Foul at SemEval-2023 Task 12: MARBERT Language model and lexical filtering for sentiments analysis of tweets in Algerian Arabic

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 389–396 July 13-14, 2023 ©2023 Association for Computational Linguistics Foul at SemEval-2023 Task 12: MARBERT Language model and lexical filtering for sentiments analysis of tweets in Algerian Arabic Faiza BELBACHIR IPSA - Ecole d’ing´enieurs A´eronautique et Spatiale Paris 63 Bd de Brandebourg Bis, 94200 Ivry-sur-Seine PHDUPS@gmail.com Abstract This paper describes the system we designed for our participation in SemEval-2023 Task 12 Track 6 about Algerian dialect sentiment analysis. We propose a transformer language model approach combined with a lexicon mix- ing terms and emojis which is used in a post- processing filtering stage. The Algerian senti- ment lexicons were extracted manually from tweets. We report on our experiments on the Algerian dialect, where we compare the per- formance of MARBERT to the one of Arabic- BERT and CAMeLBERT on the training and development datasets of Task 12. We also an- alyze the contribution of our post-processing lexical filtering for sentiment analysis. Our sys- tem obtained an F1 score equal to 70%, ranking 9th among 30 participants. 1 Introduction AfriSenti-SemEval shared task (Task 12) (Muham- mad et al., 2023b) aims at sentiment analysis on African dialectal languages of different African countries. It covers three main tasks and twelve lan- guages with a corpus made of messages collected from the micro-blogging service Twitter (Muham- mad et al., 2023a). Our participation is for the first task (Task A): monolingual opinion detection, in the specific case of the Algerian dialect. In case a tweet conveys both positive and negative sentiment expression, the strongest one must be selected for the whole message. Algerian dialect is considered a less-resourced language for sentiment analysis as few works have addressed the task in the past despite the relatively large number of speakers of the language1 unlike other languages like English (Belbachir and 1Algerian Arabic is spoken by the majority of Alge- rian population, counting around 32M speakers of the di- alect. Source: https://www.visualcapitalist. com/100-most-spoken-languages/ Boughanem, 2018; Pak and Paroubek, 2010). Various works on opinion detection have addressed Arabic in the past (Mohammad et al., 2018), but most of them concern classical Arabic language the exception of the work of Touileb and Barnes (2021). However everyday life communication uses the dialect. Each Arabic country is character- ized by its dialect different from the others by a large number of aspects: phonology, orthography, morphology, lexicon,...(Saˆadane et al., 2018). It can be written in different ways as shown as follow Table 1: ������ �� ��� ����� �� ����� ��� �� ������ ���� les fruits 3 �� �� ���� ���� Table 1: Various writings in the Algerian dialect for the sentence I like fruits. The importance of social media in everyday com- munication and the fact that Algerian Arabic is used by a majority of the population warrants our interest in developing an opinion-mining algorithm able to handle the dialect. To do so we need to solve the problems specific to the dialect, augmented by the peculiarities of the form …

Content truncated to 500 words