Facebook's created a dataset of 10,000 ‘hateful’ memes and is offering prizes totaling $100,000 for developers who can use it to detect hate speech in memes.
The problem is 'multimodal' content containing both images and words; while the two elements of a meme may be individually acceptable, the combination might not, meaning the two elements need to be analyzed together.
For example, 'Love the way you smell today' superimposed on a picture of a rose isn't hateful - but the same words with a picture of a skunk don't carry quite the same message.
"To address this challenge, the research community is focused on building tools that take the different modalities present in a particular piece of content and then fuse them early in the classification process," the team explains.
"This approach enables the system to analyze the different modalities together, like people do."
Currently, according to a research paper, the best deep learning models achieve a success rate of only 64.7 per cent when identifying hateful memes, compared with a human success rate of 84.7 per cent.
The company defines 'hateful' content as a direct or indirect attack based on characteristics such as ethnicity, race, nationality, immigration status, religion, caste, sex, gender identity, sexual orientation and disability or disease.
"We define attack as violent or dehumanizing (comparing people to non-human things, e.g. animals) speech, statements of inferiority, and calls for exclusion or segregation. Mocking hate crime is also considered hate speech," Facebook adds.
The dataset has been created using existing memes, but with the original images replaced by similar pictures from Getty. Participants have until October to explore the image bank data and starter code and build models. They can then take part in a final competition at the NeurIPS machine learning conference in December, at which they will be presented with a new set of memes and asked to identify the 'hateful' ones.
The winner will receive $50,000, with smaller prizes for other participants. There's more information here.
Over the last few years, Facebook has invested heavily, to say the least, in automating content removal. In its latest Community Standards Enforcement Report (CSER), it says that 88.8 per cent of the hate speech content it removed in Q1 2020 was detected using AI, up from 80.2 per cent in the previous quarter.
If Facebook still feels its systems are incapable of dealing effectively with hateful memes, it's not clear what it expects from the competition; anybody coming up with something really effective could surely sell it to the company for a lot more than fifty grand.
Source: https://www.forbes.com/sites/emmawoollacott/2020/05/13/facebook-offers-50k-for-ai-that-can-detect-hateful-memes/?fbclid=IwAR0MfgB0eJ6osx9_Pvg5W4KuHs5bbNbSiehFLkkbmHc6mCrjpat9131hBLc#328f99b813d8
Comments
Post a Comment