Gary McGraw, John Rehling & Robert Goldstone. (1994) Letter Perception: Toward a conceptual approach. In the Proceedings of the Sixteenth Annual Conference of the Cognitive Science Society, pages 613-618, Atlanta, GA, August 1994.

Author of the summary: Patrawadee Prasangsit, 1999, pp@cc.gatech.edu

Cite this paper for:

• Conceptual-level representations of letter parts, namely roles, are important in human's process of letter recognition.
• Analogical reasoning occurs at roles and spatial relations among them.

Summary:

The concept, for instance, lowercase-'a' can be thought of as a component composed of two smaller ideas: (1) the idea of a curved umbrella-handle-like bar on the right, and (2) the idea of a small c-like curve on the left.  These two conceptual components, which are called roles, are not explicit shapes per se but are ideas about what the acceptable bounds for letter-part shapes are, how far such shapes can be stretched before they lose their interpretation, and how they interact with other roles to form a complete object.  Roles and relationships between roles make up the internal structure of a letter category.  Category membership at the whole-letter level is determined by category membership at the lower level of roles.

Slippage involves allowing certain descriptions in a mental representation to "slip, " or be replaced by related descriptions according to contextual pressures brought to bear by the situation.

Each letter is formed by a set of short line segments, called quanta, on a fixed grid of dimension 3x7.  See figure 3.

Subjects are presented with a number of lowercase letters, one at a time, both in normal and styled shapes (called NORMALS and FONTS, respectively), and asked to identify them.  Based on the results, the authors make some predictions:

Prediction 1

• Average reaction time increases while accuracy decreases, which can be explain as follows.  Difficulty is directly determined by how easy a given letterform can be broken into parts that fill roles.  Complicated letterforms are ambiguous (thus accuracy decreases) as well as take longer to parse into roles (thus reaction time increases).
• NORMALS letterforms are far more easily recognized than FONTS.  Figure 4 and 5 show results in terms of reaction time and accuracy.
• The number of quanta does not alone determine its difficulty.

• "Blurred prototypes" of letters are created by averaging the quanta lists of NORMALS and FONTS letters in a given category together.
• A hierarchical cluster (figure 7) is created using the Euclidean distances between the blurred prototypes.
• When compare this cluster with the  errors made by human, they do not agree.  For instance, 'e' and 's' are very closely related in the hierarchical cluster (hence probably have higher chance of being mistaken as one another) but errors made by human mistaking one as another is only 0.20%.
• This contrary suggests that human's process of recognition has little to to do with low-level information such as quanta.  Literal physical properties are often less important in similarity judgments than are higher-level relations.

• Some pair of letters look similar but their higher-order concepts are too different for people to miscategorize one as another.  On the other hand, some pair of letters have similar concepts that people tend to slip them them since human perception is flexible.  In fact, all of the top human errors in the experiments can be explained by conceptual slippage (such as, unclear role fillers).

• A plot of the number of categories in which each token was found versus its average reaction time.  They strongly correlate.  As the number of categories increase, so does the reaction time.
• This result can be used to classify the tokens into two sets: those with ambiguous role fillers (i.e., few possible categories) and those with strange role fillers (i.e., large number of possible categories).  The latter takes longer reaction time.  When a token is ambiguous, its parts can fill two or more sets in a fairly reasonable fashion and hence takes subjects some time to decide.  As more choices are considered, it takes longer.  When a token is strange, its parts do not really fit any sets of roles well, so subjects make a guess after an extended period of time.

Summary author's notes:

• none