To the study of character pattern recognition by the automatic information processing system, psychologists can contribute in two ways at least. One is to analyse the human recognition process experimentally and to clarify its intrinsic nature, and the other is to offer some useful data about the ways to compose the optimal character patterns for the mechanical recognizers. Several years ago we showed one possibility of reasonably composing 46 Japanese Kata-Kana letters with 9×5 cells, and also ascertained that these quantized characters had tolerable readability as compared with typed letters. The purpose of this paper is to describe quantitatively the features of each cell which is the minimum component of the quantized characters, and to clarify the process by which human Ss recognize these characters when only part of the cell information was given. The Kata-Kana letters tentatively selected and used in this study were only 10 of all. First, two corresponding cells aij(aij=1 or 0, i=1......9, j=1......5) of each pair of letters are compared, and if two cells are equivalent, that is, if both are either 1 or 0, Yij is set to 0, while Yij=1 if different. Since there are 45 pairs of letters, we get 45 sets of value of Yij. Next, for each letter the sum Zij is obtained by summing 9 Yij's in which the letter in question is paired with the other 9 letters (Table 2). Zij's thus obtained show the degree to which the cells of that letter are structurally characterized with respect to the other letters. We can also sum up 45 Yij values of all pairs to obtain Wij (Table 3). Wij reflects the extent to which 10 letters are efficiently categorized into two groups according to the cell information. The cell with the maximum Wij is the most important one in the sense that it would help the mechanical recognition system with maximum efficiency. Excluding all pairs that gave Yij=1 in the most important cell and summing Yij's for the remaining pairs, we obtain second set of values of Wij, and determine the second important cell. Repeating these procedures until all pairs are exhausted, progressively less important cells are determined. In experiment 10 Ss were asked to guess what the character was when partial information was given by one or more cells; in Exp. I only one cell with varing value of Zij or Wij was presented, and in Exp. II cells were presented one after another to accumulate information in the order of either computed importance (F-condition), or randomness (F′-condition). Based on the analysis of response distributions, the following were concluded. 1. When a cell with larger Zij, was presented as a cue, the degree of correct identification was higher (Fig. 3), and the larger the Wij value, the more the cell information character recognition. But the processing of information was not perfect and there seemed to be considerable loss of information. 2. The loss of information was less in F′ which contained much useless information than F which agreed with the theoretically efficient system, and increased as the possibility of correct identification became larger, especially in F. 3. It was presumed that more than 25 cells or 16 bits of cell information have to be presented in order that Ss can identify 10 letters perfectly. 4. It was also suggested that the redundancy of Kata-Kana letters quantized in 9×5 cells was slightly more than 0.7. 5. The efficient information that would help the mechanical system to recognize letters is not always efficient and important for the human
In the preceding study (I), a model of information transmission had been applied in connection with two test models, the bivariate-normal model and the binomial error model, but here the rate of information transmission in hypothetical tests consisting of items of various characteristics were estimated. The logistic test model was used for computational conveniences. Among others it is interesting that items with lower discriminability, in some cases, yielded higher rates of information transmission than those with higher discriminability. This phenomenon is closely analogous to “attenuation paradox”. There were no monotonic relationships between test reliability and information transmission rate as shown in the bivariatenormal model.
The purpose of the present study is to classify types of learning process in the mirror drawing task by applying the factor analytic technique. Using 35 undergraduate students as Ss, the following two measures were observed; i.e., (1) accuracy, and (2) speed on the mirror drawing task. The whole learning process consisting of 20 trials was divided into nine stages and mean scores of these measures mentinned above were calculated for each stage. Next, the Pearson's product moment correlations (r) were computed. between these 13 mean scores (namely, two means for each of the nine stages). From the resulting correlation matrix five centroid factors were extracted, and then, the axes of them were rotated orthogonally by the subjective method. According to this rotated matrix, the followings were ascertained. (a) In both accuracy and speed data, either one or two different factors explained most of the total variance. (b) As learning proceeded, the ratio of contribution of one major factor increased gradually, and the factor structure became simpler. As the next step of analysis, the regression coefficients were calculated for the two main factors by Ledermann's method, and multiplying them by accuracy and speed scores, the factor-scores for these two factors were obtained for each subject. In accordance with the pattern of these factor-scores, all subjects were eventually classified into nine types. In order to examine the effectiveness of the classifications, all subjects were divided further into three groups (i.e., +, _??_, and - groups) according to the factor-score distributions, and differences in two measures were compared respectively. Analysis of variance of the data revealed that the differences among these groups were highly significant (p<.001), in both measures. It has been confirmed that a factor analytic technique, as a tool of classifying these types, was effectively applicable to such learning process data as mirror drawing.