Abstract
The connection between messaging and action is fundamental both to web applications, such as web search and sentiment analysis, and to economics. However, while prominent online applications exploit messaging in natural (human) language in order to predict non-strategic action selection, the economics literature focuses on the connection between structured stylized messaging to strategic decisions in games and multi-agent encounters. This paper aims to connect these two strands of research, which we consider highly timely and important due to the vast online textual communication on the web. Particularly, we introduce the following question: can free text expressed in natural language serve for the prediction of action selection in an economic context, modeled as a game? In order to initiate the research on this question, we introduce the study of an individual's action prediction in a one-shot game based on free text he/she provides, while being unaware of the game to be played. We approach the problem by attributing commonsensical personality attributes via crowd-sourcing to free texts written by individuals, and employing transductive learning to predict actions taken by these individuals in one-shot games based on these attributes. Our approach allows us to train a single classifier that can make predictions with respect to actions taken in multiple games. In experiments with three well-studied games, our algorithm compares favorably with strong alternative approaches. In ablation analysis, we demonstrate the importance of our modeling choices—the representation of the text with the commonsensical personality attributes and our classifier—to the predictive power of our model.
Introduction
In modern web applications, predicting user behavior—such as clicking an advertisement or determining the sentiment of a review—is a well-established field. These are typically "non-strategic" actions. In contrast, economic theory focuses on "strategic" behavior, where decisions depend on the anticipated actions of others. Traditionally, the link between communication and strategic choice has been studied through "cheap talk," which involves highly structured, game-specific messages.
This research bridges the gap between natural language processing (NLP) and strategic economic modeling. We investigate whether "free text"—unstructured natural language produced in daily life—can predict how individuals navigate conflict, cooperation, and coordination in economic games. The core hypothesis is that personal narratives contain latent personality signals that manifest as consistent strategic tendencies across different scenarios.
Methodology
Experimental Framework
The study employed a two-step experimental design with 271 university students. First, participants wrote a personal story of at least 1,000 characters on a topic of their choice, unaware of any subsequent tasks. Second, they played three one-shot strategic games: Chicken (a conflict game), Box (a congestion/resource game), and Door (a coordination game). The payoff matrices for these games are illustrated in Figure 1.
Feature Engineering and Learning
To translate text into a predictive signal, the researchers bypassed standard automated sentiment analysis in favor of a more nuanced approach:
- Commonsensical Attributes: 24 personality traits (e.g., adventurous, cynical, rational, insecure) were identified by graduate students as relevant to strategic decision-making.
- Crowd-Sourced Annotation: Human annotators rated each personal story against these 24 attributes.
- Transductive Attribute Clustering (TAC): A transductive learning model was used to cluster individuals based on their personality features before assigning action labels. This allows the model to leverage the feature distribution of the entire population (training and test sets) simultaneously.
Results and Key Insights
The "Unaware" Signal
A primary finding is that text written without knowledge of the future game still contains sufficient signal to predict strategic choices. For example, a narrative about a vacation can reveal whether a participant is likely to "speed" or "stop" in a game of Chicken. Interestingly, the 24 "commonsensical" attributes outperformed professional psychological frameworks like the "Big Five" in predictive accuracy for these specific tasks.
Performance Benchmarks
The TAC model consistently outperformed the Majority Vote Classifier (MVC) on Macro Average F1 scores, which is particularly significant for imbalanced datasets where one action is dominant.
- Chicken Game: TAC achieved a Macro Average F1 of 57.82, whereas the MVC fails to identify minority class actions.
- Ablation Analysis: Replacing the 24 custom attributes with IBM Personality Insights' 13 traits (TAC-IBM-13) resulted in a performance drop, validating the relevance of the chosen commonsensical features.
- Outlier Detection: While MVC may have high raw accuracy due to "focal" actions (common choices), TAC was significantly better at identifying individuals who make unconventional or aggressive strategic moves.
Discussion and Future Work
While this study demonstrates the feasibility of predicting strategic moves from free text, several challenges remain. The current reliance on human crowd-workers for attribute extraction creates a manual bottleneck; future systems would require automated NLP classifiers trained on these specific attributes. Additionally, the sample size (271 students) and demographic homogeneity limit generalizability. Future research should explore larger, more diverse datasets and the impact of "strategic writing," where individuals might intentionally use text to deceive opponents in a competitive environment.
Summary updated