Most large text-understanding systems have been designed under the assumption that the input text will be in reasonably "neat" form (for example, newspaper stories and other edited texts). However, a great deal of natural language text (for example, memos, messages, rough drafts, conversation transcripts, etc.) have features that differ significantly from "neat" texts, posing special problems for readers, such as misspelled words, missing words, poor syntactic construction, unclear or ambiguous interpretation, missing crucial punctuation, etc. Our solution to these problems is to make use of expectations, based both on knowledge of surface English and on world knowledge of the situation being described. These syntactic and semantic expectations can be used to figure out unknown words from context, constrain the possible word senses of words with multiple meanings (ambiguity), fill in missing words (ellipsis), and resolve referents (anaphora). This method of using expectations to aid the understanding of "scruffy" texts has bee incorporated into a working computer program called NOMAD, which understands scruffy texts in the domain of Navy ship-to-shore messages.

This paper presents a process model of the acquisition and operation of early predictive behavior in young children, i.e., children's ability to accurately anticipate recurring sequences of events. The principal question that the model addresses is: how do children acquire predictive behavior from experience? The model presented here, called CEL (Components of Episodic Learning) provides an effective procedure for performing this acquisition process, and has been used as the basis for a prototype computer system running at the UCI Artificial Intelligence Project. The CEL model conforms to the constraints provided by relevant results in psychology and neurobiology; some observed stages of early child learning are explained in terms of the model, and theoretical lesions to specific parts of the model are used to predict particular behavior deficits that correspond well to documented defecits associated with lesions to the hippocampus of human patients.

Results in cognitive psychology have shown that readers can be steered away from an otherwise plausible interpretation of a story by extra-textual factors such as the source of the text, the stated reading purpose, interruptions and repetition of questions about the text. For instance, successive repetitions of the same question about a given text will often elicit a series of alternative interpretations of the text. This effect cannot be accounted for by established principles of text processing behavior, such as people's preference for cohesive and parsimonious representations of text. This paper presents a computer program called MACARTHUR, which models this behavior by varying the depth and direction of its inference pursuit in response to re-questioning, resulting in a series of markedly different interpretations of the same text. In light of the results, some new experiments are suggested in hopes of arriving at a new principle, beyond cohesion and parsimony, to account for the observed text processing behavior.

Most large text-understanding systems have been designed under the assumption that the input text will be in reasonably "neat" form,e.g., newspaper stories and other prepared texts, which typically consist of well-formed sentences and a logical order of presentation of concepts. However, many everyday uses of natural text (e.g., business memos, phone messages, notes) are very ill-informed, containing misspelled words, ungrammatical or only partially complete sentences, and poorly organized ideas.

Many large organizations employ humans whose sole job it is to take ill-formed messages and encode them into machine-readable form to be entered into a database. In such cases, there would be great potential benefit if that encoding process could be partially automated. This requires a "scruffy text understander"; i.e., a system that has the ability to correctly analyze the content of such texts in spite of their scruffiness.

This paper describes the design and implementation of the NOMAD system, which partially automates the encoding of poorly written Navy messages into well-formed formats. A number of problems are described which arise in the scruffy text domain that have not been dealt with in previous text-understanding systems.

A general purpose learning program is described which demonstrates a latent learning ability by operating at two separate goal pursuit levels. At one level are the constant, implicit goals associated with the system's memory management mechanisms. At the higher level are the dynamic, explicit behavioral goals which the implicit goals enable by manipulating memory representations to conform to the external surroundings. The program is shown to negotiate a simulated maze environment by the step-wise refinement of its latently learned experiences.

In the course of understanding a text, a succession of decision points arise at which readers are faced with the task of choosing among alternative possible interpretations ofthattext. We present new experimental evidence that different readers use different inference strategies to guide their inference behavior during understanding. The choices available to an understander range from various alternative inferential paths to the option of making no inference at a particular point, leaving a 'loose end'. Different inference strategies result in observably different behaviors during understanding, including consistent differences in reading times, and different interpretations of a text. The preliminary experimental results given here so far consistently support a previously published set of hypotheses about the inference process that we have called Judgmental Inference theory.

The CEL model of learning and memory (Components of Episodic Learning) [Granger 1982, 1983a, 1983b] provides a process model of certain aspects of learning and memory in animals and humans. The model consists of a set of asynchronous and semi-independent functional operators that collectively create and modify memory traces as a result of experience. The model conforms to relevant results in the learning literature of psychology and neurobiology. There are two goals to this work: one is to create a set of working learning systems that will improve their performance on the basis of experience, and the other is to compare these systems' performance with that of living systems, as a step towards the eventual comparative characterizations of different learning systems.

Parts of the model have been implemented in the CEL-0 program, which operates in a 'Maze-World' simulated maze environment. The program exhibits simple exploratory behavior that leads to the acquisition of predictive and discriminatory schemata. A number of interesting theoretical predictions have arisen in part from observation of the operation of the program, some of which are currently being tested in neurobiological experiments. In particular, some neurobiological evidence for the existence of multiple, seperable memory systems in humans and animals is interpreted in terms of the model, and some new experiments are suggested arising from the model's predictions.