Cleaning the Digital Source: IIT Guwahati Research Fixes Wikipedia’s Hidden AI Data Errors

Author : Yasaswini Sampathkumar

Published: 20 February 2026 Category: Research & Partnerships Office: Computer Science and Engineering

In an era where both humans and Artificial Intelligence rely on Wikipedia as a primary source of truth, even minor errors can have massive downstream consequences. At the India AI Impact Summit 2026, researchers from IIT Guwahati showcased a method to identify and correct Surface Name Errors (SNEs)—the subtle typos or incorrect links that currently affect 3% to 6% of all Wikipedia entity mentions.

Why "Small" Errors are a Big Problem

Whether it is a simple misspelling like “Parise” instead of “Paris” or extra words are mistakenly included in a link, these mistakes do more than just frustrate readers. Because Wikipedia is a core dataset for training machine learning models, these errors can degrade AI performance and spread misinformation at scale.

"We should not be trusting data from the web blindly," explains Prof. Amit Awekar of the Department of Computer Science and Engineering. "Good data is the beginning of any effective AI model."

A Data-Driven Solution for a Multilingual Web

The method developed by Prof. Awekar and Mr Anuj Khare (M.Tech, 2022) uses mathematical frequency patterns rather than language-specific rules. This makes it uniquely scalable across different alphabets and dialects.

The three-step "Quadruplet" approach scans links based on:

1. Context: Where the link appears and where it points.

2. Frequency: Flagging names that appear rarely or inconsistent with established patterns.

3. Classification: Categorizing errors into simple typing mistakes or complex structural link errors.

Proven accuracy across eight languages

The team tested their method on English, Sanskrit, German, Italian, Urdu, Hindi, Marathi, and Gujarati with high success. Most notably, the global Wikipedia community has already accepted more than 99% of the manual corrections suggested by the researchers—proving that this tool is a practical asset for the world’s most famous encyclopedia.