Grand Challenge AI Applications

1 Introduction

Grand challenge AI applications are applications with significant social, engineering and scientific im­ pacts, whose development will require substantial scien­ tific and engineering break through, and use of high­ performance computing and large-scale data resources. In the high performance computers and communication act of 1991, the term grand challenge is defined as a fundamental problem in science or engineering, with broad economic and scientific impacts, whose solution will require the application of high-performance comput­ ing resources. Since this definition assumes grand chal­ lenges for scientific computing, or numeric computing field, it focuses on a fundamental problem in science or engineering rather than applications. The definition deliniated at the outset of this paper would clearify what are grand challenge AI applications.

In Feburary 1992, the National Science Foundation (NSF) hosted a workshop, chaired by Professor Ben­ jamin Wah, titled The Workshop on High Performance Computing and Communications for Grand Challenge Applications: Computer Vision, Natural Language and Speech Processing, and Artificial Intelligence.

In October 1992, Kitano hosted a private workshop called The Workshop on Grand Challenge AI Applica­ tions, participated in by major Japanese AI researchers with several Japanese, American, and European ob­ servers [Kitano, 1992]. The next workshop will be held in October 2000, as The International Workshop on Grand Challenge AI Applications: Accomplishments in the 20th Century and Prospects for the 21st Century.

Apart from these workshops, there are several projects which are engaged in grand challenge AI applications. For example, Carnegie Mellon University, ATR Inter­ preting Telephony Research Laboratory, and other re­ search institutions have been working on developing speech-to-speech translation systems- a representative Walther v .Hahn University of Hamburg Ryuuichi Oka Real World Computing Program Toshio Yokoi Japan Electric Dictionary Research Inst.

grand challenge AI application. Already, several pro­ totype systems, such as SpeechTrans, Dm-Dialog, SL­ Trans, JANUS, ASURA, and Intertalker have been de­ veloped and publicly demonstrated. Recently, increasing numbers of research institutions and national projects have decided to tackle the grand challenge AI applica­ tions. Professor Hahn will describe the project verbmo­ bil, a german national project to develop a speech-to­ speech translation system.

Grand challenge AI applications often requires sub­ stantial involvement with non-AI, or even a non­ computer science community. One typical example is the Human Genome project where molecular biologists play a major role. AI, however, is one of the critical aspects achieving success in the project . Dr .. Hunter describes the international human genome proJect and how AI can contribute to the project, as well as getting benefits from the project.

Computational resources and technical breakthroughs would be necessary to implement the challenge. In the U.S., the high performance computing and communica­ tion project aims at development of hi~h-perform~nce massively parallel computers and ultra h1gh-band w1dth communication networks. In Japan, MITI sponsered the Real World Computing Program which was started in 1992. The projects aims at the development of high performance massively parallel computers and basic re­ search for grand challenges.

High-performance computing and communication alone are not sufficient for the challenge. Availability of massive data resources is a critical factor in bringing the efforts to success. There are already several efforts to build large-scale knowledge-bases and data-bases. Dr. Yokoi describes the EDR's electric dictionary projects, and. argues the need for the NOAH Knowledge Archive proJect.

Architectural Supports

Benjamin W. Wah Coordinated Science Laboratory University of Dlinois Urbana, IL 61801 USA vahGcrhc.uiuc.edu

Many grand challenge applications in AI are related to the representation, acquisition, storage, and processing of knowledge. As computer systems supporting these ap­ plications are generally built from knowledge and search, high performance computing and communication sys­ tems can make a dramatic difference in the amount of knowledge represented and the search performed. At the NSF sponsored HPCC Workshop on Computer Vision, Speech and Nat ural Language Processing, and Artificial Intelligence last year (Wah, et al, 1993), participants identified many grand challenge applications that involve the use of AI methods. Examples of these applications include the following.

1. Electronic librarian. This involves the design of an electronic librarian that can navigate through a large amount of stored knowledge and information in an elec­ tronic library, interact with users in a natural language interface, and update the library automatically when new information is received. Relevant AI methods in­ clude knowledge representation, non-monotonic reason­ ing, and inexact searches. These methods must operate in large databases and adapt to different needs of users.
2. Nation-wide job-bank. This provides an online sys­ tem that matches the skill profiles of job applicants against jobs listings. Relevant AI methods include con­ straint satisfaction, agent modeling, and possibly intel­ ligent and interactive animation.
3. Electronic market place. This system provides a way for suppliers and customers to negotiate prices and conclude transactions.
4. Intelligent planning and scheduling. This involves the planning of resources in commercial applications, such as flight scheduling, and in manufacturing. It re­ quires the modeling of a large amount of inexact informa­ tion and the optimization of a multi-objective problem.

To understand how HPCC systems can be designed to support these applications , we must first understand the characteristics of these applications. In the follow­ ing we first discuss the characteristics of AI application tasks. This is followed by a discussion of architectural requirements on HPCC systems for supporting these ap­ plications.

1. Vast amount of knowledge and information. Many AI applications deal with a large amount of knowledge and data. Examples include an electronic library and large corpora for supporting studies in this area. This in­ formation may be represented in many different schemes and may be interdisciplinary. It is anticipated that more than one internal knowledge representation schemes will be needed in an integrated system. Related architec­ tural issues include the design of suitable representa­ tion schemes, hardware storage mechanisms, and storage methods that can adapt to different application require­ ments.
2. Large number of users with different background. A robust system supporting AI applications will need to have an intelligent interface that allows users with differ­ ent background and goals interact with it . Advances in this area will require the design of natural language in­ terfaces, interactive HDTV and computer graphics, and systems exploiting multimedia, visualization , and virtual reality technologies. These devices provide user friendly interfaces between users and systems.
3. Inexact knowledge and information. By defini­ tion, an intelligent system evolves with time. Conse­ quently, knowledge stored in such systems will be up­ dated incrementally, and existing knowledge may be in­ validated with new knowledge acquired in the future. Non-monotonic reasoning is, therefore, an essential com­ ponent of an intelligent system. Architectures suitable for such representation and reasoning may involve asso­ ciative search . A familiar example is a hardware imple­ mentation of an artificial neural network.
4. Knowledge acquisition and machine learning. Ma­ chine learning is an important process for acquiring

knowledge automatically in an intelligent system. Tech­ niques used can be simple and involve the adaptation of parameters in an algorithm (such as a vision system), or can be very complex. Examples of the latter include the synthesis of vision algorithms (such as stereo) from ex­ amples of desired performance, the acquisition of models and problem solving heuristics, planning, and the acqui­ sition of lexical, grammatical and discourse knowledge in natural language processing. Important issues to be con­ sidered in learning include the scaling and generalization of knowledge learned to more general application prob­ lems and a trade-off between the quality of the knowl­ edge acquired and the amount of time taken to learn it. Computer systems supporting machine learning may include high speed massively parallel systems for induc­ tive experimentation and a general purpose system for supporting high level reasoning.

1. Multi-objective optimization. Many AI applications involve the optimization of subgoals that are not pre­ cisely specified. Such problems are often posed as op­ timization problems with ill-defined objectives and im­ precise constraints. Further, objectives and constraints may depend on measurable attributes that are too many to be included in the formulation. This ill-formedness leads to imprecise feedback (credit assignment) between actions and results. The problem here is analogous to a non-convex optimization problem over many hundreds or thousands of parameters. Techniques for solving these problems are often formulated as searches under syntac­ tic and semantic constraints. Architectural studies in this area are directed towards load balancing, schedul­ ing, and scalability.
2. Integration of existing results. In many AI appli­ cations known today, results have been demonstrated in a small scale in individual components. A lot of work remains to integrate these results into a single frame­ work for real-world applications. For example, in spo­ ken language translation , there is active research in inte­ grating component modules, such as speech recognition, language analysis, message-content manipulation, mes­ sage planning, language generation, and speech synthe­ sis. Such integration may complicate the problem in an exponential fashion, demanding massive parallelism and high speed communication.
3. Large computational requirements and distributed processing. Many AI applications are characterized by algorithms that require large computational power, high communication bandwidth, and extensive memory. Ex­ amples include the search of large knowledge bases and the processing of a large number of distributed transac­ tions and updates. Tasks may originate from applica­ tions with a large variety of features and objectives and exist in different granularity, requiring heterogeneous ar­ chitectures that can support fine-grain as well as coarse­ grain parallelism and that are connected by high speed communication networks. Moreover, these architecture may need to evolve with time when emerging technolo­ gies are available. Related and difficult issues include the design of software systems for supporting these diverse applications and the intelligent scheduling of computa­ tional resources among tasks.

The following are some architectural issues that are important in addressing the needs of grand challenge ap­ plications (Siegel, Abraham, et al., 1992) .

• Model formulation. It is important to develop a model of computation that can abstract operational features and costs of parallel computation. Such a

• model can form the foundation for software devel­ opers and programmers to build their applications.
• General purpose supercomputing. Supercomputers are important tools for simulating prototypes be­ fore they are actually implemented. Researchers can test their algorithms on more realistic data before migrating to massively parallel systems and special purpose VLSI architectures.
• Homogeneous architectures. To effectively use mas­ sively parallel systems with homogeneous proces­ sors, it is critical to provide global (logical or phys­ ical) virtual memory that simplifies programming, language tools for expressing and for supporting parallelism and the interaction of concurrent activ­ ities, and compilers for optimizing system perfor­ mance.
• Special purpose VLSI architectures. These are im­ portant when it is necessary to deploy systems that have stringent real-time requirements. In addition, emerging technologies may allow new problems to be solved in a reasonable amount of time.
• Infrastructure support for prototyping architectures. This consists of various simulation tools and proto­ typing facilities for computer architects to test their ideas . A particularly difficult problem involving AI applications is that it is often hard to instrument an operational system in order to test new ideas and algorithms.

Grand Challenge AI Applications Proposed in Japan

Hiroaki Kitano NEC Corporation Software Engieering Lab.

Shibaura, 2-11-5, Minato

Tokyo, 108 Japan hiroakiiDcs. emu. edu

At the workshop on grand challenge AI applications, which was held 6 October 1992, in Tokyo, various ap­ plications were proposed, which can be roughly cate­ gorized into six areas: games, science, evolutional sys­ tems, robots and replicant, communication, and infras­ tructure.

Games: Grand Master Computer Shogi and Go Sys­ tems.

Science: Anatomical Mapping of total brain wiring.

• Evolutional Systems: Systems which evolve by them­ selves for use in building large-scale distributed and open systems, and to allow robots to be operated in a dynamically changing environment.
• Robots and Replicants: Robots and programs which undertake daily matters, such as making coffee, an­ swering the phone, cleaning floors, and delivering mails.
• Infrastructure: Knowledgeable community, Global Architecture for Information Access (GAIA), and auditory scence analysis system.

Several interesting observations have been made. For example, the importance of Shogi and Go has been em­ phasized. As Chess has almost finished its role as a major technical challenge, Shogi and Go emerge as new challenges. While estimated search space for Chess is around 10 120 , Shogi has about 10 200 and Go has about 10 230 . Rules for playing Shogi prevents search space to be narrowed down, even at the very end of the game. Go is a completely different game from Chess or Shogi . Developments of Grand Master Shogi and Go systems will be significant accomlishments, and the challenge is expected to produce various new technologies.

The goals for replicants and robots have been to carry out daily tasks. This is integrated with the challenge of achieving a highly flexible user interface research, since replicants and robots communicate using natural lan­ guage and other media. While the goal of undertaking daily tasks does not have the flavor of big science, its accomplishment would bring major impacts to business and consumer markets. Even premature systems, if ap­ propriate marketing can be carried out, could result in yet another workman.

Knowledgeable community and GAIA discuss the pos­ siblity of building large scale and multi-purpose informa­ tion infrastractures. While knowledgeable community stresses formal knowledge and multi-agent approach to support expert reasoning, GAIA aims at a more gen­ eral information access infrastruct.ure, reinforced by AI technologies.

Various breakthroughs in AI are necessary to meet the challenge. Basic technologies needed for the challenge are: massively parallel artificial intelligence, genetic al­ gorithms, memory-based reasoning, parallel search algo­ rithms, real-time planning and reasoning, integration of symbolic and connectionist processing, constraint-based processing, distributed AI , robust parsing, active vision, language learning, learning theory integrated with rea­ soning, situated theory, and flexible robot mechanism. This list contains a broad range of AI technologies, but it is not a random list. It selectively focuses on technolo­ gies to tackle real-world problems. In addition, technolo­ gies to handle incomplete and unstructured knowledge have been commonly recognized as an important factors . Building and sharing of large-scale data resources such as speech database, corpus, electric dictionaries, and image database, are considered to be of prime importance.

The list can be summarized as follows:

• Central Nervous System: Massively parallel artifi­ cial intelligence technologies for reasoning, learning, search, real-time processing, integration of symbolic and numeric processing, and evoluticaal systems.
• Knowledge Source: Very Large Knowledge-Bases, building and sharing of large-scale data resources. Technologies to represent and process incomplete and unstructured knowledge.
• Input/Output: Presentation technologies, such as vir­ tual reality. Recognition technologies, such as speech recognition, auditory scence recognition, and image understanding.
• Moti~n Control: AI and robot mechanism integra­ tiOn.
• Memory Storage: Tera-Peta byte size mass s~orage systems and efficient access methods.

Computers and communication systems should aim at achieving the following specifications, to be used as a basis for meeting the challenge.

• Computers: Massively parallel machines with tera­ fl'ops computing speed, tera-byte memory-space, peta or exa-byte scale mass storage, and tera­ bits/second I/0. The CM-2/-5 equivalent systems should be desk-side size in 10 years.
• Devices: Special purpose devices, such as WSI-MBR, WSI-NN, WSI-based audio stream separator, and

vision devices, as well as Giga-byte memory chips and holographic memory.

Communication: Giga-bits/second end-to-end com­ munication network. The first step should be B­ ISDN with 622 Mbits/seconds. This should be up-graded to OC-192 of synchronous optical net­ work (SONET) . Then, it should be up-graded into a Giga-bits/seconds end-to-end network as the third phase.

AI and Grand Challenges in Biotechnology Computing

Lawrence Hunter Director, Machine Learning Project National Library of Medicine Bethesda, MD 20894 USA hunterGnlm.nih.gov

Several HPCC Grand Challenge Problems involve biomedical applications, and Artificial Intelligence meth­ ods are an important component of these approaches. The "Grand Challenges 1993" report identifies biotech­ nology computing and digital imaging as the two focus areas for development of advanced software technology and algorithms. It is in the area of biotechnology com­ puting that AI has the potential for greatest impact.

The international Human Genome Project , underway since 1990, plans to produce complete genetic maps and sequences for human beings and a variety of model organisms by the year 2000. Currently, the sequence databases alone contain more than 500MB of sequence data, and are doubling roughly every 18 months. A DNA sequence, which can be represented as a string drawn from a four letter alphabet, is the blueprint of every liv­ ing organism. The genome (complete DNA sequence) of a simple laboratory bacterium contains about 10,000,000 elements; human DNA contains about 3,000,000,000. Even the partial results obtained so far in this effort have had significant social and scientific importance. The causes of several inherited diseases have been identified, including Cystic Fibrosis, the most common inborn error of metabolism. These results have also led to improve­ ments in therapy for such diseases, as well as for cancer, AIDS, heart disease and other major human health prob­ lems. In addition to health care, knowledge of genetic mechanisms and gene sequences has led to revolutionary changes in agriculture, materials science, pharmaceuti­ cals and even law enforcement. It is probably not an exaggeration to say that the coming biotechnology rev­ olution will be as rapid and significant as the computer revolution.

In fact, an important part of the underlying infras­ tructure for the explosive growth in our biological un­ derstanding and ability to do bioengineering involves the use of high power computation. Biotechnology comput­ ing involves a wide variety of specific applications, rang­ ing from laboratory robotics in support of genome se­ quencing projects to the model-based simulation of cel­ lular function and evolution. Many of these applications are well suited to AI techniques; they require finding patterns in large amounts of data, or qualitative mod­ eling, or representation and reasoning about complex, structured knowledge. The connection is quite deep; renown biophysicist Harold Morowitz claims that" com­ puters will be to biology as math is to physics." It is beyond the scope of this abstract to summarize all of this work, but see (Hunter, 1993), or the special issues of the Communications of the ACM (1991) or IEEE Ex­ pert (forthcoming) for surveys.

The amounts of data involved are often quite large, so that AI approaches must be scaled up in order to be effective. That is why the biotech area is such fertile ground for AI/HPCC research. Here I mention as ex­ amples three brief cases from my own work. First is the application of Bayesian classification to classifying pro­ tein structural elements (Hunter & States, 1992). This research involved the largest application of the AUTO­ CLASS III program, developed at NASA. The inputs to be classified were more than 10,000 protein fragments, each described by as many as 72 real numbered features . Running on a powerful SGI 4D280 server, the classifi­ cation took more than three weeks of CPU time. Sev­ eral other promising approaches to understanding pro­ tein structure with AI techniques are as yet computa­ tionally intractable. One approach has been the devel­ opment of a massively parallel version of AUTOCLASS ~hich should significantly expand the potential applica­ tion area.

Another large scale AI application involved the de­ velopment of a novel unsupervised learning algorithm that could be applied to massive, unsegmented datas­ treams (Harris, Hunter & States, 1992). This project involved finding clusters within the protein sequence databases. These sequence databases contain more than 70,000 sequences containing more than 150,000,000 el­ ements. The cluster works by looking first at all the pairwise similarities (more than 2,000,000) of them, and then does an approximate transitive closure of the bi­ nary similarity measures. This application takes more than 10 hours of supercomputer time to complete, and needs to be updated periodically as the databases grow. The resultin~ classification is supplanting manually de­ rived definitiOns of protein families and domains, and provides a valuable tool for automatically annotating the results of large scale genomic sequencing projects.

A final example identifies the potential synergistic role of AI techniques and other advanced algorithms. The bone disease Osteogenesis Imperfecta (01) is caused by mutation in the gene for the protein collagen. The range of manifestations of the disease is quite broad: some mu­ tations are lethal, and others are nearly asymptomatic. A goal of 01 researchers is to understand the molecular mechanisms that determine the significance of a partic­ ular mutation. Molecular dynamics simulations of colla­ gen in its entirety are computationally intractable, and none of the small models proposed have led to significant insights into the mechanisms of the disease. In current work (Klein, Hunter, Byers, unpublished) we have ap­ plied machine learning techniques to sort through the hundreds of potentially relevant features of the environ­ ment of the mutation to identify which are in fact rel­ evant to lethality. The results of the machine learning application are then used to guide molecular dynamics experiments to simplified models that are both compu­ tationally tractible and capture the important molecu­ lar interactions. In this way, heuristic machine learning methods can be used to focus attention on important subproblems where more formal but computationally in­ tensive methods can be applied.

A significant impediment to progress in this area is the lack of support for AI research on very high speed com­ puters. For example, there are very few available super­ computer LISP environments available to researchers. This is due in large part to lack of demand from AI

researchers themselves. Although *LISP was one of the first languages available for the massively parallel Con­ nection Machine, demand for other languages (especially parallel fortran and C) have dramatically outstripped LISP users, and support at Thinking Machines for *LISP development has dropped off. Also, many of these prob­ lems require massive memory as well as high speed com­ puters. The sequence clustering work is constrained by space as well as time tradeoffs. The method requires ap­ proximately 700M of memory for the current databases, and the databases are doubling in size every 18 months. The program is much faster if it doesn't have to page, so in the near future we will require machines with a giga­ byte of RAM. We have found that large scale AI prob­ lems tend to require more memory than CPU speed, and that large memory machines are an important need.

In summary, both scaling up AI applications and ap­ plying intelligent techniques to guide the use of non-AI methods are likely to be important in addressing grand challenge problems. AI researchers must work hard at scaling their approaches to grand challenge problems, and demand the supercomputing resources (including memory and software, not just megaflops) necessary to accomplish their visions.

The Real World Computing Program

Ryu-ichi Olea

Tsukuba Research Center Real World Computing Partnership Tsukuba Mitui Bldg., Floors.17 1-6-1 Takezono,Tsukuba City, lbaraki 305, Japan

The Real World Computing Program (RWC) is a new national research program funded by Japan's Ministry of International Trade and Industry (MITI) at a level $500 million for the next 10 years. RWC has as its over­ all technical objective the development of flexible and advanced information technologies that are capable of processing a variety of diversified information (such as images, speeches, texts, and so forth). RWC emphasizes technologies that match the flexibility of human infor­ mation processing capabilities such as pattern recogni­ tion, handling incomplete information, and learning, all of which are manifested in the way people solve problems in the real world.

The approach to the objective of RWC research pro­ gram is structurally explained as follows. The flexible processing will be based on four types of function; un­ derstanding, learning, inference, and control. Patterns and symbols are basic and common materials laid in the real world. The functions will be required to carry out the integration of symbols and patterns. The integration will be realized by exploring both problems of represen­ tation and algorithm which work in the real world. The representation is the framework for modeling the real world and also strongly related to parallel systems in­ cluding optical systems. The algorithm is the dynamics for solving problems formulated in terms of the represen­ tation. Neural models are candidates that are expected to create both the representation and the algorithm.

Five major research themes in the following give more concrete descriptions of the technical objective.

Theoretical foundations: The theoretical foundations for flexible information processing will be a general framework for solving problems by the integration of symbols and patterns in conjunction with learning and self-organizational ones. In order to reveal relationships between the integration and conventional methods, the­ ories for pattern recognition, multivariate data analysis, probabilistic and statistic inference, neural computing, and machine learning will be deepened and unified.

Novel functions: The goal is to develop the technolo­ gies for the flexible information processing that conven­ tional technology lacks in terms of robustness, open­ ness, and real-time capability. Specific research goals include (i) establishing the schemes fr,r flexible recogni­ tion and understanding of multi-modd.l information, in­ cluding moving images, voices, texts, and gestures, and developing interactive information systems by which hu­ mans can communicate with such information media, and (ii) developing flexible systems that are able to au­ tonomously sense and understand and also control the environment through interactions with the real world.

Massively parallel computing: RWC-1/2, massively parallel systems which will be developed to support the "real world computation". Research goals include (i) providing multiple computation paradigms such as shared memory, message passing, data parallel, multi­ threading, neural networks, (ii) establishing flexible adaptation to the execution status of problems and com­ putation resources, i.e. they will have the mechanisms for being adapted to the system load condition, vari­ ety of properties of information, requirements of real­ timeness, robustness both for hardware and software, etc., (iii) providing computation and space capacity to permit an unpredictable amount of computation which will permit programmers the naive description of their problems, and (iv) providing fancy fast interconnection network and synchronization mechanisws both of which are essential for massive and unstable activities like neu­ rons and elements of other novel models.

The software architecture of the massively parallel system consists of the kernel, implementation language, base language, operating system, and programming en­ vironment layers. The base language supports the notion of object-orientation, reflection, and time dependent pro­ gramming. Using reflection, the programming language semantics as well as the resource management strategy is modified by the programming language itself. The new language functionalities might be easily constructed by the base language. Since we have to consider "time" in the real world, the base language provides the facility to describe the real-time system.

Neural network systems: Here we wish to establish a quite new type of neural model or computation which is different from conventional models such as Back­ propagation, Boltzmann machine, and so forth. Con­ nectionist models will be pursued as candidates that will provide a new scheme for representations of the real world. The domain of "neural" research should be ex­ tended to deal with materials in the real world.

Optical computing systems: This part of RWC aims at establishing basic technology for (i) optical intercon­ nection devices and networks, (ii) optical neural models, devices, and systems, (iii) optical logic devices, circuits, and digital systems, and (iv) advanced opto-electronic integrated circuits development environments.

The objective of RWC meets the grand challenge prob­ lems in AI in both conceptual and computational as­ pects. First, the approach from the pattern side toward the integration of symbols and pattern should reveal new models or algorithms suitable to obtain new function with robustness and openness beyond ones of conventional pattern recognition and AI. Scalable algorithms with simplicity and transparency will be also expected to process a large mount of raw data such as images, voices, texts, and so forth. Second, real-time realization of those models or algorithms will require a new archi­ tecture of massively parallel computation which has a wide spectrum of I/0 channels that contact tightly with the real world.

The typical application of the RWC will include the re­ alization of humanoid assistant which is capable to sup­ port human activities interactively between users and the real world. The support means that i) to suggest a new viewpoint to a human user in his creative activities in terms of automatic search and inference of items in a large amount of data base including images and voices, ii) to take care of handicapped peoples in daily life by means of intelligent monitoring and so forth.

In conclusion, RWC is not a simple successor of the Fifth Generation Computer Project , but will promote challenging researches to solve intrinsic problems related to AI and pattern recognition. The key concept of RWC is the integration of symbols and patterns in the best sense of the words.

Very Large-Scale Knowledge Bases - From Lexical Knowledge to World Knowledge-

Toshio Yokoi

Japan Electronic Dictionary Research Institute (EDR)

Mita-Kokusai Building Annex, 4-28, Mita 1-chome, Minato-ku, Tokyo 108, Japan yokoi41edr . co. jp

The major themes of very large-scale (massive) know l­ edge, very large-scale knowledge acquisition and very large-scale knowledge bases, are the basic "Grand Chal­ lenge" common to all attempts in the AI field . Two projects are presented here: the EDR Electronic Dictio­ nary Project and the NOAH Project . The EDR Elec­ tronic Project, which is presently under way, deals with lexical knowledge, while the NOAH Project that is be­ ing planned will deal with world knowledge. The NOAH Project will pursue a new development in the technology by uniting knowledge processing and natural language processing. This project will need to be carried out un­ der appropriate international cooperation. An interna­ tional conference of KB&KS '93 will be held in December of this year in order to create such worldwide consensus.

1. EDR Electronic Dictionary Project

The EDR Electronic Dictionary Project is a nine-year project that started in 1986. It aims to research and de­ velop large-scale computer dictionaries required for en­ abling computers to process and understand natural lan­ guages. The target languages are Japanese and English. Total funding for research and development is 14 bil­ lion yen (approximately $100 million) . Seventy percent of this amount is financed by the Japanese Government while the remaining 30% by eight representative com­ puter manufacturers. The Japan Electronic Dictionary Research Institute, Ltd. serves as the promoting nucleus.

All the results of the EDR project will be sold at reasonable prices. The same conditions regarding the use of the EDR Electronic Dictionaries will be applied to both domestic and overseas users. We will set the prices much lower than those of the machine-readable dictionaries that are currently on sale. Naturally, for academic users, such as universities and public research institutions, special measures will be arranged, includ­ ing prices, etc. Discussions for realizing sound mutual cooperation with various organizations overseas have fi­ nally started. To ensure that this cooperation continues permanently and produces fruitful results, the results of the Electronic Dictionaries must be inherited by an or­ ganization established to develop them. Discussion of this issue has just started and has not yet reached the stage where any official comments have been expressed. Yet , a su~estion would be to establish an appropriate organization to expand and promote better use of the Electronic Dictionaries. Any corporation or person who wishes to use the Electronic Dictionaries should be a member of this consortium. The members would pay an entrance fee and an annual membership fee. The consor­ tium would be run using the funds gathered from these fees and with the assets inherited from EDR. The mem­ bership would be roughly divided into two classes: one for commercial users and the other purely for academic users. The members, no matter whether they are do­ mestic or overseas, should equally enjoy the benefits of the consortium. Furthermore, the consortium should ac­ tively promote the creation of an international network of electronic dictionaries by, for example, establishing a working relationship with similar overseas organizations.

2. NOAH (Knowledge Archives for Building and Sharing Very Large-Scale Knowledge) Project

In March of last year, the first edition of its plan was drawn up and official examinations of the project started. The NOAH Project centers around knowl­ edge processing and natural language processing, and aims to research and develop new technolo~ies related to knowledge itself. Natural language is constdered knowl­ edge representation language used widely and steadily. Documents expressin~ organized information in natu­ ral language are constdered the basic source of knowl­ edge. Therefore, natural language processing and text processing are considered the basic technology of knowl­ edge processing. Based on these views, this project aims to develop various new functions that acquire and col­ lect massive amounts of knowledge, share the knowl­ edge, reuse the knowledge and support the creation of new knowledge. It is a project related to technologies for software, knowledge, and information widely asso­ ciated with such fields as natural language processing, text processing, knowledge engineering, hypermedia, ad­ vanced database, software engineering, etc., and with the knowledge and information technologies. The sharing of knowledge is naturally to be carried out on a global scale. The NOAH Project aims to research and develop new technologies for realizing global-scale knowledge sharing environments. The foundation for this would be the sharing of knowledge related to the various languages, that is, the sharing of linguistic data such as the elec­ tronic dictionaries. By developing mutually cooperative international networks achieved by the Electronic Dic­ tionaries project, the NOAH Project strives to seek the ideal form of new international cooperation. For this reason, an international conference is scheduled to open in Tokyo this year. This conference will be promoted from its planning stage with wide-range participation of governmental organizations and researchers from vari­ ous countries. We strongly hope that the Knowled~e Archives Project will stimulate similar projects withm each country and play a significant role in the research and development of knowledge processing and natural language processing.

THE PROJECT VERBMOBIL: Mobile speech-to-speech translation

Walther v. Hahn University of Hamburg, Germany vhahn0nats4.informatik.uni-hamburg.de

The project verbmobil is a German national project for the research and development of speech-to-speech translation system. The project overview is as follows:

Background

• Increasing Language problems in the European Commumty
• Foreign language technology as an economical factor to raise the international market accessability of the industry.

Ambitious tasks are a challenge and will increase the overall awareness for the field as well as concrete re­ search activities in natural language and speech process­ in~ . The project takes up previous and on~oing work bemg performed at CMU and ATR, and will work in direct cooperation with these institutions.

Specific Goals:

Mobile speech-to-speech translation device from and to German. Two steps are planned:

1. the contributions of the dialogue partners are both translated to written English on a screen ,
2. additionally there is a target language generation and speech output.

Scenario: Negotiation dialogues starting with time schedule tasks.

Size and schedule: 20 partners from universities and industry will participate with a total funding budget of 10 - 15 million DEM per year . Additional financial in­ house support of the industrial partners is expected. The run-time of the project will be 8 - 10 years, beginning with 1993. Mile-stones are every 3 years.

Research Issues

• Softw~re architectures for speech and language in­ tegratiOn.
• Cog!litive models of human translation and inter­ prel.mg.
• Development of functional specifications for differ­ ent types of MT systems.
• Description and processing of spontaneous language (hesitations, ellipses, prosody etc.) .
• Evaluation and integration of symbolic and neural network models for different tasks (hybrid modules) .
• Integration of domain knowledge, cultural knowl­ edge, and linguistic knowledge.
• Large corpora research.
• Discourse representation

Central Design Decisions

• English as intermediate language in the sense that both partners are expected to have elementary pas­ sive skills and all contributions are translated into English (additionally) to provide control facilities.
• Transfer techniques, no interlingua.
• Interactive architecture for top-down and bottom­ up processes

Expected Technical Effects

• Concentration of national efforts by one research is­ sue in speech and language,
• Reinforcement of international cooperation,
• Opportunity for natural language processing to solve real-world tasks,
• Evaluation of what is feasible at the moment under real-application conditions,
• Professionalizing of the natural language processing work,
• Linguistic robustness for handling spontaneous lan­ guage.