Meta разработала Sphere AI algorithm, which automatically fact-checks citations in Wikipedia. It can also propose alternatives if a passage is found to be inaccurate.
(1/2) Today, we’re announcing the first model capable of automatically verifying hundreds of thousands of citations at once. Read more: https://t.co/c0nnvhuQx7 pic.twitter.com/UTSEYUmCRZ
— Meta AI (@MetaAI) July 11, 2022
Sphere uses a natural-language understanding (NLU) model that attempts to understand the relationships between words and phrases in a sentence.
Sphere’s database comprises 134 million web pages. The natural-language understanding algorithm analyses the citations in encyclopedia articles and searches for a single source to verify each assertion.
According to company representatives, Sphere can analyse “tens of thousands” of citations simultaneously. Meta noted that the encyclopedia is updated with around 17,000 articles at once, containing numerous citations that are hard to verify manually.
To illustrate the AI’s capabilities, Meta showed an example of an incomplete quote found by the model on the Wikipedia page about the Blackfeet. In the English-language version of the article, Joe Hipp is mentioned — the first member of the nation to contend for the world heavyweight title according to the WBA.
However, corroboration for this fact could not be found on the linked site. A search of Sphere’s database yielded a more suitable quote in a 2015 article in the Great Falls Tribune. The algorithm highlighted the following excerpt:
In 1989, late in his career, [Marvin] Camel fought Joe Hipp of the Blackfeet nation. Hipp, who became the first Native American to challenge the world heavyweight champion, said that the fight was one of the strangest in his career.
In the excerpt, boxing is not explicitly mentioned. Sphere identified a suitable link thanks to its natural-language understanding capabilities, the developers said.
The company believes that in the future the tool will help combat misinformation on Facebook.
In broader terms, we hope that our work can be used for fact-checking and increasing the overall reliability of information on the Internet,
— the model’s creators said.
In addition, Meta plans to create a platform that editors of Wikipedia can use to systematically verify and correct citations.
The model’s source code is available on GitHub.
In July Meta unveiled an AI translator, capable of supporting 200 languages.
In April, the tech giant spoke about the development of the open-source language model OPT-175B.
Subscribe to ForkLog news on Telegram: ForkLog AI — all the news from the world of AI!