Skip to yearly menu bar Skip to main content


Poster
in
Workshop: NeurIPS'24 Workshop on Causal Representation Learning

LLMs as Emotion Analyzers for Causal Models: Partial Identification with Fuzzy Interval Data

Huidi Ma · Wendao Xue · Yifan Yu


Abstract:

Information systems (IS) researchers often use machine learning algorithms to recognize emotions in massive and unstructured online content and then study the causal effect of emotions on various business outcomes. Large language models (LLMs) as automatic emotion analyzers offer new opportunities for IS researchers to advance the causal understanding of emotions. Nevertheless, we show that directly plugging LLM-generated emotional variables into econometric models can induce bias in causal estimation because LLMs are imperfect in emotion recognition, which adds noise to the causal models. We propose a novel algorithm to correct such bias. A key feature is that the algorithm considers predictions generated by different LLMs to form fuzzy interval data. Then, partial identification of causal parameters is achieved. The algorithm meets three design requirements. First, it is unsupervised, meaning it does not need additional labeled data to correct the causal estimation. Second, it is flexible in incorporating the predictions of different LLMs (leveraging the "wisdom of the AI crowd") and can be easily adapted to various causal models. Third, the causal estimators are theoretically guaranteed to achieve consistency. This work provides important implications for causal inference with LLMs, the causal study of emotions, and prescriptive analytics for fuzzy interval data.

Chat is not available.