Estimating Context-Specific Values from Natural Language

Values are the abstract motivations that justify opinions and actions. The pursuit of values drives human behavior and promotes cooperation. Existing research is focused on general (e.g., Schwartz) values that transcend contexts. However, the context-specific nature of values must be considered to (1) understand human decisions, and (2) engineer intelligent agents that can elicit human values and take value-aligned actions. Further, in practical applications (e.g., to conduct meaningful conversations or to identify online trends), artificial agents should be able to understand values on the fly from natural language.

We outline an approach for estimating context-specific values from text. At first, the values relevant to a context must be identified. To this end, we propose Axies, a hybrid (human and AI) methodology to identify context-specific values. Then, we examine the effectiveness of NLP models in classifying values in text. As context influences how we express values in natural language, we investigate the extent to which the learned value rhetoric can be transferred across contexts. Subsequently, we propose explainability techniques to inspect whether value classifiers have learned the context-specific connotations of values. Finally, we combine the steps above into a single method for swiftly estimating context-specific values from users.

Enrico Liscio is a PhD candidate in the Interactive Intelligence Group at TU Delft and part of the Hybrid Intelligence Centre. He obtained cum laude MSc. in Systems and Control from TU Delft (the Netherlands, 2017) and cum laude BSc. in Automation Engineering from the University of Bologna (Italy, 2015). Between his MSc. studies and the current position, he worked for 2.5 years as deep learning developer and technical project lead at Fizyr (the Netherlands).