TL;DR
Researchers have introduced ArabGend, the largest manually annotated dataset for gender and location analysis on Arabic Twitter, featuring ~166K accounts. By leveraging linguistic markers in usernames and profile descriptions, they developed an inference model that achieves an 82.1% F1 score, significantly outperforming previous benchmarks and providing deep insights into the socio-cultural dynamics of the Arabic Twitter-sphere.
What Problem Does This Paper Address?
While gender inference on social media is a well-trodden path for English, the "Arabic Twitter-sphere" has remained relatively under-explored. This is a missed opportunity: Arabic is a morphologically rich language where nouns and adjectives often carry explicit gender markers (like the Taa Marbouta suffix for feminine forms).
However, identifying gender isn't as simple as looking at a profile. Twitter doesn't provide gender data, and self-reported information in the Arab region can be misleading or missing. Existing datasets were often small or biased toward specific dialects (like Egyptian). The authors sought to bridge this gap by creating a massive, diverse dataset and testing which features—usernames, tweets, profile pictures, or friend networks—actually move the needle for automatic gender classification.
Methods
The authors developed a multi-stage pipeline to build and validate their dataset (see Figure 1).
- Word List Creation: They built a dictionary of 713 gender-marked words (e.g., "lawyer (f)" vs "lawyer (m)") by analyzing the first words of user descriptions.
- Automatic Labeling: They matched these markers against 100M tweets to initially label 167K users.
- Manual Verification: A native Arabic-speaking expert reviewed the labels to filter out organizations, spam, and ambiguous terms (e.g., "Doctor" which is gender-neutral in some contexts).
- Classification: They trained Support Vector Machines (SVMs) using character n-grams (2-5) as features. They experimented with different inputs: single tweets, aggregated tweets, user descriptions, and usernames.
- Multi-modal Exploration: Beyond text, they tested deep learning tools for profile image analysis and explored "homophily"—the idea that users tend to follow others of the same gender.
Datasets
- ArabGend (New): The primary contribution, containing 166,540 users. The distribution is skewed: 80% male and 20% female. About 56% of these users (92,654) are mapped to specific Arab countries.
- ARAP Dataset: A smaller, existing dataset of 1,600 accounts used as a benchmark for comparative evaluation.
Key Insights and Ideas
- Usernames are King: Surprisingly, usernames proved to be the strongest signal for gender inference in Arabic, outperforming even the text of the tweets themselves. This suggests that Arabic users often incorporate gendered names or markers directly into their handles.
- The Engagement Paradox: While the dataset is 80% male, female accounts are significantly more "social." On average, female users have more than double the followers of their male counterparts and roughly 30% more friends (see Figure 4).
- Geographic Skew: The data highlights the dominance of the Gulf region on Twitter, with Saudi Arabia (SA) accounting for over 54% of all users in the dataset (see Figure 2).
- Professional Gaps: The analysis revealed stark gender divides in professions mentioned in bios. For instance, 98% of users identifying as "players" (sports) were male, whereas "translators" had the highest female representation at 36%.
- Topic Divergence: Using valence scores, the authors found that males tweet significantly more about politics and sports, while females focus more on family, society, and feelings.
Key Results
- Inference Accuracy: The best model, using usernames from the ArabGend training set, achieved an F1 score of 82.1% on the ARAP test data. This is a 47.3% improvement over the majority baseline.
- Profile Images: Manual annotation of profile images reached 81% accuracy, but automated deep learning tools lagged behind at 64%, often misclassifying females as males.
- Network Effects: Incorporating the gender of a user's friends (homophily) provided a modest boost, raising the F1 score from 82.1% to 82.9% in cases where the primary classifier was uncertain.
- Validation: The authors compared their findings with the Twitter Ads API and found a very close match in the overall female user percentage (19% vs 20%), validating their data collection method.
Limitations and Future Work
The authors acknowledge a significant geographic bias; because Twitter usage varies wildly across the Arab world, the statistics are heavily influenced by the Gulf region (especially Saudi Arabia) and may not perfectly represent the Maghreb.
Future work aims to move beyond the "first word" of a bio to analyze the entire user description for gender markers. They also plan to integrate profile image features more deeply into the textual models and explore age group prediction using similar self-disclosure patterns.
Summary updated