, etc.) is a common Natural Language Processing (NLP) task, known as Part-of-Speech tagging (PoS-tagging). PoS-tagging precedes many other Natural Language Processing tasks, such as Text Classi- fication, Named Entity Recognition, Sentiment Analysis, Question Answering, etc. Computer programs that perform this task, the so-called ‘taggers’, can be based on lookup-tables, regular expres- sions, linguistic (morphological, semantic and syntactic) rules, machine learning methods (Gim´enez and M`arquez, 2004; Denis and Sagot, 2009; Manning et al., 2014) or state-of-the-art Deep Neural Networks (DNNs) (Huang et al., 2015; Choi, 2016; Akbik et al., 2018). The first operational tagger and lemmatization model for Serbian were produced as a parameter file (TT11) for the TreeTagger model (Schmid, 1999) after a thorough anal- ysis of the state-of-the-art solutions (Popovi´c, 2010), and subsequently used for various purposes. The research pre- sented in this paper was motivated primarily by the need to enrich the tagset with new tags, such as the gram- matical category of gender, harmonize resources manually annotated within different projects over a long period of time, as well as to develop tools for supporting prepara- tion of training sets to be used for different taggers and tagsets in the future. The research was focused on anno- tation schemata alignment between Serbian morphological dictionaries tagset (presented briefly in Subsection 2.1.), MULTEXT-East tagset (Erjavec, 2012), and the Universal Part-of-Speech tagset (Petrov et al., 2012), for the purpose of preparing datasets for training the TreeTagger and spaCy taggers. The main goal was to automate the process of an- notation schema harmonization …

Content truncated to 500 words