Michio Sugeno

A systemic-functional approach to Japanese text understanding

We have implemented a Japanese text processing system, combining the existing parser and dictionary with the linguistic resources that we developed based on systemic functional linguistics. In this paper, we explain the text understanding algorithm of our system that utilizes the various linguistic resources in the Semiotic Base suggested by Halliday. First, we describe the structure of the SB and the linguistic resources stored in it. Then, we depict the text understanding algorithm using the SB. The process starts with morphological and dependency analyses by the non-SFL-based existing parser, followed by looking up the dictionary to enrich the input for SFL-based analysis. After mapping the pre-processing results onto systemic features, the path identification of selected features and unification based on O’Donnell are conducted with reference to the linguistic resource represented in the system networks. Consequently, we obtain graphological, lexicogrammatical, semantic and conceptual annotations of a given text.

A model of rhetorical structure analysis of Japanese instructional texts and its application to a smart help system

The purpose of this paper is to propose an algorithm of rhetorical structure analysis of Japanese texts with special reference to wordprocessor help texts, which is to be realized in an interactive help system. The model of the rhetorical structure analysis presented in this paper is applicable to 25 texts (118 sentences) extracted from in the help texts accompanying Microsoft Word 2000. Also, several ways of utilizing the rhetorical structure of the help texts in information retrieval and text planning in the course of interaction with the user are proposed, so that it is possible to bring the computer closer to the people and make effective use of the existing resources.

Optimization of Fuzzy If-Then Rule Bases by Evolutionary Tuning of the Operations

This paper discusses optimization of fuzzy if-then rule bases by an evolutionary adjustment of the operations considered to use the rules. Adjustment of the operations is made by first parameterizing the operations with the help of either a contracting function or an order automorphism, followed by evolutionary tuning of the parameters. Optimization of a system based on Takagi-Sugeno rules and its distributed implementation are discussed in detail.

k-Additivity and C-decomposability of bi-capacities and its integral

k-Additivity is a convenient way to have less complex (bi-)capacities. This paper gives a new characterization of k-additivity, introduced by Grabisch and Labreuche, of bi-capacities and contrasts between the existing characterization and the new one, that differs from the one of Saminger and Mesiar. Besides, in the same way for capacities, a concept of C-decomposability, distinct from the proposal of Saminger and Mesiar, but closely linked to k-additivity, is introduced for bi-capacities. Moreover, the concept of C-decomposability applies to the Choquet integral with respect to bi-capacities.

Everyday language help system based on software manual

This paper presents a method for providing an everyday language smart help using an application software manual. This system takes three phases in the process: understanding of a users input, matching the result of understanding with manual texts of application software, and paraphrasing the matched texts to output for the user. We focus on the output phase, in which a manual text matched in the matching phase is paraphrased using the analysis result in the understanding phase. Four paraphrasing methods for a novice user are identified by a subject experiment, and rules for these methods are applied in respect to the linguistic features extracted in the understanding phase. An example scenario is shown, and the applied rules and their effects are discussed.

Script generation using rhetorical information in a task specification text

In order to extend the application domain of natural language interfaces to more realistic tasks without the decrease of users performance, it is desirable for users to be able to specify their requests as coherent texts consisting of more than one sentence. This paper presents a processing model of a natural language interface that accepts such an input text. Rhetorical information in an input text is used to generate a control structure of an output script program. Detection of loop structures, paraphrasing and dialogue are also incorporated. The algorithms explained in this paper have been implemented and evaluated in a personal email management domain.

Toward Elucidating Language Functions in the Brain

Human intelligence is characterized by the use of language rather than the brain hardware. The human brainware consists of a neural system as hardware and a language system as software. Language was created by the brain hardware, and the human brain evolved together with language over millions of years. It is, therefore, necessary to take two approaches to create the human brain: a hardware-centered approach and a software-centered approach. While the hardware-centered approach is based on computational neuroscience, it is possible to base the software-centered approach on linguistics.

Towards Programming in Everyday Language: A Case for Email Management

In order to extend the application domain of natural language interfaces to more realistic tasks without the decrease of user’s performance, it is desirable for users to be able to specify their requests as coherent texts consisting of more than one sentence, in other words, to write a program in everyday language. In this paper, we present a processing model of a natural language interface that accepts such an input text. It consists of the text understanding process using a systemic functional linguistic resource called the Semiotic Base, and the mapping process from the structure of the input text to the structure of an output computer program. The algorithms explained in this paper have been fully implemented in our everyday language programming system that deals with personal email management tasks.

What charaterizes intelligent systems: Issues of language

The human brainware consists of hardware as a neural system and software as a language system. It is said that protolanguage was born two millions years ago. Since then the brain and language have evolved together. It is considered that protolanguage had developed to language with grammar by a hundred thousand years ago. In such a way, language is the most complex system at the highest level among other systems on our earth. Without doubt, human intelligence is based on the functions of language. From this fact it is very natural to characterize an intelligent system by language: capability of handling language or capability of working with language. Discussing some issues of language related to intelligence, we focus on language functions of the brain in this talk. Why can a human speak? It is because the brain contains the system of language in it as a neurocomputational system: not only lexicon and grammar, but also meaning resources, context and mechanisms of understanding and generating texts. The system of language in the brain can be well described in terms of system networks by Systemic Functional Linguistics initiated by Halliday. He points out “four principles to design the language brain”: Stratification, Metafunction, Instantiation and Composition. Stratification is concerned with the hierarchical structure of language as phonology, lexicogrammar, semantics, and context. Metafunction, for example in the semantic stratum, is concerned with ideational, interpersonal and textual meanings. Instantiation is a relation as understanding and generation between meaning resources and texts. Composition is concerned with the logical structure of language.

Fundamental research for brain machine interface by use of EEG from right frontal gyrus

In order to develop a brain machine interface, the authors have investigated the brain activity during human recognition of characters and symbols representing directional meaning. The authors have recorded electroencephalograms (EEGs) from subjects viewing four types of Kanji (Chinese characters being used currently in the Japanese language) and arrows that were presented on a CRT. Each denoted direction for upward, downward, leftward and rightward, respectively. Subjects were asked to read them silently. Regardless of the directions, the reaction time was almost equal. EEGs were averaged for each stimulus type and direction, and event related potentials (ERPs) were obtained. The equivalent current dipole source localization (ECDL) method has been applied to these ERPs. In both cases, no large difference was observed until 250 ms at their latencies, and after that ECDs were localized to areas related to the working memory for the spatial perception. Taking into account these facts, the authors have investigated a single trial EEGs precisely after the latency at 400ms, and have determined effective sampling latencies for the discriminant analysis on four types of arrows: ↑, ↓, ← and →. By a discriminant analysis, the results of discriminant rate are 100% for each subject and for each trial. These results show the possibility of using EEGs for a brain machine interface in four type controls.