An Overview of Computational Approaches for Analyzing Interpretation

Philipp Blandfort, Jörn Hees, Desmond Upton PattonarXiv 2018 · OpenReview

Abstract

It is said that beauty is in the eye of the beholder. But how exactly can we characterize such discrepancies in interpretation? For example, are there any specific features of an image that makes person A regard an image as beautiful while person B finds the same image displeasing? Such questions ultimately aim at explaining our individual ways of interpretation, an intention that has been of fundamental importance to the social sciences from the beginning. More recently, advances in computer science brought up two related questions: First, can computational tools be adopted for analyzing ways of interpretation? Second, what if the "beholder" is a computer model, i.e., how can we explain a computer model's point of view? Numerous efforts have been made regarding both of these points, while many existing approaches focus on particular aspects and are still rather separate. With this paper, in order to connect these approaches we introduce a theoretical framework for analyzing interpretation, which is applicable to interpretation of both human beings and computer models. We give an overview of relevant computational approaches from various fields, and discuss the most common and promising application areas. The focus of this paper lies on interpretation of text and image data, while many of the presented approaches are applicable to other types of data as well.

Introduction: The Subjectivity of the Beholder

Interpretation is inherently subjective. Whether it is a human judging the aesthetic quality of a photograph or a machine learning model classifying a sentiment, the process involves mapping raw data to a specific meaning or outcome. In the social sciences, understanding these individual perspectives has been a cornerstone of research. However, as artificial intelligence becomes increasingly integrated into decision-making processes, a dual challenge has emerged: we need computational tools to analyze human subjectivity, and we need to understand the "subjectivity" or internal logic of the AI models themselves.

Historically, these two problems have been treated separately. Explainable AI (XAI) researchers focus on the technical mechanics of neural networks, while social scientists employ qualitative methods or manual coding. This paper proposes a unified framework to bridge this gap, providing a formal language to analyze and compare the "points of view" of both humans and machines.

A Unified Framework for Interpretation

The core of the proposed approach is a mathematical abstraction known as the Interpretation Function ($f_b$). In this framework, any entity that interprets data—whether a person, a demographic group, or a computer model—is defined as a Bearer ($b$). The interpretation function maps a set of inputs ($I_b$) to a set of meanings or outputs ($M_b$).

The proposed framework for interpretation analysis, characterizing the latent interpretation function from given inputs and outputs.

By adopting a "black-box" assumption, the framework focuses on the relationship between inputs and outputs rather than the internal architecture of the bearer. This allows for a consistent analysis across different types of bearers. As shown in Figure 1, the goal of interpretation analysis is to characterize this latent function using observed data, effectively reverse-engineering the "logic" behind a particular perspective.

Taxonomy of Computational Methodologies

The paper categorizes existing computational approaches into four primary domains, each offering different strengths for characterizing interpretation functions.

  1. Statistical Methods: These utilize classical tools like correlation coefficients and Canonical Correlation Analysis (CCA) to identify linear relationships between data features and interpretation results. They are robust for testing significance but may struggle with complex, non-linear data.
  2. Pattern Mining: This approach seeks to discover "association rules"—explicit "if-then" statements that describe conditions under which a certain interpretation occurs. While highly interpretable, the number of rules can become overwhelming in large datasets.
  3. Model-based Approaches: These involve training a "proxy" model to emulate the bearer's behavior. Once the proxy is trained, it can be analyzed using techniques like heatmapping (saliency maps), prototype identification, or ablation studies to see which features are most influential.
  4. Visualization Techniques: Methods like t-SNE or PCA are used to project high-dimensional interpretation data into two or three dimensions, allowing researchers to visually inspect clusters and outliers in perspectives.

Practical Applications and Data Contexts

While the framework is general, the paper highlights its utility in text and image analysis. For instance, in image preference datasets, the framework can help identify whether "Humor" or "Explosions" are the primary drivers for a user's "Like" rating. In the realm of XAI, it can be used on standard benchmarks like ImageNet to visualize why a model confuses certain dog breeds.

Beyond these, the paper discusses applications in:

  • Social Media: Analyzing the nuances of sarcasm, aggression, or identifying potential for violence in online discourse.
  • Scientific Research: Interpreting complex data such as gene expression patterns, medical imaging, or EEG signals.

Critical Insights and Trade-offs

A key insight of this work is that the most profound understanding often comes from Sample-wise Merging—comparing the outputs of two different bearers on the same data. By analyzing the points of disagreement, researchers can pinpoint the specific features that define a unique perspective or a systematic bias.

However, the paper also notes significant challenges. There is a "Reliability Gap" where proxy models might rely on statistical artifacts or "adversarial noise" that the original human bearer would never consider. Furthermore, while heatmaps are popular, they often indicate how to change an input to alter a classification rather than explaining the current classification itself.

Ethical Implications and Future Directions

The ability to decode individual interpretations carries significant ethical weight. The authors warn of "dual-use" risks: while these tools can uncover hidden biases and increase AI transparency, they could also be used for hyper-targeted manipulation of human behavior.

Looking forward, the field must move beyond simple correlation toward Causal Inference to prove that specific features actually cause an interpretation. Additionally, there is a pressing need for standardized, automatic evaluation metrics for "explanations" to move beyond subjective qualitative assessment.

Summary updated

Abstract

It is said that beauty is in the eye of the beholder. But how exactly can we characterize such discrepancies in interpretation? For example, are there any specific features of an image that makes person A regard an image as beautiful while person B finds the same image displeasing? Such questions ultimately aim at explaining our individual ways of interpretation, an intention that has been of fundamental importance to the social sciences from the beginning. More recently, advances in computer science brought up two related questions: First, can computational tools be adopted for analyzing ways of interpretation? Second, what if the "beholder" is a computer model, i.e., how can we explain a computer model's point of view? Numerous efforts have been made regarding both of these points, while many existing approaches focus on particular aspects and are still rather separate. With this paper, in order to connect these approaches we introduce a theoretical framework for analyzing interpretation, which is applicable to interpretation of both human beings and computer models. We give an overview of relevant computational approaches from various fields, and discuss the most common and promising application areas. The focus of this paper lies on interpretation of text and image data, while many of the presented approaches are applicable to other types of data as well.
BibTeX
@misc{blandfort2018overview,
  author = {Philipp Blandfort and Jörn Hees and Desmond Upton Patton},
  title = {{An Overview of Computational Approaches for Analyzing Interpretation}},
  year = {2018},
  eprint = {1811.04028},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1811.04028},
}
Figure