CS491 / CS492 Senior Design Project · Bilkent University

TRACED

Turkish Retrieval-Augmented Claim Evaluation and Detection

Pronounced like the English word “traced” (/treɪst/)

About

We propose an analyst web application that discovers Turkish claims in X (Twitter) posts, investigates them using external evidence, and tracks how they spread and change over time. It shows fact-checkers which claims are spreading fastest, so they know what to check first, while a person always makes the final decision. The project combines a working platform with a Turkish evidence-ranking model that we train ourselves and a measured evaluation against simpler baselines.

How it works

Each claim reaches the analyst with its evidence and its spreadPostsrecorded collections,live X if available,or claims typed inClaim detectioncheckable claimssplit from opinions;date, place, entitiesGroupingsentence embeddingsmerge paraphrasesinto one clusterSpread analysisgrowth sets priority;repost graph andcoordination signalsEvidence searchfact-checks, officialsources and news,repeated if neededEvidence rankingour fine-tunedTurkish model ordersthe passagesAssessmentan existing LLM picksone of four labels,citing the passagesAnalyst reviewpriority queue,evidence and spread;a human decidesThe Turkish ranking model we train ourselves

Team

Supervisor
Prof. Dr. Özgür Ulusoy
Innovation expert
Mesut Gözütok

Reports

Reports will be published here as they are submitted.