How Well Do Tweets Represent Sub-Dialects of Egyptian Arabic?

Mai Mohamed Eida, Mayar Nassar, Jonathan DunnNAACL 2024

TL;DR

This paper investigates whether tweets accurately represent sub-dialects of Egyptian Arabic (EA), focusing on Cairene Egyptian Arabic (CEA) and Sa'idi Egyptian Arabic (SEA). The study reveals that while CEA is well-represented in tweets, SEA is not, with SEA tweets often showing a prevalence of CEA markers and Modern Standard Arabic (MSA). This highlights a significant gap in current digital resources for less prestigious dialects and questions the reliability of geo-referenced tweets for capturing dialectal diversity.

What Problem Does This Paper Address?

The field of Arabic Natural Language Processing (NLP) has largely focused on dialectal Arabic (DA) at a country level, neglecting the rich variation within dialects. Egyptian Arabic (EA), for instance, has several distinct sub-dialects, with Cairene Egyptian Arabic (CEA) being the most widely studied and represented. However, less prestigious sub-dialects, such as Sa'idi Egyptian Arabic (SEA), are significantly under-represented in both research and digital resources. This poses a problem for developing inclusive NLP technologies, as the training data used for downstream applications (like machine translation or dialect identification) directly impacts their performance on less represented dialects. Furthermore, the interplay between dialect and register (e.g., spoken vs. written language) is not well understood across different EA sub-dialects. This paper aims to bridge this gap by assessing how well existing digital corpora, specifically tweets, represent EA sub-dialects, and whether register variation affects this representation.

Methods

The researchers employed a multi-faceted approach to analyze the representation of EA sub-dialects in tweets. They utilized morphological and grammatical features identified from ground-truth dialect surveys as a basis for measuring dialectal distance. These features, covering demonstratives, interrogatives, prepositions, adverbs, and negation particles, were selected for their distinct orthographic representations across sub-dialects.

To quantify representation, they compared the prevalence of these dialectal features in geo-referenced tweet datasets against expected patterns derived from dialect surveys. They also analyzed the usage of Modern Standard Arabic (MSA) versus Dialectal Arabic (DA) within tweets to understand register variation. For quantitative analysis, they developed regular expressions to capture feature variations and account for spelling mistakes. Qualitative analysis was also used for features where orthographic distinctions were less clear.

Datasets

The study utilized several datasets:

  • Cairo Baseline Corpus: A large corpus of geo-referenced tweets from Cairo (Dunn, 2020). This served as a baseline to validate the extraction method and assess the representation of CEA features.
  • Sub-Dialect Datasets: Three datasets of geo-tagged tweets from Arabic micro-dialect identification shared tasks:
  • MicroDialect Identification: (Abdul-Mageed et al., 2020b)
    • NADI 2020: (Abdul-Mageed et al., 2020a)
    • NADI 2021: (Abdul-Mageed et al., 2021)

These datasets were further categorized into those representing CEA cities and SEA cities, allowing for a comparative analysis of dialectal feature usage.

Key Insights and Ideas

  • CEA Dominance: Cairene Egyptian Arabic (CEA) features are overwhelmingly prevalent in tweets originating from Cairo, aligning with expectations.
  • SEA Underrepresentation: Conversely, Sa'idi Egyptian Arabic (SEA) features are significantly underrepresented in tweets from SEA regions. Instead, these tweets often exhibit a higher usage of CEA markers and Modern Standard Arabic (MSA).
  • Register Variation Matters: The study highlights that register variation is not uniform across sub-dialects. While CEA speakers freely use their dialect in tweets, SEA speakers appear to revert to MSA more frequently in this digital written context.
  • Questioning Geo-Referencing: The findings cast doubt on the sole reliance on geo-referenced tweets for accurately capturing dialectal diversity, especially for less prestigious dialects.
  • Resource Scarcity: The research underscores that EA sub-dialects, beyond CEA, are low-resource, necessitating the development of more inclusive datasets and NLP tools.

Key Results

The study found that:

  • In the Cairo Baseline Corpus, CEA features were highly prevalent, with many features showing over 80% usage, while SEA features were minimal.
  • Across the NADI2020, NADI2021, and MicroDialect datasets, SEA cities showed a significantly lower prevalence of expected SEA features compared to CEA cities. For instance, in the NADI2020 corpus, SEA features were only present in a small fraction of alternations, while CEA features dominated even in SEA locations.
  • Qualitative analysis revealed that SEA speakers in tweets often use CEA features or MSA, rather than distinct SEA markers, suggesting a shift in dialectal expression in digital written contexts.
  • Analysis of MSA vs. DA usage in tweets showed that SEA users utilized MSA approximately as much as DA, whereas CEA users favored DA significantly more.

Limitations and Future Work

The authors acknowledge several limitations:

  • Orthographic Representation: The selected morphological markers' orthographic representation is based on frequent patterns, not necessarily ground truth, which can lead to inconsistencies.
  • Feature Scope: The analysis was limited to specific morphological features, potentially missing significant lexical markers of SEA.
  • Dataset Size and Overlap: Some features were underrepresented or overrepresented due to dataset size, and the analysis was limited to cities explicitly identified with CEA and SEA, potentially missing nuances from other rural/urban sub-dialects.
  • Influence of Media: The popularity of media focusing on SEA settings might introduce external influences into SEA datasets.

Future work could involve:

  • Exploring lexical choices to better capture SEA dialectal markers.
  • Investigating the impact of media and cultural influences on SEA dialect representation.
  • Conducting larger-scale studies with more comprehensive datasets to explore possibilities of language change, internal migration, and demographic variables influencing dialect usage.
  • Developing more robust methods for dialect identification that account for register variation and the complexities of Arabic orthography.

Summary updated

BibTeX
@inproceedings{eida2024how,
  author = {Mai Mohamed Eida and Mayar Nassar and Jonathan Dunn},
  title = {{How Well Do Tweets Represent Sub-Dialects of Egyptian Arabic?}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
}

Loading related content...

Figure