Automatically Explaining a Model: Using Deep Neural Networks to Generate Text From Causal Maps

Simulation models start as conceptual models, which list relevant factors and their relationships. In complex socio-environmental problems, these conceptual models are routinely created with participants, via a ‘participatory modeling’ approach. Transparency is a tenet of participatory modeling: participants should easily provide their input into the model-building process and see how that input is utilized. Although several elicitation methods are transparent, the resulting conceptual model can become too large and difficult to interpret. Usability studies have shown that participants struggle to interact with such large conceptual models, even if they contributed to creating parts of it. In this paper, we propose to automatically transform these large conceptual models into a more familiar format for participants: textual reports. We designed and implemented a process combining Natural Language Generation (via the deep learning GPT-3 model) and Network Science. Two case studies demonstrate that our prototype generates sentences that perform satisfactorily on several metrics.