SarcParsBERT-CRNN: Sarcasm Detection in Persian Tweets Using ParsBERT and a CNN-RNN-Attention Head (Technical Note)

SarcParsBERT-CRNN is a technical note describing a Persian sarcasm detection pipeline for social-media text. The approach combines a pretrained Persian Transformer encoder (ParsBERT) with a convolutional–recurrent classification head (1D CNN followed by bidirectional recurrent layers such as BiLSTM/BiGRU) and an attention mechanism to emphasize informative tokens/spans. The preprocessing pipeline is tailored to noisy Persian tweets and includes emoji-to-text conversion, removal of URLs/mentions/hashtags and punctuation, number normalization (e.g., replacing digits with a dedicated token), whitespace cleanup, and Persian normalization (e.g., unifying Arabic/Persian character variants and handling half-spaces / ZWNJ).In experiments on a manually labeled dataset of 2,940 Persian tweets (sarcastic vs. non-sarcastic), the proposed configuration reports 76% accuracy under a fixed train/validation/test split.This Zenodo record provides the technical note and is intended to be linked to the corresponding open-source implementation archived on Zenodo via the GitHub release.