SenZi: A Sentiment Analysis Lexicon for the Latinised Arabic (Arabizi)

Proceedings of Recent Advances in Natural Language Processing, pages 1203–1211, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_138 1203 SenZi: A Sentiment Analysis Lexicon for the Latinised Arabic (Arabizi) Taha Tobaili1, Miriam Fernandez1, Harith Alani1, Sanaa Sharafeddine2, Hazem Hajj3, and Goran Glavaš4 1Knowledge Media Institute, The Open University 2Department of Computer Science and Mathematics, Lebanese American University 3Department of Electrical and Computer Engineering, American University of Beirut 4Data and Web Science Group, Universität Mannheim {taha.tobaili, miriam.fernandez, h.alani}@open.ac.uk, sanaa.sharafeddine@lau.edu.lb, hh63@aub.edu.lb, goran@informatik.uni-mannheim.de Abstract Arabizi is an informal written form of di- alectal Arabic transcribed in Latin alphanu- meric characters. It has a proven popularity on chat platforms and social media, yet it suffers from a severe lack of natural language process- ing (NLP) resources. As such, texts written in Arabizi are often disregarded in sentiment analysis tasks for Arabic. In this paper we de- scribe the creation of a sentiment lexicon for Arabizi that was enriched with word embed- dings. The result is a new Arabizi lexicon con- sisting of 11.3K positive and 13.3K negative words. We evaluated this lexicon by classify- ing the sentiment of Arabizi tweets achieving an F1-score of 0.72. We provide a detailed er- ror analysis to present the challenges that im- pact the sentiment analysis of Arabizi. 1 Introduction Arabizi, a portmanteau of Arabic and Englizi (En- glish), is a written form of dialectal Arabic (DA) often used by Arabic speakers for informal com- munication in messaging applications and on so- cial media enabling them to type Arabic words using Latin letters (Yaghan, 2008). Arabizi lacks a consistent orthography and reflects the various dialects of Arabic, which differ from one Arab region to another and from the formal Modern Standard Arabic (MSA) phonetically, morpholog- ically, and syntactically. Studies show that Arabizi has reached 12% of the Latin script tweets in Lebanon and 25% of the Latin script tweets in Egypt (Tobaili, 2016). It is a common way of communication among the youth (Keong et al., 2015; Muhammed et al., 2011; Allehaiby, 2013) and has been actively used during relevant events in the Arab world such as the Arab spring (Basis-Technology, 2012). De- spite the growth of this newly born written lan- guage, Arabic sentiment analysis approaches of- ten disregard Arabizi text due to the challenges it poses and the scarcity of NLP resources to process it (Bies et al., 2014). In this work we contribute to the sentiment anal- ysis of Arabizi by creating a new Arabizi sen- timent lexicon (SenZi) for the Lebanese dialect. We annotated a 3.4K Arabizi Twitter dataset to evaluate the lexicon and to train an Arabizi lan- guage identifier. We used this identifier to create an Arabizi corpus of 1M public Facebook com- ments. We widened the coverage of SenZi by en- riching it with inflectional and orthographic forms for each sentiment word using word embeddings on the corpus reaching 11.3K positive and 13.3K negative words. All resources and detailed de- scription are made public and freely accessible on the project’s webpage1. The rest …

Content truncated to 500 words