A Lightweight Contrastive System for Misinformation Detection in Social Media Tweets

Abstract

Priyam Saha*

A compact classification system was developed and submitted to Prompt RecOvery for MisInformation Detection (PROMID) Subtask 3 for the detection of misinformation in tweets about the Russia-Ukraine conflict on the Twitter platform, as provided by the workshop organizers at Forum for Information Retrieval Evaluation (FIRE) 2025, held at the Indian Institute of Technology (BHU) Varanasi, India. The proposed solution combines a frozen RoBERTa encoder, a small projection head trained with a supervised contrastive objective, and a lightweight classifier trained jointly with binary cross-entropy. Design choices were driven by compute and memory constraints; several practical implementation details and evaluation outcomes are reported to support reproducibility of results. The submission of predictions computed on the test dataset as provided by the organizers was made on the Codabench platform as team 'priyam_saha17' and submission ID as 431064. On the official test set, the methodology produced a weighted F1 score of 0.82 (precision 0.87, recall 0.80), thereby securing the 5th rank in the track leaderboard, accessible at Link. For comparison, the leaderboard was topped by team 'ClimateSense', who achieved a weighted F1 score of 0.91 (precision 0.91, recall 0.91). The approach, training pipeline, and error analysis are documented in order to assist future participants and applied researchers working under limited resource conditions.

HTML PDF