A connectionist network with recurrent connections among units was designed to simulate human reader's performances in naming Japanese Kanji words. The network was trained to map the orthography of two-character Kanji words onto their pronunciations or phonology. After learning with the training corpus including approximately 4,000 two-character Kanji words, the network showed, in its naming latency, frequency and consistency effects and an interaction between these effects, largely comparable to these effects in naming latency of Japanese skilled readers. On the other hand, when naming nonwords, the network's accuracy was substantially worse than that of skilled readers. The results of the present study indicate that such a kind of network, although originally developed to simulate human readers' performance in alphabetic writing systems, can apply to Kanji as well, with caution that further elaboration is indispensable to cope with poor performance for nonwords.
This paper proposes a model that can learn the meanings of instructions (for example, “good” and “bad”.). This model assumes that an advisee learns the meanings of instructions in parallel with learning the evaluation of its action experience. The reinforcement learning algorithm is adopted for the action learning. We conducted experiments with a robot simulator. The result of the experiments suggests that our model can learn not only evaluation-instructions but also two types of instruction (evaluation-instructions and direction-instructions) simultaneously. This model can be thought as a basic model of an intelligent agent that can learn the meanings of instructions.
The main objective of this paper is, on one hand, to introduce a viewpoint of dynamical system theory into the study of social surveys that is now confronting many problems in terms of sampling data collection, and on the other hand, to exemplify the utility of some principles and the framework of quantum theory in the cognitive sciences. First, after the theory on a psychometric model called “wave-scaling” based on social quantum theory (Yoshino, Behaviormetrika, 1998, 25, 2, pp. 111-132) is summarized, the theory is modified and applied to the analyses of practical social survey data: longitudinal survey data on Japanese national character and on voting rate in the election of senator in Japan. Second, although the original model is invented for the case where the number of categories under consideration is 2, it is now extended as a model named HiBiD (hierarchical binary decomposition) to cover more general types of social survey data where the number of categories can be more than 2. This model is tested in the applications of the analysis of recent nationwide elections data. Finally, some comments and suggestions are provided for the future development of this model.
We obtained fifty-two recall data when the titles of twenty well-known Japanese songs that most Japanese people learn in school were given. Song memory is based on the associative chaining of units of melody and lyrics. The parts found to be well memorized were the beginning of the songs, parts that share words with the titles, and the end of the songs. We found various kinds of errors in the recall of song lyrics, which we classified into four groups: song confusion, lyrics confusion, word construction, and misused characters. The primary constraints on word construction are the number of syllables, vowels, parts of speech, and syntactic contexts. Word construction is based on rationalization that transfers things reasonable and natural to information receivers. We found some interesting errors that change the image of the original song lyrics. Based on these findings, we present a model for recalling songs from a title and a model of error recall in Japanese songs.
We investigated the effect of category learning in visual pattern recognition by psychological experiment and neural network model simulation. In a psychological experiment, we used tachistoscopic presentation task, and we found that pattern discriminability was enhanced by category knowledge. We constructed a neural network model with three layers and reciprocal connections. We used Wake-Sleep algorithm for network learning and the network made its internal representation by the interaction of bottom-up and top-down processes. The network model can simulate the profit of having category knowledge observed in psychological experiment. Furthermore we considered the computational explanation of the profit of having category knowledge from the viewpoint of MDL (Minimum Description Length). Category knowledge helps the network to construct efficient (shorter description length) representation of patterns. In conclusion, category knowledge has a functional profit not only in visual object identification but also in efficient processing of pattern recognition.
When they use daily electronic appliances to achieve their goals, users need to decompose their goals to a set of subtasks in a specific way. A previous study revealed that users who have extreme difficulties in using such appliances fail to decompose the task, or that their decompositions are different from the one that designers assume. Based on this view, we propose a new method to help users understand “task decomposition” by providing all operating functions in a tree structure. We conducted three experiments. In the first and second experiments, we built support panels corresponding to the operating panels of two different copiers, which visualize all operating functions of respective copiers to facilitate users' understanding task decomposition. We compared the copying performance of subjects who used the support panels with that of those who did not. Both the support panels were found to be useful for technologically inept users. We then built an interface by integrating the operating panel into the support panel, both of which were provided in the first experiment: We conducted an experiment similar to both of the previous ones and evaluated its effectiveness. The result showed that this interface facilitated users' task decomposition. We conclude that the supporting method we proposed is effective for technophobic users' understanding task decomposition.
Using the “matchstick” algebra problem, in which an incorrect arithmetic equation, given in Roman numerals, must be made true by moving a single “stick” of a Roman numeral (e.g., IV=III+III→VI=III+III), two experiments were conducted to investigate the role of constraint relaxation and different problem representations in insight problem solving. Participants tried to solve matchstick problems, presented in either Roman or Arabic numerals, three times. The first experiment showed that participants were significantly less successful when the problems were presented in Arabic numerals than when Roman numerals were used. The second experiment showed that participants who had been unable to solve the problem in the first and second trials were significantly more likely to solve the problem at the third trial when they switched from Arabic to Roman numerals between the second and third trials. Two theories of constraint relaxation and the role of different problem representations in insight problem solving are also discussed.
In this study, we proposed general goal-setting strategies in a problem of collaboratively achieving a common goal by two problem solvers. As a framework for analysis, we utilized the recursive problem space of the Tower of Hanoi puzzle. We described subjects' goal setting processes by the protocol analysis, and discussed its transition patterns using two indexes for the estimation on the recursive problem space: abstraction level of each goal and distance between two goal states. We identified two kinds of goal setting strategies. One was called the HG strategy (Higher-abstraction-oriented Gradual-goal-setting strategy) in which subjects first set a goal whose abstraction level was the highest then moved to more specific goal settings as the problem was being solved. The second was called the landmark strategy in which subjects used a prominent goal state as a landmark. We confirmed, through our cognitive psychological experiment and analysis, how those two strategies were actually utilized in the subjects' collaborative problem solving processes.