Sophisticated hostile social media messages are being ignored by artificial intelligence tools

A report by NATO’s Center for Strategic Communications of Excellence (StratCom COE) has warned that many of the artificial intelligence (AI) tools used to monitor social media posts are too obvious and met with confusion. difficulty in detecting misinformation and sophisticated hostile messages.

Many machine learning models allow platforms, companies, and governments to estimate the emotions of online posts and videos.

However, a week before the Cannes World Festival of Artificial Intelligence (February 9-11), a group of experts said the majority of these AI-based systems rely on understanding the emotions behind a message. , which is not as obvious as initially thought.

Their research, published as part of a collaboration between the University of Portsmouth and the NATO Center for Strategic Communications Excellence research report, explores trends in AI. It outlines the limitations of these open source sentiment strategies and suggested ways to improve them.

The goal of online breach text is to attack an individual, group, organization or country in a way that is difficult for AI to detect upon analysis.

Dr Alexander Gegov, Computational Intelligence Reader and University of Portsmouth Research Lead for the NATO StratCom COE, said: “Small breaches are very dangerous for social media platform because they can often resonate with people of similar beliefs and help spread malicious or hostile messages.”

“Evaluating emotions online is challenging, but in this report we’ve demonstrated that there are ways we can improve our usual pre-processing. It’s time to go beyond that. beyond simple polarizing emotions and teach the AI ​​to assess the context of a conversation.”

The authors say Google Jigsaw’s emotion classifier is an interesting addition to online polar emotion analysis, but its classifier cannot distinguish between readers’ reactions to malicious comments harm or someone is spreading. hate speech.

They found another approach, called rhetorical structure theory (RST), to be a more powerful and effective way to analyze violations. In a way, it mimics how the brain unconsciously considers different parts of a sentence by assigning importance to certain words or phrases.

For example, ‘Today sucks’ and ‘It was a beautiful dress’ both contain the positive word ‘beautiful’. But ‘pretty’ can also enhance the sentiment of words around it, for example ‘bad’.

Study co-leader Djamila Ouelhadj, Professor of Operations and Analytical Studies at the University of Portsmouth, explains: “It is clear that text analysis alone is not enough when trying to classify forms of hate speech. more sophisticated.

“Our study with the NATO StratCom COE suggested several recommendations on how to improve artificial intelligence tools to address these limitations.

“Understanding how an individual assembles a message yields a rich, untapped source of information that can provide the analyst with the ‘story’ of how and why the message was assembled.

“For example, when analyzing messages and tweets from offensive or anti-Western groups and individuals, the RST model can tell us how radical a group is, based on how confident they are about it. the topic they are broadcasting.

“It can also help detect if someone is being groomed or radicalized by measuring the level of insecurity that person expresses when communicating their ‘opinion’.”

The team created a series of datasets to understand the violations and tested them in English and Russian texts.

They took a sample of 500 Russian-language messages from a Kremlin-linked Telegram channel discussing the Ukraine war and analyzed their level of hostility using a Google Jigsaw model.

The translated text is less malicious than the original Russian text. This highlights that when the message is translated by the AI ​​from its original language, some malicious inferences are ignored or ignored. The effect can be even stronger in breach analysis, where hostility is not apparent.

To overcome this, the paper says online translators can be fine-tuned and tailored to specific countries and languages ​​by region.

Dr Gundars Bergmanis-Korāts, Senior Specialist at NATO StratCom COE, said: “Governments, organizations and agencies of NATO and allied nations must address current AI challenges and focus on focusing on AI challenges. the adjustment of local language specifics to ensure equal IEA capabilities.”

“Military and government organizations leverage machine learning tools to detect, measure, and mitigate misinformation online, and measure the effectiveness and reach of communications. Therefore, it is very important to understand the audience by analyzing the communication context.”

Last year, the United States spent more than half a million dollars developing a artificial intelligence model can automatically detect and prevent violations on social media.

Dr. Gegov added, “Often, the overall nature of a social media message is hidden between less relevant sentences, which is why the manual filtering and post-processing steps on the platform is necessary.”

“This probably won’t change anytime soon, but while there’s no ‘one tool that does it all’, we explore some simple tricks that data analysts and AI enthusiasts can take to potentially increase the performance of their word processing workflows.”

“We also encourage social media custodians to become more transparent about what systems they currently use.”