"Artificially Generated Causal Data Across Domains"

"Causal domain knowledge, which refers to information about cause-and-effect relationshipsexpressed in text, is often not fully aligned with guidelines for generating causal models. In this paper, weevaluate the ability of LLMs to transform textual causal domain knowledge bases into a formal causalrepresentation, by following established causal data science guidelines. To this end, we introduce twonovel tasks. The first task is concerned with transforming causal domain knowledge, described as textentities, into a set of modular causal variables. The second task is concerned with identifying interactionentities that could potentially violate the transitivity principle of the modeled causal relations. We evaluatea diverse set of LLMs across multiple domains and datasets."