---
title: 'Tunizi: Tunisian Arabizi Sentiment Data'
url: https://www.emergentmind.com/topics/tunizi
type: topic
---

# Tunizi: Tunisian Arabizi Sentiment Data

Searching arXiv for the specified paper and closely related Tunisian Arabizi/Tunizi work to ground the article with current citations.
TUNIZI is a manually annotated sentiment analysis dataset of Tunisian Arabizi, the informal way Tunisians write their dialect using Latin characters and numerals instead of Arabic script. It consists of \(N = 9210\) user comments written only in Tunisian Arabizi, each labeled with binary sentiment, positive or negative, and is described as the first publicly available sentiment dataset specifically for Tunisian Arabizi [2004.14303]. More broadly, the term “Tunizi” is also used in subsequent work to denote Tunisian Arabic written in Latin script, an orthographic practice that is prevalent in informal digital communication and that poses distinct NLP challenges because it is dialectal, non-standardized, and frequently code-switched [2010.06857, 2511.16683].

## 1. Definition, linguistic scope, and motivation

Tunisian Arabizi refers to Tunisian Arabic written with Latin letters and numerals. The dataset description gives examples such as `lkolna m3ak`, and specifies numerals representing Arabic consonants, including `3` for ع, `7` for ح, and `5` for خ. Example lexical items include `3asslema`, `chna7welek`, `sou2el`, and `5dhit`, paired respectively with Arabic-script forms and English glosses such as “Hello,” “How are you,” “Question,” and “I took” [2004.14303]. Related work further characterizes TUNIZI/Tunizi as Tunisian Arabic written in Roman script with numerals and multigraphs, including correspondences such as “ch” for ش, “gh” or “8” for غ, and “9” for ق, and highlights a French-oriented spelling tendency that distinguishes Tunisian Arabizi from Egyptian Arabizi [2010.06857].

This variety is important because it is not Modern Standard Arabic; it is a spoken dialect with distinct vocabulary, morphology, and syntax. The dataset paper states that a multidialectal parallel corpus shows Tunisian dialect has the least lexical overlap with other Arabic dialects, reinforcing that Tunisian is particularly distinct and low-resource [2004.14303]. Related sentiment-analysis work reiterates that Tunisian has the lowest lexical overlap with other major Arabic dialects, which limits transfer from Egyptian or Levantine resources [2010.06857]. A plausible implication is that script conversion alone does not eliminate the need for Tunisian-specific lexical and semantic modeling.

The motivation for TUNIZI is explicitly framed around a lack of resources, the low-resource status of African and Arabic dialects, and the need for sentiment analysis on social media. The authors state that no annotated Tunisian Arabizi dataset existed, that deep learning methods are data-hungry, and that Tunisians express opinions in Arabizi on products, services, politics, social issues, entertainment, and culture [2004.14303]. This suggests that TUNIZI was conceived not merely as a corpus release, but as a foundational resource for dialect-specific opinion mining.

## 2. Data acquisition and corpus composition

TUNIZI was collected exclusively from YouTube comments. The paper states: “TUNIZI is collected from comments on Youtube social network” [2004.14303]. To ensure representativeness, comments were scraped from videos across sports, politics, comedy, TV shows and TV series, arts, and Tunisian music videos. The stated purpose of this diversity was to capture different age groups, varied socio-cultural backgrounds, and different writing styles and registers [2004.14303].

The collection and filtering process involved automated scraping with a custom script developed by the iCompass team, followed by a second script that removed comments written in French and English. Because automatic filtering did not capture all non-Arabizi comments, the authors then manually reviewed the data to ensure that the final dataset only contains Tunisian Arabizi comments, excluding comments fully in French or English, comments in other Arabic dialects or MSA, and non-textual or irrelevant content [2004.14303]. The result is a curated set of 9210 comments.

The core descriptive statistics reported for the dataset are as follows.

| Characteristic | Number |
|---|---:|
| \#Comments | 9210 |
| \#Negative comments | 4838 |
| \#Positive comments | 4372 |
| \#Words | 79,862 |
| \#Unique words | 22,850 |

The sentiment distribution is described as balanced, with 47% positive and 53% negative, slightly skewed toward negative but close to 1:1 and suitable for binary classification [2004.14303]. Using the reported counts, the average comment length is approximately \(79{,}862 / 9{,}210 \approx 8.67\) words, indicating short social-media utterances. The large number of unique words relative to corpus size is presented as evidence of non-standard spelling, heavy variation in Arabizi orthography, and slang or domain-specific terms [2004.14303].

A later paper uses a “TUNIZI Dataset (Fourati et al.)” with 9,911 comments, 4,679 negative and 5,232 positive, 82,384 total word tokens, and 30,635 unique word types, split into 8,616 train and 1,295 test examples [2010.06857]. Because these counts differ from the 9,210-comment release described in the original dataset paper, the safest interpretation is that the term “TUNIZI” came to denote both the original corpus release and a closely related dataset configuration used in downstream sentiment experiments [2004.14303, 2010.06857].

## 3. Preprocessing and annotation protocol

The preprocessing applied to TUNIZI is deliberately light. The authors explicitly state that “TUNIZI was preprocessed by removing links, emoji symbols and punctuation” [2004.14303]. The paper does not describe an explicit Arabizi normalization function and does not formalize numeral-to-phoneme mappings. Instead, Arabizi spelling is left as originally written by users, preserving numerals as part of tokens and preserving non-standard spellings and variations [2004.14303]. This preserves actual user-generated language, though a plausible implication is that downstream models must absorb substantial orthographic entropy.

The paper does not discuss systematic handling of mixed-language comments or Arabic-script/Latin-script mixtures. The automatic scripts removed comments in French and English, and manual filtering aimed to keep comments that are predominantly Tunisian Arabizi [2004.14303]. This suggests that significant code-switching is reduced but not formally modeled in the dataset construction.

Annotation was carried out by five Tunisian native speakers: three males and two females, all at higher education level. The profiles listed are two male PhD holders aged 43 and 42, one female and one male aged 25 who were R&D engineers at iCompass, and one female aged 23 who was a software engineering student [2004.14303]. The label scheme is binary sentiment, positive or negative. The paper gives example labeled comments such as `lkolna m3ak w msendinek` (“We are all with you and supporting you”) for Positive and `nakrhek 5atrek kadheb` (“I hate you because you are a liar”) for Negative [2004.14303].

The full set of comments was divided equally among the five annotators, each annotator labeled their subset as positive or negative, and an evaluation or review was then performed by the two female annotators to verify and correct labels as needed [2004.14303]. The paper does not detail formal annotation guidelines and does not report inter-annotator agreement metrics such as Cohen’s \(\kappa\) or Fleiss’ \(\kappa\). No neutral, mixed, or fine-grained labels are included in this version of TUNIZI [2004.14303].

Later work on a separate 100-example Tunizi benchmark adopts a different setup: a three-way parallel dataset of Tunizi, Arabic-script Tunisian Arabic, English, and sentiment labels POSITIVE, NEGATIVE, and NEUTRAL, annotated manually by a single annotator [2511.16683]. That benchmark is not TUNIZI in the narrow dataset sense, but it shows how the notion of Tunizi expanded from binary sentiment classification to transliteration, translation, and three-class sentiment evaluation.

## 4. Orthographic and computational specificities

TUNIZI’s central linguistic challenge is the absence of a conventional orthography. The dataset description emphasizes non-standard, highly variable spelling, common use of numerals for consonants, script and language mixing, and marked dialectal specificity [2004.14303]. Related work adds that Tunisian Arabizi is influenced by French spelling conventions, that comments often mix Tunisian Arabizi with French, English, and Spanish, and that there is no official orthography, so users improvise with competing numeral and multigraph choices such as `ghalia` and `8alia` [2010.06857].

These properties create standard NLP difficulties: language identification, normalization, tokenization, and transfer across dialects. The TUNIZI paper does not specify a specialized tokenizer; tokenization appears to be simple space-based segmentation after cleaning, inferred from the reported word and unique-word counts [2004.14303]. The lack of predefined train/dev/test splits in the original release means users are expected to create their own splits for experiments [2004.14303].

Subsequent work on Tunisian Arabizi provides a broader technical context. A multi-task sequence prediction system for Tunisian Arabizi, implemented in Fairseq, was proposed to produce multi-level annotation: text classification of tokens as arabizi, foreign, or emotag; encoding into CODA* Arabic orthography; morphologically aware tokenization; and PoS tagging [2011.05152]. That work explicitly cites “TUNIZI: a Tunisian Arabizi sentiment analysis Dataset” as a lexical resource merged with the TLD dataset for pretraining the classification stage [2011.05152]. This indicates that TUNIZI’s importance extends beyond sentiment labels: it also contributes lexical coverage for broader Tunisian Arabizi processing pipelines.

A plausible implication is that TUNIZI occupies a foundational position in a layered Tunisian NLP stack. In such a stack, raw Tunisian Arabizi can be normalized or CODAfied, morphologically segmented, PoS-tagged, and then used for sentiment or other downstream tasks. However, the original TUNIZI release itself remains intentionally minimal and does not include these additional annotation layers [2004.14303].

## 5. Modeling use and empirical baselines in later work

The original TUNIZI paper is primarily a dataset introduction. It does not describe specific baseline models, deep learning architectures, feature engineering, loss functions, training procedures, or official benchmark metrics [2004.14303]. Its emphasis is on dataset construction and public release rather than benchmarking.

A later study, however, uses TUNIZI for comparative evaluation of unsupervised word representations and deep neural architectures [2010.06857]. That work evaluates three embedding approaches: TUNIZI-specific word2vec trained from scratch on the TUNIZI corpus with 300-dimensional vectors; pretrained French frWaC word2vec embeddings, also 300-dimensional; and multilingual BERT, used as a fixed embedding source with WordPiece tokenization [2010.06857]. The authors compare CNN and Bi-LSTM architectures without handcrafted features or preprocessing.

On the TUNIZI dataset used in that study, the reported results are:

| Model | ACC | F1 micro | F1 macro |
|---|---:|---:|---:|
| Word2vec + CNN | 67.2% | 67.2% | 67.1% |
| frWaC + CNN | 70.2% | 70.2% | 69.0% |
| M-BERT + Bi-LSTM | 76.3% | 76.3% | 74.3% |
| M-BERT + CNN | 78.3% | 78.3% | 78.1% |

The best result on TUNIZI is M-BERT + CNN, with ACC 78.3%, F1 micro 78.3%, and F1 macro 78.1% [2010.06857]. The same paper attributes the superiority of French embeddings over TUNIZI-trained word2vec to the much larger French training corpus and better coverage of French segments in code-switched comments, while M-BERT’s advantage is linked to subword tokenization and multilingual coverage [2010.06857]. This suggests that TUNIZI sentiment modeling benefits strongly from subword-aware multilingual representations, especially under OOV pressure and code-switching.

The paper also evaluates a filtered Romanized subset of TSAC, denoted TSAC-TUNIZI, and reports even higher results there, with M-BERT + CNN achieving ACC 93.2% and F1 93.0% [2010.06857]. These experiments do not change the content of the original TUNIZI release, but they establish TUNIZI as a benchmarkable resource for modern representation learning.

## 6. Broader significance, extensions, and misconceptions

TUNIZI matters because Tunisian dialect is both distinct from other Arabic dialects and severely low-resource, especially in Arabizi form. Before TUNIZI, no annotated Tunisian Arabizi sentiment dataset existed, limiting the development of deep learning and other data-driven NLP methods for this variety [2004.14303]. The dataset therefore contributes directly to research on low-resource sentiment analysis, social media monitoring in Tunisian dialect, and cross-dialect or cross-script studies [2004.14303].

A common misconception is to treat Tunisian Arabizi as merely Arabic transliterated into Latin script. The data and related papers do not support that reduction. Tunisian Arabizi is described as dialectal, lexically divergent, morphosyntactically distinct from MSA, and shaped by French-oriented orthographic conventions and multilingual borrowing [2004.14303, 2010.06857, 2511.16683]. Another misconception is that transliteration to Arabic script is sufficient for robust NLP. Related work on transliteration, CODA* conversion, and LLM evaluation shows that mapping Tunizi to Arabic script is itself difficult and error-prone, particularly because there is no unified orthography and many spellings are contextual or human-driven [2011.05152, 2511.16683].

Later research broadens the Tunizi landscape in two directions. First, a 2025 benchmark introduces a manually curated 100-example parallel Tunizi–Arabic-script–English dataset with POSITIVE, NEGATIVE, and NEUTRAL labels and evaluates LLMs on transliteration, translation, and sentiment analysis. It reports that Gemini 2.5 Flash performs best on transliteration and translation metrics, while GPT-4o Mini performs best on sentiment classification with Accuracy 0.60, Weighted F1 0.60, and Macro F1 0.45 [2511.16683]. Second, the multilingual fake-news corpus BOUTEF explicitly labels Tunisian Arabic in Arabic script as TUNDIALECT and Tunisian dialect written in Latin script as TUNARABIZI, showing that Tunizi appears in fake comments, genuine comments, and posts within a larger North African misinformation ecology [2606.00193]. These later developments do not redefine TUNIZI as a dataset, but they show that the category of Tunizi has become central across sentiment analysis, transliteration, LLM evaluation, and misinformation studies.

The original paper states that the dataset is publicly available on GitHub at `https://github.com/chaymafourati/TUNIZI-Sentiment-Analysis-Tunisian-Arabizi-Dataset` [2004.14303]. It does not specify a license, and it does not define predefined splits or additional metadata such as timestamps, user identifiers, or video IDs [2004.14303]. Natural future extensions identified in the source materials include larger corpora, additional sentiment labels such as neutral, aspect-based sentiment or sarcasm markers, Arabizi datasets for other North African dialects, and possible normalization layers from Arabizi to standardized Tunisian dialect, MSA, or Arabic script [2004.14303]. This suggests that TUNIZI should be understood both as a first-generation sentiment dataset and as a point of departure for a broader research program on Tunisian Arabic in Latin script.

Source: https://www.emergentmind.com/topics/tunizi