Using word embeddings to generate data-driven human agent decision-making from natural language

Abstract

Generating replicable and empirically valid models of human decision-making is crucial for the scientific accuracy and reproducibility of agent-based models. A two-fold challenge in developing models of decision-making is a lack of high resolution and high quality behavioral data and the need for more transparent means of translating these data into models. A common and largely successful approach to modeling is hand-crafting agent decision heuristics from qualitative field interviews. This empirically-based, qualitative approach successfully incorporates contextual decision making, heterogeneous preferences, and decision strategies. However, it is labor intensive and often leads to models that are hard to replicate, thereby limiting the scale and scope over which such methods can be usefully applied. A potential solution to these problems is provided by new approaches in natural language processing, which can use textual sources ranging from field interview transcripts to unstructured data from the web to capture and represent human cognition. Here we use word embeddings, a vector-based representation of language, to create agents that reason using similarity comparison. This approach proves to be effective at mirroring theoretical expectations for human decision biases across a range of natural language decision-making tasks. We provide a proof-of-concept agent-based model that illustrates how the agents we create can be readily deployed to study cultural diffusion. The agent-based model replicates previously found results with the added benefit of qualitative interpretability. The agent architecture we propose is able to mirror human likelihood assessments from natural language and offers a new way to model agent cognitive processes for a broad array of agent-based modeling use cases.