A Retrospective View of the Hearsay-II Architecture

The Hearsay model has heen presented as a paradigm for attacking errorful knowledge-intensive problems requiring multiple, cooperating knowledge sources. The Hearsay-II architecture is the latest attempt to explore the model. This paper describes experiences gained while successfully applying this architecture to the problem of speech understanding. The major conclusions are: 1. The paradigm of viewing problem solving in terms of hypothesize-and-test actions distributed among distinct representations of the problem has been shown to be computationally feasible. 2. A global working memory (the "blackboard"), in which the distinct representations are integrated in a uniform manner, has made it convenient to construct and integrate the individual sources of knowledge needed for the problem solution. 3. The use of a uniform data-directed structure for controlling knowledge-source activity has made the system easy to understand and modify. 4. A solution has been demonstrated to the problem of focusof-attention in this type of control environment. This solution does not need to be modified when the sources of knowledge in the system are changed. INTRODUCTION The Hearsay model [Red73Mo] has been developed for problem-solving in domains which must use large amounts of diverse, errorful, and incomplete knowledge in order to search in a large space. The Hearsay-1 architecture and system [Red73Hx and Erm74En] represented a first (and successful) attempt to apply that model to the problem of understanding connected speech in specialized task domains. In this first application, the size of the vocabulary (less than 100 words) and complexity of the grammar were very limited. Experiences with Hearsay-1 led to the more generalized Hearsay-II architecture [Les750r and Erm75Mu] in order to handle more difficult problems (e.g., larger vocabularies and lessconstrained grammars). The first configuration of knowledge sources (KSs) for Hearsay-II -configuration CI — was complete in January, 1976 [CMU76W4]. This implementation had poor performance (e.g., 107 sentences correct in 85 MIPSS (million instructions per second of speech) on a 250-word vocabulary). Experience with this configuration has led to a substantially different set of KSs — configuration C2 [CMU77Su]. This configuration performs substantially better (e.g., 857. correct in 60 MIPSS on a 1,000-word vocabulary). The Hearsay-II system, with the second configuration, has been successful: it comes close to the original performance goals set out in 1971 to be met by the end of 1976 for the ARPA speech understanding effort [New73Sp] and does so with a system organization that is of interest because of the potential for its application to other problem areas. Several other problems have been attacked with organizations strongly influenced by the 1 Author's current address: Department of Computer and Information Science, University of Mass., Amherst, Mass. 01003. 2 This work was supported by the Defense Advanced Research Projects Agency (F44620-73-C-0074) and is monitored by the Air Force Office of Scientific Research. 3 Other approaches for solving this class of problem include production systems, frames [Min74Fr], heterarchical structures [Wal770v and Woo76Fi], relaxation techniques [Bar76MS and Ros76Sc), Planner [Hew72De], QA4 [Rul73Qa], and the Locus model [Low76Ha and Rub77Lo]. Hearsay-II structure: image understanding [Pra77Se], reading comprehension [Rum76To], protein-crystallographic analysis [Eng77Kn], signal understanding [Nii77Ru], and complex learning [Sol77Kn]. This paper is divided into two major parts. The first part presents an overview of the Hearsay model, the Hearsay-II architecture, which is a further specification of this model, and the two KS configurations. (More detailed descriptions of these configurations are contained in the appendix.) The second part of the paper discusses the implication of these experiences for the Hearsay model and the Hearsay-II architecture. In particular, those aspects of the architecture are identified that have contributed most strongly to the success of the system, as well as those parts that need the most future work.** This discussion is structured around two themes -the multi-level global data base (blackboard) for KS cooperation, and the asynchronous, data-directed control structure for KS activation. OVERVIEW OF THE HEARSAY MODEL A number of characteristics of the problem drive the Hearsay model: 1. Large search space. 2. Diverse sources of knowledge. Many of the KSs are large; some have large internal search problems of their own. 3. Error and variability. These are characteristics of both the input data (the acoustic signal) and the processing of knowledge sources. 4. Experimental approach needed for system development. This implies the need for iterating the system and running over large amounts of data. 5. Performance requirement -accuracy and speed. This is true of any practical solution to the problem as well as during development (because of the experimental nature). The basic notions of the Hearsay model [Red73Mo] were developed in response to the requirements just stated: 1. The KSs are kept separate, independent, and anonymous. This separation is felt to be a decomposition which is natural and also can help make the combinatoric problems more tractable. For development purposes, the separation should help with system modifications (especially adding and modifying KSs) and evaluation. 2. A global data structure -the blackboard -is the means of communication and interaction of KSs. This provides an hypothesize-and-test means of interaction. Each KS accesses and modifies the blackboard in a uniform way. 3. A KS responds to changes to the blackboard which it is concerned with; it applies its knowledge within the context' 4 The fact that certain parts of the implementation need further work does not necessarily indicate deficiencies with the basic Hearsay model, but rather points out inadequacies in the Hearsay-II implementation of the model. It is to the model's credit that even though some of its more sophisticated capabilities are not implemented effectively, it still provides an appropriate framework for the successful solution of a complex task. Thus, one of the intents of this paper is to define some of the major design goals for the next iteration in the implementation of the Hearsay model. 5 While this paper discusses the means of organizing the knowledge and applying it to the problem, it does not describe in detail nor quantify the knowledge in the system. At least as much work has been expended on specifying and debugging the knowledge in the system as on building and refining the structure to hold and apply that knowledge. Specialal ized Svstems-2: 790 Lesser of such a change. This implies data-directed activation of KSs. OVERVIEW OF THE HEARSAY-II ARCHITECTURE The Hearsay-II architecture is one framework for implementing the Hearsay model. In this section, a very brief overview of that architecture is given. More details are described in [Les750r and Erm75Mu].