Social Embeddings: Concept and Initial Investigation

Abstract

We introduce social embeddings as a compact, yet semantics-preserving, mathematical representation of social situations. Social embeddings are constructed by leveraging pre-trained large language models: we automatically generate a textual description of the social environment of a robot, and use pre-trained text embeddings to generate a vector representation of the social scene. The article presents the details of the methodology, and analyses key properties of these embeddings, including their ability to measure social ‘similarity’. We argue that social embeddings are a quantitative pseudo-metric for social situations, we demonstrate their operationalization on actual social robots, and discuss their potential applications.