Mika Fukui

Development of document architecture extraction

The purpose of this study is the reduction of the burden in the document structurization process. A technique is presented for extracting the document architecture. As the technical document, 12,000 articles are extracted from the proceedings of a national convention. A summary of sample sentences as well as approximately 500 office documents within the organization also are examined as business documents. The rules for extracting the architecture are derived. The technique developed for document architecture extraction can extract such hierarchical structures as chapters and sections, as well as the reference structure to figures and tables from the technical document. The technique can also extract the hierarchical structure such as communications and reports from the business document. The technical and business documents can be discriminated by analyzing the character strings. As a result of evaluation using proceedings and in-office documents other than those used for deriving the rules, the error rate is 10.0 percent for the technical document and 23.0 percent for the business document. The error in extracting the reference structure is 8 percent. A field test is executed after improving the method so that the equations, figures and tables embedded in the text can be handled. The error rate is 5.4 percent for the technical document and 15.4 percent for the business document. It is verified through examples that the structurization can be achieved in a considerably shorter time than by manual processing. The developed document architecture extraction technique is commercialized as an automatic system by combining the technique with the layout attribute. The developed extraction technique will be utilized effectively in the hypertext conversion of the existing document and other problems, in addition to the layout processing.

Abbreviation Generation for Japanese Multi-Word Expressions

This paper proposes a novel method for generating Japanese abbreviations from their full forms with the Log-Linear Model (LLM) in order to take advantage of characteristic patterns of Japanese abbreviation. Our experimental results show that the method is effective for TV program titles that contain colloquial expressions. The proposed method achieved 78.8% recall for the top 30 candidates, whereas a baseline method using Conditional Random Fields (CRFs) achieved 68.3% recall. Moreover, from the results of experiments using six data sets classified according to types of character and semantic categories, we show that each performance of the above two methods depends on the types of the full forms.

Research on Model Based Document Processing System DARWIN

This paper describes a newly developed automatic document architecture extraction system with a document entity model. Hierarchical and anaphoric structural knowledge for chapter and anaphora is represented as a grammar in the document entity model. Automatic document architecture extraction is the first phase of a model based document processing system. DARWIN (Document Architecture Realization for Well-informed INterface). The purpose of DARWIN is to provide a natural and comfortable document processing environment for both authors and readers. Two other models, an author model and a reader model, are going to be embedded. Experimental results have shown that the automatic document architecture extraction system is sufficiently powerful to detect hierarchical and anaphoric structures. This paper is structured and formatted by the current DARWIN system. in which a simple reader model is implemented

Improving the robustness to recognition errors in speech input question answering

In our previous work, we developed a prototype of a speech-input help system for home appliances such as digital cameras and microwave ovens. Given a factoid question, the system performs textual question answering using the manuals as the knowledge source. Whereas, given a HOW question, it retrieves and plays a demonstration video. However, our first prototype suffered from speech recognition errors, especially when the Japanese interrogative phrases in factoid questions were misrecognized. We therefore propose a method for solving this problem, which complements a speech query transcript with an interrogative phrase selected from a pre-determined list. The selection process first narrows down candidate phrases based on co-occurrences within the manual text, and then computes the similarity between each candidate and the query transcript in terms of pronunciation. Our method improves the Mean Reciprocal Rank of top three answers from 0.429 to 0.597 for factoid questions.