Computational Methods in the Analysis of Verbal Behaviour

R. M. Frumkina, P. F. Andrukovich, A. Yu. TerekhinaCOLING 1973

TL;DR

This paper introduces multidimensional scaling techniques to model verbal perception, moving beyond traditional text analysis. By analyzing how people perceive similarities between Russian letters, the authors demonstrate a method to uncover underlying perceptual dimensions and individual differences in how these dimensions are weighted, offering a more nuanced view of verbal behavior.

What Problem Does This Paper Address?

While computational linguistics has largely focused on automatic text analysis and generation, the broader application of computational methods to understand the perception of language and verbal behavior has been slower to gain traction. Psychologists have been more proactive, but linguists have lagged in adopting computational tools for developing and evaluating mathematical models of how humans process and understand speech. This paper aims to bridge this gap by demonstrating how multidimensional scaling can be used to create differential representations of verbal perceptual structures, essentially trying to open the "black box" of speech perception.

Methods

The core methodology employed is multidimensional scaling (MDS). The authors utilize a modification of "individual multidimensional scaling" methods, which allows for the analysis of individual differences in how subjects perceive stimuli. The general approach involves:

  1. Determining Dimensionality: Identifying the number of underlying dimensions that best explain the perceived similarities between stimuli.
  2. Stimulus Positioning: Mapping each stimulus (in this case, Russian letters) as a point in a multi-dimensional perceptual space, where the distance between points reflects perceived similarity.
  3. Subject Positioning: Representing each subject's perceptual space by assigning weights to the identified dimensions, reflecting their individual salience.

Specifically, they first use the principal components method to identify the most discriminative dimensions. Then, they apply individual scaling to uncover how each subject differentially weights these common dimensions. A key aspect is the development of a criterion to minimize discrepancies between observed similarity judgments and predicted distances in the perceptual space, with a specific formula provided for this purpose.

Datasets

The study uses data from an experiment involving 32 block letters of the Russian alphabet. These letters were presented pairwise to subjects, who were asked to judge whether each pair was "alike" or "not alike." Each comparison was made by at least 50 subjects. The resulting data for each subject is a similarity matrix.

Key Insights and Ideas

  • Beyond Text: Perceptual Space of Language: The paper highlights that understanding verbal behavior requires looking beyond the written text itself to the underlying perceptual and cognitive structures that shape how we interpret language.
  • Uncovering Hidden Dimensions: MDS can reveal latent dimensions of perception that are not immediately obvious. For Russian letters, the authors identified dimensions related to "straight elements vs. non-straight elements" and "acute-angular elements vs. non-acute-angular elements."
  • Individual Differences Matter: A crucial finding is that while common perceptual dimensions exist, individuals weight these dimensions differently. This "individual scaling" approach allows for the mapping of subjects into a perceptual space, revealing how their subjective interpretations vary.
  • "Alike" vs. "Not Alike" as a Window: The simple binary judgment of "alike" or "not alike" provides rich data for inferring perceptual structures, demonstrating that even basic experimental designs can yield significant insights.
  • Potential for Broader Applications: The authors suggest this method has significant potential for analyzing other verbal phenomena, such as confusion data in language acquisition, to understand internalized linguistic knowledge.

Key Results

The analysis of the Russian block letters revealed two primary psycholinguistically interpretable dimensions:

  1. Dimension 1: Opposition between letters with straight elements and those without.
  2. Dimension 2: Opposition between letters with acute-angular elements and those without.

The individual scaling analysis showed that subjects could be grouped based on how they weighted these dimensions. Some subjects weighted the first dimension more heavily, while others prioritized the second. A significant portion of subjects showed little difference in their weighting, suggesting a more uniform perception. The paper presents figures (Fig. 3, 4, 5, 6) illustrating these stimulus and subject configurations, and the contrasting perceptual spaces of individual subjects.

Limitations and Future Work

  • Illustrative Example: The authors acknowledge that the classification of letters might seem trivial, and the study serves primarily as an illustrative example of the method's application.
  • Further Research: The paper strongly advocates for the broader application of this method to verbal behavior research, particularly in analyzing confusions data for children at different stages of language acquisition. This could shed light on the distinction between linguistic "competence" and "performance."
  • Generalization: The authors express hope that this approach can be generalized to discover socially determined cognitive superstructures underlying individual verbal behavior.

Summary updated

BibTeX
@inproceedings{frumkina1973computational,
  author = {R. M. Frumkina and P. F. Andrukovich and A. Yu. Terekhina},
  title = {{Computational Methods in the Analysis of Verbal Behaviour}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1973},
}
Figure