市场营销学系学术讲座（11月11日）

From Hidden Neural Representations to Agentic Insight:

Discovering Beyond-Semantic Drivers of Review Helpfulness through Mechanistic Interpretability

Speaker: Xin (Shane) Wang

Professor of Marketing

Pamplin College of Business, Virginia Tech

Time: 13:30-15:00, Nov.11th(Wednesday), 2026

Venue: Siyuan Building, Room 624

MC: Prof. Liyin Jin

Abstract:

Text is central to marketing communication, yet research has focused largely on what text says rather than how it is said. This leaves important textual paralinguistic cues, such as formatting, structure, spacing, and punctuation, underexamined. These cues are difficult to extract, code, and analyze at scale because they are often context-dependent, difficult to predefine, and commonly removed during text preprocessing. To advance understanding of how textual paralinguistic features shape marketing-relevant outcomes, we develop an automated feature-discovery process that uses the internal representations of large language models (LLMs) as a source of data. The framework applies mechanistic interpretability techniques to extract sparse, interpretable features from LLM representations; uses supervised learning to identify features associated with focal outcomes; and deploys a multi-agent process to interpret candidate features. Applying the framework to Yelp restaurant reviews, we uncover a broad set of paralinguistic features associated with review helpfulness. We then focus on two organizational paralinguistic cues—formatted lists and structural discourse markers — and through controlled experiments, establish their positive causal effects on perceived review helpfulness and identify processing fluency as an underlying mechanism. Our findings deepen our understanding of textual paralanguage by uncovering context-dependent paralinguistic features that are difficult to detect using existing methods.

Bio:

Shane Wang is a Professor of Marketing at Pamplin College of Business, Virginia Tech. His research examines how emerging technologies transform business strategy and consumer behavior. Currently, his work and teaching focus on AI agents and synthetic data, with an emphasis on their strategic impact on businesses and consumer markets.His work has appeared in Marketing Science, Journal of Marketing Research, Journal of Marketing, Journal of Consumer Research, Management Science and International Journal of Research in Marketing.