Discovery of Multiword Expressions with Loanwords and Their Equivalents in the Persian Language

Proceedings of Recent Advances in Natural Language Processing, pages 918–928 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_105 918 Discovery of Multiword Expressions with Loanwords and Their Equivalents in the Persian Language Katarzyna Marszałek-Kowalewska YUKKA Lab AG / Berlin, Germany k.marszalek.kowalewska@gmail.com Abstract This paper presents an attempt at multiword expressions (MWEs) discovery in the Persian language. It focuses on extracting MWEs con- taining lemmas of a particular group: loan- words in Persian and their equivalents pro- posed by the Academy of Persian Language and Literature. In order to discover such MWEs, four association measures (AMs) are used and evaluated. Finally, the list of ex- tracted MWEs is analyzed, and a compari- son between expressions with loanwords and equivalents is presented. To our knowledge, this is the first time such analysis was provided for the Persian language. 1 Introduction Today, almost 19 years after the seminal paper “Multiword Expressions - A Pain in the Neck for NLP” by Sag et al. (2002), multiword expres- sions (MWEs) are still an interesting and challeng- ing aspect of many Natural Language Processing (NLP) tasks, which is reflected in the number of papers addressing this phenomenon as well as the number of people contributing to, and attending workshops, conferences, and initiatives such as SIGLEX-MWE1 or PARSEME.2 MWEs are very frequent in language and range over a number of different linguistic constructions, from idioms, e.g., to pay an arm and a leg, to fixed expressions, e.g., rock and roll, light verb constructions, e.g., take a shower, to noun compounds, e.g., golf club. Biber et al. (1999) claim that the number of MWEs in spoken English is 30% – 45% and 21% in academic prose. Jackendoff (1997) suggests that the number of MWEs in a speaker’s lexicon is the same as simple words. Nevertheless, if we take into consid- eration the domain-specific lexicons, this number 1http://https://multiword.org 2https://typo.uni-konstanz.de/parseme/ seems to be an underestimation (Sag et al., 2002). Indeed, the research conducted by Ramisch (2009) suggests that the MWEs ratio can be between 50% and 80% in a corpus of scientific biomedical ab- stracts. Research by Krieger and Finatto (2004) estimate that MWEs can constitute more than 70% of the specialized lexicon. MWEs have received considerable attention in recent years and it has been suggested (Sag et al., 2002) that their proper treatment could make a sig- nificant improvements in a number of NLP tasks, e.g., lexicography (Church and Hanks, 1990; Gan- tar et al., 2018; Fellbaum, 2016), word sense disam- biguation (Finlayson and Kulkarni, 2011), part-of- speech tagging and parsing (Baldwin et al., 2004), information retrieval (Newman et al., 2012), lan- guage learning (Christiansen and Arnon, 2017), machine translation (Carpuat and Diab, 2010) or sentiment analysis (Berend, 2011; Williams et al., 2015). The research on MWEs in Persian has so far fo- cused mainly on verbal multiword units and light verb constructions (LVCs) in particular. Taslim- ipoor et al. (2012) adopted a method originally pro- posed by Fazly et al. (2007) for identifying LVCs in Persian. They extended existing statistical mea- sures …

Content truncated to 500 words