TArC: Tunisian Arabish Corpus, First complete release

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 1125–1136 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 1125 TArC: Tunisian Arabish Corpus First complete release Elisa Gugliotta, Marco Dinarelli Laboratoire d’Informatique de Grenoble (LIG), LIDILEM, “Sapienza” University of Rome Campus Universitaire de Saint-Martin-D’Hères, 38000 Grenoble, Circonvallazione Tiburtina 4, 00185, Rome {elisa.gugliotta,marco.dinarelli}@univ-grenoble-alpes.fr Abstract In this paper we present the final result of a project on Tunisian Arabic encoded in Arabizi, the Latin-based writing system for digital conversations. The project led to the creation of two integrated and independent resources: a corpus and a NLP tool created to annotate the former with various levels of linguistic information: word classification, transliteration, tokenization, POS-tagging, lemmatization. We discuss our choices in terms of computational and linguistic methodology and the strategies adopted to improve our results. We report on the experiments performed in order to outline our research path. Finally, we explain why we believe in the potential of these resources for both computational and linguistic researches. Keywords: Tunisian Arabizi, Annotated Corpus, Neural Network Architecture 1. Introduction In this paper we describe the methodology we adopted for building a dialectal Arabic Corpus from scratch. Our motivations for building a corpus from scratch are related to the utility we envisioned with its release, both from a linguistic and a NLP points of view. Along with the path identified to achieve our goal, we give the linguistic motivations and describe the computational experiments that guided the choice of our approach, leading us to the final corpus structure. The corpus is the result of a semi-automatic annotation procedure carried out using an NLP tool based on neural models, that we developed specifically to create the corpus. Our architecture produces in cascade the different levels of annotation that we decided to have in the corpus. For many reasons, mentioned below, our approach is hybrid. First of all, the project lies at the intersection of different research fields: Arabic dialectology, corpus linguistics and deep learning. Secondly, the texts collected in our corpus are in an Arabic dialect. Arabic dialects are notoriously under-resourced linguistic systems. In particular, the texts collected in our corpus are encoded in a script that, on the one hand represents some phonetic phenomena of Tunisian Arabic (e.g., the article assimilation, unlike the encoding in Arabic script), on the other hand is a writing system. Such script arose in a diamesically influenced context, namely digital environments. The encoding we refer to uses the Latin alphabet, as well as some numbers, for Arabic phonemes, without correspondence in the Latin script. This encoding is known as Franco-Arabic, Arabizi, Arabish et alias, depending on the Arabic country. We focused on the system in use for writing Tunisian Arabic. The structure of the paper is as follows: we describe the state-of-the-art in section 2. In section 3 we explain the reasoning behind the planning of our work. We will de- scribe our corpus building steps in section (4). In sec- tion 5 we …

Content truncated to 500 words