From Embeddings to Explainability: A Tutorial on Transformer-Based Text Analysis for Social and Behavioral Scientists

Abstract

Large language models and their use for text analysis have had a significant impact on psychology and the social and behavioral sciences in general. Key applications include the analysis of texts, such as social media posts, to infer psychological traits, as well as survey and interview analysis. In this tutorial paper, we demonstrate the use of the Python-based natural language processing software package transformers (and related modules from the huggingface universe) that allow for the automated classification of text inputs. In doing so, we rely on pre-trained transformer models which can be fine-tuned to a specific task and domain. The first proposed application of this model class is to use it as a feature extractor, allowing for the transformation of written text into real-valued numerical vectors (called "embeddings") that capture a text's semantic meaning. These vectors can, in turn, be used as input for a subsequent machine-learning model. The second presented application of transformer models is the end-to-end training (so-called "fine-tuning") of the model. This results in a direct prediction of the label within the same model that directly maps the text to the embeddings. While in the second case, results are usually better and training works more seamlessly, the model itself is often not directly interpretable. We showcase an alleviation of this issue via the application of post-hoc interpretability methods by calculating SHAP values and applying local interpretable model-agnostic explanations (LIME) in an attempt to explain the model's inner workings.