A Hard Look at Soft Concepts

Dafna ShahafIJCAI 2016

Introduction

As artificial intelligence masters "hard" logic-based tasks such as Chess, Go, and Jeopardy, the frontier of human uniqueness has shifted toward "soft" concepts: creativity, humor, and the synthesis of complex, non-linear narratives. These concepts lack a clear "objective function," making it difficult to define mathematically what makes a story coherent or a joke funny. This research bridges the gap by providing a framework to formalize the intuitive through two primary methodologies: an axiomatic approach and a data-driven approach.

Methodology: Information Cartography

The first approach, Information Cartography, uses an axiomatic framework to help users navigate complex news events. The author developed "metro maps" that visualize intersecting narrative threads rather than linear lists of articles. This system formalizes a "good map" using three mathematical axioms:

  • Coherence: Ensuring a narrative thread makes sense globally, modeled as a linear programming problem to avoid "associative drift."
  • Coverage: Using submodular functions to ensure the map captures diverse and important aspects of the story.
  • Connectivity: Revealing how different storylines interact through structural intersections.
A metro map of the 2014 Crimean crisis, showing intersecting narrative lines.

Methodology: Humor Recognition

In contrast to the top-down axiomatic approach, the author employed a data-driven methodology to tackle humor. By analyzing thousands of entries from The New Yorker Caption Contest, the system identified linguistic features—such as sentiment, readability, and part-of-speech frequency—that correlate with "funniness."

A sample cartoon from The New Yorker used to evaluate humor recognition algorithms.

Empirical Validation and Results

The research utilized diverse datasets, including news corpora for Crimean crisis maps and scientific corpora for literature surveys. In a study comparing metro maps to Google Scholar, map users achieved significantly higher precision (84.5% vs. 74.2%) and identified more seminal papers on average (1.62 vs. 1.2).

In the domain of humor, the classifier correctly identified the funnier of two captions 64% of the time. Furthermore, the algorithm's top-10 rankings consistently included captions selected by human judges, suggesting its utility in automating the initial screening process for editors.

Key Insights and Discussion

A critical insight from narrative mapping is that coherence is a global property; local similarity between adjacent sentences can lead to drift, whereas a "line" must maintain a consistent theme. In humor, the data revealed that simplicity often wins: funnier captions tend to use fewer proper nouns and third-person references, relying instead on relatable language.

While these models provide powerful tools for navigating information and identifying humor, they remain limited. The humor model relies on linguistic "traces" rather than a deep semantic understanding of incongruity. Future work will focus on generalizing these axioms to other domains, such as legal or fictional texts, and enhancing human-AI collaboration.

Summary updated

BibTeX
@inproceedings{shahaf2016hard,
  author = {Dafna Shahaf},
  title = {{A Hard Look at Soft Concepts}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
}
Figure