Abstract:

Generative theory holds that the recognition of visual graphics is the inverse reasoning of its generation process. Chinese characters are hieroglyphs formed by interlacing strokes according to orthographic rules. Chinese character recognition can be regarded as the reverse reasoning of the generation process of Chinese characters. Based on the typical generative model -- Bayesian program learning model, the recognition of Chinese characters starts from recognizing the strokes. Firstly, the nodes are extracted based on the intersection of lines, and then all the stroke combination modes that can generate the node are enumerated to obtain the generation mode of Chinese characters. According to the above prediction, the number of nodes and node complexity are important factors in the process of Chinese character recognition. This study investigated the role of nodes in Chinese character recognition through three experiments.

If the nodes provide guidance information for stroke segmentation, the more nodes, the better the performance of Chinese character recognition. In Experiment 1, we tested whether characters with more nodes have recognition advantages by adopting a 2 × 2 within-subjects design and using 76 single characters as the materials (Figure 1). Characters were chosen from two groups (high node-count and low node-count) of true characters, and two groups (high node-count and low node-count) of fake characters. The characters were briefly presented (10 ms, 20 ms, 30 ms, 40 ms, 50 ms, 60 ms) and appeared once at each presentation time. The presentation order of stimuli was completely random. Each participant completed a total of 456 trials. Twenty-six participants joined in the experiment. After observing each character, the participants reported whether it was a true character or a fake one.

If high complex nodes in a larger stroke space provide more information, covering high complex nodes will cause greater interference to character recognition. In Experiment 2, we tested whether characters covered the high complex nodes are harder to recognize by adopting a 2 × 2 × 4 within-subjects design and using 160 compound characters covering a node as the materials (Figure 2). Characters were chosen from four groups (covering the first node with high or low complexity and the fifth node with high or low complexity) of true characters, and four groups of fake characters with the same conditions. The process was the same as that of Experiment 1, except the presentation time (60 ms, 70 ms, 80 ms, 90 ms). Twenty-nine participants joined in the experiment. Each participant completed 640 trials. Experiment 3 adopted a task similar to Experiment 2, and added two variables: component type and node generation method. The presentation time was 60 ms. Characters were chosen from eight groups of true characters, and eight groups of fake characters with the same conditions (Figure 3). Each stimulus is presented once. Twenty-six participants joined in the experiment. The accuracy and reaction time (RT) of true characters were analyzed in all experiments.

The results showed that the participants had a better recognition performance for the characters with more nodes (node number effect), and covering the high complex nodes significantly damaged their performance (node complexity effect). In Experiment 1, the overall mean of accuracy was 63%, and the standard deviation was 29%. The accuracy of recognizing characters with more nodes was higher (Figure 4). The repeated-measures ANOVA of accuracy found that the main effect of the number of nodes was significant, F₁ (1, 25) = 9.65, p = 0.005, η_p² = 0.28, 95% CI = [−5%, −1%]; F₂ (1, 18) = 4.56, p = 0.047, η_p² = 0.20, 95% CI = [−5%, −1%]. The interaction between the number of nodes and presentation time was significant, F₁ (5, 125) = 1.998, p = 0.083, η_p² = 0.07; F₂ (5, 90) = 1.12, p = 0.355. When the stimulus presentation time were 40 ms and 50 ms, node number effect was more pronounced.

In Experiment 2, the overall mean of accuracy was 79%, and the standard deviation was 21%. The accuracy of recognizing characters covering the high complex nodes was lower (Figure 5). The repeated-measures ANOVA of accuracy found that the main effect of the complexity of nodes was significant, F₁ (1, 28) = 6.93, p = 0.014, η_p² = 0.20, 95% CI = [−4%, −1%]; F₂ (1, 19) = 0.73, p = 0.404. The interaction between node complexity and node order is significant, F₁ (1, 28) = 11.56, p = 0.002, η_p² = 0.29; F₂ (1, 19) = 2.49, p = 0.131. Node complexity effect was more pronounced when covering the fifth nodes.

In Experiment 3, the overall mean of accuracy was 72%, and the standard deviation was 18%. we also found that the main effect of the complexity of nodes was significant (Figure 6). The repeated-measures ANOVA of accuracy found that the main effect of the complexity of nodes was significant, F₁ (1, 200) = 8.32, p = 0.004, η_p² = 0.04; F₂ (1, 112) = 5.69, p = 0.019, η_p² = 0.05.The interaction between node complexity and node generation method is significant, F₁ (1, 200) = 3.87, p = 0.050, η_p² = 0.02; F₂ (1, 112) = 1.96, p = 0.165, η_p² = 0.02. Node complexity effect was more pronounced when covering the nodes generated by strokes connection, t (102) = −3.67, p < 0.001, Cohen’s d = 0.73, 95% CI = [−19%, −6%].

These findings support the nodes provide bottom-up stroke separation guidance information. Stroke separation was performed in parallel, for the more nodes a character has, the more information provided for the stroke segmentation, and therefore the character would be easier to recognize. And stroke separating began with the extraction and analysis of nodes, for the more complex nodes are, the greater impact on recognize. This study deepens the understanding of the early visual process of Chinese character recognition and supports Chinese character recognition is a generative reverse reasoning process, which could contribute to develop a complete cognitive model of Chinese character recognition.

Key words: handwritten Chinese character recognition, node, stroke, generative model