Technical Challenges of AI in Moderating Hate Speech Even in 2021

Why AI is still facing technical challenges in moderating hate speech in 2021?

The spread of misinformation and hate speech is increasing on multiple social media platforms affecting a certain group of people. Celebrities and politicians are experiencing the most as primary targets but that is affecting the minds of common people as well. The malicious digital content also contains hate speech regarding different ethnicity and minorities like LGBTQ. Hate speech travels faster than light on social media platforms. This can develop violence, riots, or other dangerous impacts in society. It is seen that AI models and deep learning algorithms are advancing as per time but it is still struggling in moderating hate speech. In this 21^st century, everyone has a smartphone and a smooth internet connection. Thus, it is very easy to spread hate speech through social media platforms within a second. AI is failing to differentiate between innocent content and online trolling/ hate speech. Let's look at the technical challenges faced by AI in moderating hate speech even in 2021.

There are multiple challenges faced by AI models to tackle hate speech efficiently. Firstly, it is an enormous responsibility to annotate a training set of data with the current real-time information. The AI models and deep learning algorithms do not have the differentiating power unless it is taught during the period of training. The training set lacks sufficient information on race, nationality, or religion and its concept on hate speech. Hate speech varies from place to place across the world because it is connected to culture. There is a change of mindset in every geographical area and one may not consider online content as hate speech that is unacceptable in another area. Thus, it is very difficult for AI to differentiate between the training data and real-time data.

Secondly, AI and deep learning algorithms fail to spot original hate speech on social media platforms due to the biases against some races. The developers in hi-tech companies are trying to automate hate speech detection in understanding the context of human language. The AI models calculate the possibility of words appearing in a specific sequence based on the training data. But the hate speech detector model utilizes only sample sentences in the training data to moderate content on social media platforms. Thus, the real-time data does not match the training data.

Since the most popular social media platform is Facebook, Facebook AI is trying to reduce the content on hate speech by investing in deep content semantic understanding with multimodal learning. The Silicon Valley giant has published a new model known as XLM-R that uses self-supervised training techniques to understand multiple language models. It is built on the Bidirectional Encoder Representations from Transformers (BERT) technique. It takes sentences, blanks out random words, and predicts alternate suitable words for the blank. It has also created NLP language models through XLM-R that consists of strengths from XLM and Robert to outperform traditional monolingual baselines to detect hate speech. The main aim of Facebook is to spot hate speech accurately for every language across the world. The tech giant is using RIO and Linformer to analyze content on Facebook and Instagram across the world. Google is also in a fight against hate speech with Conversational AI to automatically spot and moderate hate speech with the help of deep learning algorithms.