For a robot working in an open environment, a task-oriented language ability will not be sufficient. To adapt to the environment, such a robot will have to learn language dynamically. In this paper we propose a system which acquires noun concepts (labels for images) based on an infant vocabulary acquisition model. In order to improve the performance, we have introduced label acquisition rules into this system. The evaluation experiment shows that the rules are effective in extracting noun concepts. We also evaluate user impressions of the system using the rating scale method. The experimental results show that the system leaves a good impression on people.
This paper deals with Japanese riddles that are created using not only different meanings of two homonymous verbs but also common knowledge of causal relation, and aims at the construction of a creation system of such Japanese riddles. The proposed system consists of a creation part and two kinds of databases. The one database has information on verbs, their use frequency and the case relation between verbs and noun phrases. The other has information on effects caused by the action expressed by a verb and the degree of causal relation between actions and effects. From these sorts of databases the creation part extracts homonymous verbs and noun phrases having case relations to extracted verbs using the following indices; the verb's use frequency and the degree of causal relations. And then the creation part chooses one homonym occurring to human easily among chosen homonyms, and creates as interesting Japanese riddles as the ones that human does. The creation part also shows their answers with explanations. This paper also shows subject experiments in order to confirm whether interesting riddles are created by the presented system.
The purpose of this paper is to develop an honorific expression e-learning system with dialogue so as to learn not only the wording of honorific but also the situation when we should use it. Our system consists of three phases: (1) preprocessing; (2) honorific processing; (3) dialogue processing, and three kinds of data: honorific judgment dictionary, honorific explanation dictionary and situation data. In the preprocessing phase, an input text from a user is morphologically analyzed to obtain the nouns and verbs. The honorific type that should be used in current stage is inferred from the feature of its context. In the honorific processing phase, our system extracts honorific errors in the input text and corrects based on the inferred honorific type. As the error correction, particular translation is first performed using the honorific judgment dictionary, and after that general translation is performed. The explanation for the wrong honorific is simultaneously retrieved. In the dialogue processing part, an output text corresponding to the input text is adequately provided from the situation data. Repeating the above procedures, i.e., user input, system processing, and system output, will ensure the dialogue for learning the honorific expression effectively. We evaluated the correction accuracy and learning effect of an implemented system.
To manage and utilize huge amount of multimedia contents, various kinds of application software have been provided. However, they manage the contents based upon the property information of a file, such as file type, size, or date, therefore, it is hard to say that they allow us to easily access to the target contents with the only information of a file. In this context, in our study, we propose a method to firstly structurize the contents by using the `time' and `place' information of an event which tends to be naturally remembered in our everyday lives (In this paper, we call this information “Everyday life Ontology”), and then to operate the contents by natural language. In addition to the basic property information of a file, by using the constraints about time and location in order to manage the contents, we will be able to manage the contents considering the situation in which the contents were obtained. And, in order to operate application software by natural language, we have developed an ontology of the operation function of application software, which makes it possible to flexibly connect the user's natural language input with the commands of application software. Furthermore, we can access to the structured contents by using the time and place information used for its structure, this makes us possible to access other contents through the hierarchy and sequence of the structure from the contents currently being focused. We conducted an experiment to test the abilities of users to deal with multimedia contents and these constraints using our natural language interface, and then compared them to manual operations, evaluating them from the viewpoints of necessary time to achieve the given task and of the correctness of result. By this, the usefulness of the system and our proposed method was confirmed.
A natural language dialogue system should correctly interpret input sentences from the users and extract information essential for proper response. Since a variety of surface dependency structures may carry the same meaning, a natural language dialog system should be able to interpret such variety of expressions. Traditional semantic representations, however, take over the variety of surface dependency structures without interpreting them. Consequently, we cannot compare semantic expressions with different dependency structure. In addition, it is practically impossible to prepare interpreting rules particular to each possible input. In our previous studies, we proposed a framework for semantic representations which circumvents the problem. The framework enables us to compare semantic representations with different dependency structures without employing structure-specific rules. In this paper, we describe how to interpret a clause based on its semantic representation and show that the interpreting processes are applicable to the interpretation of multiple clauses and/or multiple sentences. We have constructed a dialogue system based on this framework and evaluated the system. As a result, we have confirmed that the system can make proper response and that the framework gives the system capacity to make proper response, as desired.
Internet users write blogs related to their personal experience, daily news, and so on. We can obtain blogs about personal experience using search engines on the Web. However, the search engines also output blogs about other topics unrelated to personal experience. Therefore, it is necessary for us to read all blogs to obtain those about personal experiences. It takes too much time. This paper proposes a support system for obtaining blogs about personal experiences efficiently. The system extracts three keywords that denote place, object, and action from a blog. The three keywords describe an event that leads a person to write a blog about personal experience. The system expresses the event with three pictures related to the extracted keywords. The pictures help users to judge whether personal experience is written about in the blog. We experimented with the system, and verified that it supports users in obtaining personal experiences efficiently.
In this article, it is used virus evolution type Genetic Algorithm searching for a solution by co-evolution of two populations. One of the population is host population to have a solution of the issue of application, virus population to have the partial solution one more. And we confirmed that evolution processes were different by a difference of the infection technique that was genetic operations of a virus. In addition, we report that the result that is better than evolution by independent infection technique was provided by putting different infection technique together.
In general, the structure of neural network, such as the layered type or the mutually connected type, etc., is chosen as to a given problem. And connection weights and thresholds of neural units are usually determined by the learning method corresponding to the structure. Several given problems, however, cannot be always solved by the neural network of the structure decided beforehand. Therefore, Flexibly Connected Neural Network (FCN) was previously proposed as a method of constructing arbitrary neural networks with optimized structures and parameters to solve unknown problems. FCN was applied to action control of an autonomous agent and showed experimentally that it is effective for perceptual aliasing problems. The number of hidden units, however, was determined experimentally in the FCN. In this paper, we propose a method based on FCN, which can determine automatically the number of hidden units without trial and error. In order to verify the effectiveness, we applied the proposed method to Tartarus Problem that requires action control of an autonomous agent in unknown maps and we analyzed the obtained action rules.
Interactive Evolutionary Computation (IEC) is one of the effective methods for optimization problems which are difficult to formulate the evaluation function such as human sensitivity. This method, however, gives a big burden to the user because he/she himself/herself has to evaluate the candidates of solutions a lot of times. This paper employs a fitness inference method to reduce the burden for evaluation. When fitness inference method is applied to IEC, it can be a problem that the evaluation criterion for user has changed with the passage of time and/or impression given by the effects of other candidates. Because the fitness inference method infers the fitness values of candidates using the information of the actually evaluated solutions and their fitness values in previous generations. This paper proposes the fitness inference method with varying the number of actual evaluation for candidates based on the accuracy of inference for effective update of database. The proposed method can reduce the burden of evaluation following the user's evaluation criterion. This paper applies the proposed method to hearing aid adjustment support system using IEC and investigates the effectiveness of this method.
In this paper, we compare the pattern classification methods to discriminate the inferior shijimi clams with aim at development of the selecting device for shijimi clams. In our framework, the feature vectors are extracted based on frequency analysis of the acoustic signal that occurs in hitting shijimi clams themselves, and then shijimi clams are classified by the pattern classification methods such as decision tree learning, multi-layer perceptron, k-nearest-neighbors classification method and support vector machine. The experimental results indicate that multi-layer perceptron shows the best performance among the four pattern classification methods. Also we confirm all four pattern classification methods show sufficiently high accuracy. Therefore, we can show the effectiveness of our framework.