This project seeks to develop methodologies and tools for the monitoring, analysis, and measurement of trends in online hate speech in different languages, countries, and contexts, with a particular focus on the emerging trends of hate speech regarding COVID-19. This involves the undertaking a multi-stage programme of research, development, and policy engagement.
Background Literature Review and Toolkit Development
The initial stage of literature review would look at the state of the art of knowledge on producers of hate speech and hateful content, and also current understanding of consumer dynamics (i.e. who is listening to hate speech and how they are affected). There would be a particular focus on exploring the current state knowledge regarding the circulation of online hate speech related to the COVID-19 pandemic.
This stage would also place a major focus on computational methods for detecting and classifying pieces of text as ‘hateful’, particularly in multi-lingual and multi-country contexts. We would comprehensively review this work, identifying both methods and techniques and also open-source libraries and existing platforms. We would also set out what is known about the accuracy of these systems and address generic means of understanding accuracy in such classification systems.
Finally, we would look at what is known about detecting prevalence on social platforms, and tackle issues such as sampling strategy, as well as summarising the latest measurements that have been released about the prevalence of hateful issues and content. Sampling on social media is a complex challenge, as social platforms tend to divide themselves into a small number of highly used groups and pages and a long tail of smaller ones
Digital Platform Development
This activity, which is the core of the project, encompasses the development of a prototype digital platform which would flexibly allow for the measurement of trends in online hate speech in different languages, countries and contexts. Such a tool could be tuned to different local contexts and environments using local political and linguistic knowledge, and be used to serve up rapid reports on trends in online hate speech.
The core of the tool would be a multi-platform monitoring system which would systematically query the Application Programming Interfaces [APIs] of major social platforms (Facebook, Twitter, Instagram, Youtube and Reddit) for relevant social media data. This tool will incorporate a keyword search-based capacity; keywords would be defined by country or region of interest, but could, for example, contain political or religious topics, or other identity based topics which are fertile grounds for hate speech, as well as coronavirus-related keywords.
By querying for content relating to these topics over an extended period of time, we can build up a dataset of political, religious, coronavirus or other content relating to the country or region of interest. Then, a limited amount of manual content analysis can be applied to determine the level of hate speech around the topic of interest. This will be combined with large group search capacity and the monitoring of media outlets known for promoting hateful or divisive content. The output of this strand of work will be a tool that, with some input from a country expert, can be tuned to detect hate speech in different regions or countries of interest.