Persian Ezafe Recognition Using Transformers and Its Role in Part-Of-Speech Tagging

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 961–971 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 961 Persian Ezafe Recognition Using Transformers and Its Role in Part-Of-Speech Tagging Ehsan Doostmohammadi♣, Minoo Nassajian♣, Adel Rahimi♠ ♣Sharif University of Technology, Tehran, Iran ♠Dathena Science Pte. Ltd., Singapore {e.doostm72,m.nassajian2016}@student.sharif.edu, adel.rahimi@dathena.io Abstract Ezafe is a grammatical particle in some Iranian languages that links two words together. Re- gardless of the important information it con- veys, it is almost always not indicated in Per- sian script, resulting in mistakes in reading complex sentences and errors in natural lan- guage processing tasks. In this paper, we experiment with different machine learning methods to achieve state-of-the-art results in the task of ezafe recognition. Transformer- based methods, BERT and XLMRoBERTa, achieve the best results, the latter achieving 2.68% F1-score more than the previous state- of-the-art. We, moreover, use ezafe informa- tion to improve Persian part-of-speech tagging results and show that such information will not be useful to transformer-based methods and ex- plain why that might be the case. 1 Introduction Persian ezafe is an unstressed morpheme that ap- pears on the end of the words, as -e after conso- nants and as -ye1 after vowels. This syntactic phe- nomenon links a head noun, head pronoun, head adjective, head preposition, or head adverb to their modifiers in a constituent called ‘ezafe construc- tion’ (Nassajian et al., 2019). Whether a word in a sentence receives or does not receive ezafe might affect that sentence’s semantic and syntactic struc- tures, as demonstrated in Examples 1a and 1b in Figure 1. There are some constructions in English that can be translated by ezafe construction in Per- sian. For instance, English ‘of’ has the same role as Persian ezafe to show the part-whole relation, the relationship of possession, or ‘’s’ construction, and possessive pronouns followed by nouns show- ing genitive cases are mirrored by Persian ezafe (Karimi and Brame, 2012). 1The y is called an intrusive y and is an excrescence be- tween two vowels for the ease of pronunciation. This affix is always pronounced but almost al- ways not written, which results in a high degree of ambiguity in reading and understanding Persian texts. It is hence considered as one of the most interesting issues in Persian linguistics, and it has been discussed in details from phonological as- pects (Ghomeshi, 1997), morphological aspects (Samvelian, 2006, 2007) and (Karimi and Brame, 2012), and syntactic aspects (Samiian, 1994; Lar- son and Yamakido, 2008; Kahnemuyipour, 2006, 2014, 2016). Nearly 22% of the Persian words have ezafe (Bi- jankhan et al., 2011), which shows the prevalence of this marker. Moreover, this construction also ap- pears in other languages such as Hawramani (Holm- berg and Odden, 2005), Zazaki (Larson and Ya- makido, 2006; Toosarvandani and van Urk, 2014), Kurdish (Karimi, 2007) etc. Ezafe construction is also similar to idafa construction in Arabic and construct state in Hebrew (Habash, 2010; Karimi and Brame, 2012) and Zulu (Jones, 2018). Ezafe recognition is the task …

Content truncated to 500 words