KUALA LUMPUR, Jan 14 — An early version of an artificial intelligence (AI) programme, developed to help identify hateful social media posts within Malaysian society was made public today through the web page #TrackerBenci.
#TrackerBenci — which translates from Malay to “Hate Tracker” — was developed by KL-based think tank The Centre, which launched the microsite via a virtual conference to demonstrate it as a “proof of concept”.
“We hope #TrackerBenci can help policymakers and members of the public recognise and better understand the trends and themes of online hate speech in Malaysia,” The Centre’s chief executive officer, Dr Khairil Ahmad, said at the launch.
Researcher Tham Jia Vern said that #TrackerBenci is designed to help fill the gaps in current social media hate speech tracking systems, which often heavily rely on undertrained human moderators.
“The lack of investment in locally-trained and diverse panels of moderators leaves a gap in understanding what is hateful in certain cultural contexts,” she said.
She added that these moderators need to sift through the massive amounts of online content daily — making a human-only system unreliable in the long term.
Thus, she said, tools such as #TrackerBenci can be paired with human moderators, allowing for a much-needed hybrid approach to detecting and classifying harmful online content.
To note, #TrackerBenci is currently only able to process Twitter’s tweets, and is also not yet able to fully comprehend all forms of hate speech; largely due to the time and manpower limit researchers had to train the AI.
However, The Centre hopes to improve #TrackerBenci’s performance over time, and have also allowed users to further improve the AI by informing it when users find the AI to be wrong about what is hateful — through a component called the Benci Calculator.
According to Tham, training the AI involved showing it some 20,000 tweets, and then teaching it to differentiate between the tweets as either “hateful”, “a snarky reply towards hateful sentiments”, or “not hateful or irrelevant to the Malaysian context”.
These tweets were chosen by a team of six researchers, led by Tham and Nelleita Omar, who trawled through the social media site for posts they curated as relevant to Malaysian society.
The AI was then further trained to determine the intensity of the “hateful” tweets, by identifying them as either an “insult” or “threat to marginalisation”.
It also learned how to distinguish hateful tweets by categories and themes such as “violent threat”, “increases intergroup tension”, “sexism”, “related to one’s religion or religious beliefs”, and “perpetuates a negative stereotype”.
Former Parliament Speaker Tan Sri Mohamad Ariff Md Yusof, who made the opening address at the launch today, lauded the team involved, and said he was confident that their initiative would garner the support of the majority of parliamentarians and politicians.
The Centre was co-founded by Umno’s Health Minister Khairy Jamaluddin and his party information chief Shahril Hamdan, who is also an economic director at the Prime Minister’s Office.
However, neither men were present at the virtual launch, and checks on their social media show they have not publicised the project, nor their involvement in the matter in any recent post.