Site: Mitsubishi Electric Corporation
Computer and Information Systems Laboratory
5-1-1, Ofuna, Kamakura
Kanagawa 247, Japan
Date of Visit: May 25, 1995
Report Author: J. Foley
Mitsubishi Electric Corporation (MELCO) is one of several major Japanese corporations in the electrical/electronic industry segment, with roughly $30 billion in annual sales at the time of the JTEC visit. There are 15 research labs; all but one is in Japan. The lab employs about 4,000 researchers, with a budget that is 7% of annual sales.
The laboratories and those areas of research directly related to HCI are as follows:
The JTEC team visited the Computer and Information Systems Laboratory: researchers from the Industrial Design Center and Personal-Use Electronics Laboratory joined the team there for the presentations.
Masahito Tsuchiya's presentation focused on the process used for designing a new subway ticket vending machine. The design was done by Mitsubishi's Industrial Design Department, which performs usability testing and has developed a set of screen design principles.
The design process began with a survey of competitive systems currently in use at subway stations, which then led to a proposed new design. The new design was evaluated in a usability study with 5 evaluators. The sessions were videotaped. Evaluation was via protocol analysis, questionnaires, task completion timings, and error rates. The entire process was quite similar to those used by companies in the United States.
According to Masahiko Sakata, the Home Electronics Products Division has a five-person group that is developing a design methodology incorporating three aspects of human-computer interaction:
The Mitsubishi representatives stated that they are now concentrating mainly on the cognitive aspects. The emotional aspects have not been investigated, though our hosts see possibilities for this in the future. A methodical approach to this could make a contribution, if it is further developed and proven to be effective.
Katsushi Suzuki demonstrated FinalFiling, a commercially-sold document filing and retrieval system. It accepts scanned-in documents, does kanji character recognition, and creates keyword indexes for the documents. A contemporary GUI interface allows users to enter search criteria for either a keyword-based or full-text search retrieval. FinalFiling uses a client-server architecture with an SQL interface. A MELCO UNIX computer was used for the demonstration.
Dr. Yoshiharu Abe surveyed MELCO's commercial products in speech recognition, and then discussed his research.
MELAVIS is a speech recognition system that has been sold since March 1995. It is speaker-independent, recognizes isolated words from a 1,000-word vocabulary (the 1,000 currently-active words come from a larger 100,000 word vocabulary), and has a recognition accuracy of 95%. Dr. Abe believes it is most appropriately used for question-answer dialogues in which the speaker is being led through a series of steps, thereby limiting the number of words available at any point in time.
Dr. Abe also indicated that work is underway on continuous speech recognizers which can robustly deal with "ungrammatical order" of words, extraneous words, hesitations, repetitions, and unknown words. He did not discuss the techniques being used. A speech- to-text typewriter system is being developed.
The research work of Dr. Abe concerns recognizing phonemes using techniques that have an error rate half that of current approaches. Following phoneme recognition, phoneme sequences are recognized using statistical techniques. Then homonyms are disambiguated using a statistical corpus. Their approach has a 90% syllable recognition rate, versus 80% for other current approaches. Details of the new approach were not discussed.
Hironao Oz and Akira Takano demonstrated ASSISTAS, a commercially-sold system. ASSISTAS runs on MS Windows, and creates hyperlinks between or within files created with various Windows-based programs, including both custom and third party software. A "hot spot" is selected as the source for the hyperlink, and then a target for the hyperlink is specified. Hot spots of each source file are selected by either positions or logical entities: hot spots are still valid, even if the source file is edited so that logical entities move about in the file.