An understander reading or listening to someone speak has to repeatedly solve the problem of word-meaning ambiguity, the selection of the intended meaning of a word from the set of its possible meanings. For example, the problem of pronominal reference can be considered as a choosing of the intended referent from the collection of entities which have already been mentioned or which can be inferred. Human understanders apply rules of syntax, surface semantics, general world knowledge, and various types of contextual knowledge to resolve word-sense or pronominal ambiguity as they process language. We describe a mechanism, called a cooperative word-meaning selector, which allows the computer to use various knowledge sources as it ``understands'' text. The word-meaning selector is part of a conceptual analyzer which forms the natural-language interface for a pair of multiprocess language processing systems. The first, called DSAM (distributable script applier mechanism), reads and summarizes newspaper articles making heavy reference to situational scripts. The second, ACE (academic counseling experiment), is a conversational program which automates certain parts of the academic counseling task. In each of these systems, a variety of knowledge sources, each managed by a distinct ``expert'' process, is brought to bear to enable the word-meaning selector to form the most plausible reading of a sentence containing ambiguous words.

The ability to explain something, e.g., the operation of a complex machine or program, is an important, but poorly understood, component of intelligent behavior. We discuss an artificial intelligence approach to the modeling of the explanation process within the framework of a graphics-based CAD system currently under development, which can describe its own use, including the common ways to make and recover from errors. With a coordinated textual and pictorial display, the system, CADHELP, simulates an expert demonstrating the operation of the graphical features of the CAD subsystem. It consults a knowledge base of feature scripts to explain a feature, generate prompts as the feature is being operated, and to give certain types of help when a feature is misused. CADHELP provides these services by summarizing a feature script in different ways depending upon what it has told the user previously. The summarization process is based upon a theory of natural-language generation, in which a concept to be expressed is replaced in a short-term memory by words spanning part of its meaning interspersed with subconcepts still to be expressed. The generation process is mediated by a series of "sketchification" strategies, which prescribe which parts of a knowledge structure, a causal chain, or a single concept need not be expressed, since the listener should be able to infer them.

This paper considers the possibilities for knowledge-based automatic text translation in the light of recent advances in artificial intelligence. It is argued that competent translation requires some reasonable depth of understanding of the source text, and, in particular, access to detailed contextual information. The following machine translation paradigm is proposed. First, the source text is analyzed and mapped into a language-free conceptual representation. Inference mechanisms then apply contextual world knowledge to augment the representation in various ways, adding information about items that were only implicit in the input text. Finally, a natural-language generator maps appropriate sections of the language-free representation into the target language. We discuss several difficult translation problems from this viewpoint with examples of English-to-Spanish and English-to-Russian translations; and illustrate possible solutions as embodied in a computer understander called SAM, which reads certain kinds of newspaper stories, then summarizes or paraphrases them in a variety of languages.

Recent research in artificial intelligence has identified a number of knowledge sources which appear to be needed for effective automatic "understanding" of connected natural language speech and text. These include word-and phrase-level semantics, models for actors and objects, and inference techniques for using script- or goal-oriented knowledge structures. A unified model of the understanding process can be defined using the distributed-computing viewpoint, provided some way can be found to integrate and control a collection of "experts," each one associated with a certain kind of knowledge source. A technique is described, called hierarchical task management, for constructing computer language-processing systems comprising an arbitrary number of distinct, potentially distributed, processes. The technique is based upon the repeated activation and expansion of data structures called tasks, which define important components of the understanding process in terms of a controlled interaction among the experts. The tasks are maintained on several "agendas," and are manipulated by a uniform monitor called the task manager. The process of task management is illustrated in a multiprocess story understander called a distributable script applier mechanism (DSAM), which reads and summarizes newspaper stories about plane crashes.