Hypothesis Testing in a Computational Theory of Visual Word Recognition

Jonathan J. HullAAAI 1987

In his 1987 paper presented at the AAAI proceedings, Jonathan J. Hull proposed a significant shift in the computational approach to reading. Moving away from the traditional, often brittle method of segmenting words into individual characters, Hull introduced a holistic, human-inspired computational theory of visual word recognition. This model prioritizes "gross visual shapes" to generate hypotheses, which are then refined through selective feature testing.

The Challenge of Traditional OCR

During the late 1980s, the primary hurdle for Artificial Intelligence in text recognition was the reliance on character-by-character analysis. These systems attempted to isolate each letter, identify it, and then reconstruct the word. However, this approach is highly susceptible to noise and formatting issues. When characters touch—a common occurrence in handwriting or tight kerning—segmentation fails, and the recognition process breaks down.

Hull observed that human readers do not process text this way. Instead, we utilize "saccades" (rapid eye movements) and peripheral vision to capture a broad visual description of a word, using context and overall shape to fill in gaps. Hull’s theory seeks to mimic this multi-stage biological process to achieve greater algorithmic robustness.

A Three-Stage Computational Theory

Hull’s proposed system operates through a structured hierarchy designed for computational economy:

  1. Hypothesis Generation: The system extracts a "gross visual description" of a word—identifying sequences of ascenders, descenders, and vertical bars. This description acts as an index into a dictionary to retrieve a "neighborhood" of visually similar candidates (e.g., "be," "he," and "has").
  2. Hypothesis Testing: Once a neighborhood is identified, the system employs a decision tree to perform selective feature analysis. Rather than analyzing every pixel, it executes only the minimum number of tests (MT) required to distinguish between the specific candidates in that neighborhood.
  3. High-Level Processing: A final stage (proposed for future development) uses syntax and semantics to resolve any remaining visual ambiguities.
A decision tree used to distinguish between the visually similar words 'be', 'has', and 'he'.

Experimental Methodology and Results

The study utilized the Brown Corpus, a massive collection of modern American English, to test the algorithm across various genres (e.g., Press, Religion, Science). Hull also generated 12,600 synthetic word images across 10 different fonts, specifically including "touching" characters to test the system's resilience against segmentation errors.

Key Findings:

  • High Accuracy: The system achieved 98% accuracy on training fonts and 95% on novel fonts when characters were separated.
  • Robustness to Noise: When characters were forced to touch, accuracy remained high at 92%, a scenario where traditional OCR typically fails.
  • Computational Efficiency: The methodology proved highly scalable. For a vocabulary of 25,000 words, only 425 unique tests were required to distinguish them, demonstrating that the number of tests scales linearly and efficiently with dictionary size.
Scalability of the hypothesis testing method: the number of tests required grows slowly relative to dictionary size.

Conclusion and Future Directions

The primary limitation of this visual-first approach is "visual ambiguity." Approximately 3% to 7% of text contains "visual twins"—words that are indistinguishable by shape alone in certain fonts. Hull concludes that high-level linguistic knowledge—incorporating grammar and meaning—is the essential third stage required to bridge this gap, allowing a reading machine to function with the same contextual fluency as a human reader.

Summary updated

BibTeX
@inproceedings{hull1987hypothesis,
  author = {Jonathan J. Hull},
  title = {{Hypothesis Testing in a Computational Theory of Visual Word Recognition}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1987},
}
Figure