Facebook & Its Tumultuous Relationship With AI-Based Content Moderation

Download our Mobile App

During a press meet recently, a Facebook spokesperson said that the social media giant would be redoubling its efforts to counter ‘harmful content’ on its platform using artificial intelligence. Reportedly, Ryan Barnes, the Facebook Product Manager of Community Integrity, said that the company would use AI to prioritise harmful content. This move is targeting at helping its over 15,000 human reviewers and moderators in dealing with reported contents.

Barnes said during the press interaction, “We want to make sure we’re getting to the worst of the worst, prioritising real-world imminent harm above all.” 

With that being said, there have been numerous attempts in the past to bring AI into the content moderation process on Facebook’s platforms. However, not all of them have met with success. We track down some of the major efforts of Facebook in the past and how it has fared in tackling the issue.

Facebook’s Efforts Towards AI-Based Moderation

In the past, Facebook has used an XML method which uses a single shared encoder to train massive amounts of multilingual data. This provided an improvement over both supervised and unsupervised machine translation of low-resource languages for better detection of hate speech and harmful contents even in languages other than English. This system enabled the quality of classifiers to apply training in one language, for most cases, English, to be applied across other languages. This method was able to proactively detect harmful language and content in about 40 languages.

This method soon succeeded by Whole Post Integrity Embeddings (WPIE). WPIE is a pre-trained representation of content for integrity problems. As compared to previous systems, the WPIE method was trained on a larger set of violations and training data. While introducing this method, Facebook said in its blog that the method “improves performance across modalities by using focal loss, which prevents easy-to-classify examples from overwhelming the detector during training, along with gradient blending, which computes an optimal blend of modalities based on their overfitting behaviour.”

Stay Connected

Get the latest updates and relevant offers by sharing your email.

Facebook claimed that upon deployment, these tools have helped in substantially improving the performance of their integrity tools. For example, this tool was able to help in detecting almost 97.6% of 4.4 million drug sale content hosted on the platform in 2019.

Earlier this year, as the COVID-19 situation was looming large, Facebook started utilising SimSearchNet, a convolutional neural net-based model, built originally to detect near-exact duplicates, for fighting misinformation. The company said that SimSearchNet was helping in end-to-end image indexing to recognise and flag near-duplicate matches.

Just recently, Facebook introduced its machine translation model called M2M-100, which has been trained on 2,200 languages — about ten times the amount of training data used in the preceding model. As per the company, this model was built as a many-to-many data set with 7.5 billion sentences for 100 languages using novel mining techniques. The resulting parameters capture information from related languages and reflect a diverse script of languages and morphology. One of the salient features of this technique was found to be the fact that it did not require English as a link between two languages. Meaning, a language can be translated to another without having to be first translated into English. The ultimate goal of this model is to perform bidirectional translation between 7,000 languages to particularly benefit low-resource languages. Apart from its obvious application in communication, Facebook anticipated that the M2M-100 model would help in content moderation across a larger set of languages.

How Successful Are These Moderation Techniques?

In ways, more…