Natural Language Explanations of Dataset Patterns

Ruiqi Zhong

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2025-78
May 15, 2025

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-78.pdf

Explaining patterns in large datasets is essential for empirical science, engineering, and business. For example, by analyzing a dataset of symptom descriptions, a doctor may discover that “tingling in the thumb” is a good explanatory variable for disease X. However, existing methods (e.g. regression) are primarily designed to analyze real-valued datasets and explain patterns in mathematical formulas (e.g. F = kx + b).

This thesis proposes metrics and methods for discovering and explaining dataset patterns in structured modalities (text/images) using natural language strings such as “tingling in the thumb”. We evaluate the explanations based on the predictive power they give to humans, which differs from common metrics based on human ratings or similarity to human demonstrations. We then generate dataset explanations by optimizing them against our evaluation metric, with the help of language models. Concretely, we sample candidate explanations from language models and select the highest-scoring one under our evaluation.

Based on these principles, we build a general framework, “statistical models with natural language parameters”, which allows us to explain distributional differences, clusters, and time-series in real-world datasets with structured modalities. Additionally, our metric can evaluate explanations of model decisions by treating them as explanations of datasets, which consist of the model’s input-output behavior. Using this approach, we show that language models are still far from explaining themselves as of 2024. Our contribution paves the way for helping humans understand complex datasets and systems, thereby accelerating scientific discovery and advancing explainable AI systems.

Advisor: Daniel Klein and Jacob Steinhardt

"; ?>

BibTeX citation:

@phdthesis{Zhong:EECS-2025-78,
    Author = {Zhong, Ruiqi},
    Title = {Natural Language Explanations of Dataset Patterns},
    School = {EECS Department, University of California, Berkeley},
    Year = {2025},
    Month = {May},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-78.html},
    Number = {UCB/EECS-2025-78},
    Abstract = {Explaining patterns in large datasets is essential for empirical science, engineering, and business. For example, by analyzing a dataset of symptom descriptions, a doctor may discover that “tingling in the thumb” is a good explanatory variable for disease X. However, existing methods (e.g. regression) are primarily designed to analyze real-valued datasets and explain patterns in mathematical formulas (e.g. F = kx + b).

This thesis proposes metrics and methods for discovering and explaining dataset patterns in structured modalities (text/images) using natural language strings such as “tingling in the thumb”. We evaluate the explanations based on the predictive power they give to humans, which differs from common metrics based on human ratings or similarity to human demonstrations. We then generate dataset explanations by optimizing them against our evaluation metric, with the help of language models. Concretely, we sample candidate explanations from language models and select the highest-scoring one under our evaluation.

Based on these principles, we build a general framework, “statistical models with natural language parameters”, which allows us to explain distributional differences, clusters, and time-series in real-world datasets with structured modalities. Additionally, our metric can evaluate explanations of model decisions by treating them as explanations of datasets, which consist of the model’s input-output behavior. Using this approach, we show that language models are still far from explaining themselves as of 2024. Our contribution paves the way for helping humans understand complex datasets and systems, thereby accelerating scientific discovery and advancing explainable AI systems.}
}

EndNote citation:

%0 Thesis
%A Zhong, Ruiqi
%T Natural Language Explanations of Dataset Patterns
%I EECS Department, University of California, Berkeley
%D 2025
%8 May 15
%@ UCB/EECS-2025-78
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-78.html
%F Zhong:EECS-2025-78