ERC Starting Grant 949944 - (INTEGRATOR) "Incorporating Demographic Factors into Natural Language Processing Models"

Principal investigator:

Dirk HOVY (Bocconi University)


The goal of INTEGRATOR is to develop novel data sets, theories, and algorithms to incorporate demographic factors into language technology. This will improve performance of existing tools for all users, reduce demographic bias, and enable completely new applications.
Language reflects demographic factors like our age, gender, etc. People actively use this information to make inferences, but current language technology (NLP) fails to account for demographics, both in language understanding (e.g., sentiment analysis) and generation (e.g., chatbots). This failure prevents us from reaching human-like performance, limits possible future applications, and introduces systematic bias against underrepresented demographic groups. Solving demographic bias is one of the greatest challenges for current language technology. Failing to do so will limit the field and harm public trust in it. Bias in AI systems recently emerged as a severe problem for privacy, fairness, and ethics of AI. It is especially prevalent in language technology, due to language's rich demographic information. Since NLP is ubiquitous (translation, search, personal assistants, etc.), demographically biased models creates uneven access to vital technology.
Despite increased interest in demographics in NLP, there are no concerted efforts to integrate it: no theory, data sets, or algorithmic solutions. INTEGRATOR will address these by identifying which demographic factors affect NLP systems, devising a bias taxonomy and metrics, and creating new data. These will enable us to use transfer and reinforcement learning methods to build demographically aware input representations and systems that incorporate demographics to improve performance and reduce bias.
Demographically aware NLP will lead to high-performing, fair systems for text analysis and generation. This groundbreaking research advances our understanding of NLP, algorithmic fairness, and bias in AI, and creates new research resources and avenues.