mining meaning: Sentence-level embeddings and semantic similarity for political texts
Natural Language Processing (NLP) techniques are an increasingly valuable component of the social science toolkit. A subset of NLP tools – embedding models – allows researchers to analyze the semantic meaning of text to a limited extent. This research note demonstrates the value of sentence-level (as opposed to word-level) embedding models, specifically for the purpose of characterizing the semantic similarity of a set of texts. I show how scholars can (1) validate qualitatively defined coding-schemes using semantic similarity analysis; and (2) test key properties of a corpus using the same method.
the neural correlates of partisanship
What does it mean to self-identify as a political partisan in the U.S. today? Is there a “partisan brain” that acts as a comprehensive filter and template for structuring information about the world into “us vs. them”? Or is partisanship still related to policy preferences in some fundamental way? Nir Jacoby (Dartmouth), Emily Falk (UPenn), Kevin Ochsner (NYU), Jacob Perelman (UPenn), Emile Bruneau and I consider these questions using fMRI.
information in story form
My interest in learning, memory, and experiments has led to preliminary investigation of the causal effect of narrative structures. While much of the priming literature as uncovered a differential effect of information structured in "story form", the underlying mechanisms, heterogeneity of the "story effect", and consequences for belief formation and strength remain under-explored.
Tools for the study of emotion in neuroimaging
In the course of developing a study that induced two emotions (fear and disgust) in participants undergoing fMRI scanning, we discovered several gaps in publicly available research tools. As such, I am preparing two manuscripts to supplement the main imaging paper: “First-Person Fear and Disgust Induction in fMRI: An ALE Meta-Analysis” and “Emotion Induction using Guided Narratives: 64 Stimuli for Fear, Disgust, and Calm.”
The meta-analysis uses activation likelihood estimation (ALE) to evaluate common activations across first-person emotion inductions for fear and disgust. We explicitly exclude studies using faces (3rd-person induction) and conditioned pain paradigms in order to focus on first-person, naturalistic stimuli.
Our study induced emotions using short, guided narratives, which I wrote. The final vignettes were selected from a larger set evaluated by human raters on MTurk as well as an automated system that rated narrativity, and several other latent properties of the text. The final product is a set of stimuli that are standardized on ease of imagining (to avoid task demand differences), emotion discriminability, and more conventional properties. Multidimensional scaling performed on the distance between vectors of active voxels within ALE-defined ROIs indicates that the stories successfully generate a space of mental representation defined by the neutral condition on one pole and the target emotion on the other. Non-target conditions generally fall in between.