TamilMemes: Troll and Not-Troll Classified Memes with Latin Transcripts

A dataset of Tamil memes annotated for troll classification, created to facilitate computational modeling of trolling in Indian languages. The dataset, named TamilMemes, contains meme images categorized as "troll" or "not_troll" and includes Latin transcriptions of the text from the memes. It was released by researchers associated with the University of Galway and presented at the WILDRE-5 workshop in 2020.

Use Cases

• Train image classifiers for troll meme detection based on annotated meme images.
• Develop multimodal models that combine visual and textual features from meme images and their Latin transcriptions.
• Benchmark troll detection systems using the shared task evaluation metric of weighted average F1-score.
• Study region-specific humor and sarcasm in memes for content moderation research.

Strengths

• Dataset is explicitly annotated with binary "troll" or "not_troll" labels for classification tasks.
• Provides both meme images and corresponding Latin transcripted text, enabling multimodal analysis.
• Was used in a shared task with 10 system submissions, providing a benchmark context.

Limitations

• Dataset size, row count, and file formats are unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• The best-performing system in the shared task achieved a weighted average F1-score of 0.55, suggesting the classification task is challenging with existing methods.

Provenance

Source

Ollscoil na Gaillimhe – University of Galway

Collection Method

Manually annotated and released by researchers; details on gathering method are not specified.

Geography

Focus on Tamil language, suggesting content is region-specific to Tamil-speaking communities.