Predicting Dialect Variation in Immigrant Contexts Using Light Verb Constructions

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1391–1395, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Predicting Dialect Variation in Immigrant Contexts Using Light Verb Constructions A. Seza Do˘gru¨oz Netherlands Institute for Advanced Study Wassenaar, Netherlands a.s.dogruoz@gmail.com Preslav Nakov Qatar Computing Research Institute Tornado Tower floor 10, P.O. Box 5825, Doha, Qatar pnakov@qf.org.qa Abstract Languages spoken by immigrants change due to contact with the local languages. Capturing these changes is problematic for current language technologies, which are typically developed for speakers of the standard dialect only. Even when dialec- tal variants are available for such technolo- gies, we still need to predict which di- alect is being used. In this study, we dis- tinguish between the immigrant and the standard dialect of Turkish by focusing on Light Verb Constructions. We experiment with a number of grammatical and contex- tual features, achieving over 84% accuracy (56% baseline). 1 Introduction Human languages are in constant evolution, driven in part by contact with other languages (Uriel, 1953; Thomason, 2008). In immigrant contexts, bilingual and multilingual speakers act as agents of change by transmitting borrowed words and ex- pressions across languages (Grosjean, 2014). De- pending on social factors such as duration and in- tensity of contact with the local languages, large- scale spread of borrowed elements could lead to differences between the contact and non-contact dialects of the same language (Winford, 2005). For example, Spanish spoken by immigrants in USA sounds different in comparison to Spanish spoken in South America (Corval´an, 2003). In this study, we focus on the immigrant di- alect of Turkish as spoken in the Netherlands (NL-Turkish), which differs from Turkish spo- ken in Turkey (TR-Turkish). In contact situa- tions, it is common for verbs to be borrowed across languages and integrated as nominal com- plements of Light Verb Constructions (LVCs) (Ed- wards and Gardner-Chloros, 2007; Butt, 2010). NL-Turkish LVCs are changing due to Dutch in- fluence (Do˘gru¨oz and Backus, 2007; Do˘gru¨oz and Backus, 2009; Do˘gru¨oz and Gries, 2012). How- ever, assessing Dutch influence is not always easy since NL-Turkish LVCs still co-exist with the TR- Turkish LVCs. This study aims to automatically identify the features that can distinguish between NL-Turkish and TR-Turkish LVCs. Our study would benefit Machine Translation systems targeting dialectal variation. It differs from studies concerning the well-established di- alectal variations of Arabic, e.g., Levantine, Gulf, Egyptian, Maghrebi (Salloum and Habash, 2012), EU vs. Brazilian Portuguese (Marujo et al., 2011) or Turkish vs. Tatar (Altintas and Cicekli, 2002). In contrast, we are interested in developing lan- guage technologies for immigrant dialects, which are often understudied and lack written resources due to their unofficial status. When immigrant speakers face communication difficulties (e.g., bu- reaucratic affairs with the local officials, teacher- parent meetings, doctor-patient conversations) in the local languages (e.g., Dutch) of the host coun- try, they are often provided with translation equiv- alents in the standard dialect (e.g., TR-Turkish) of their native languages. However, these trans- lations ignore the evolution …

Content truncated to 500 words