According to the influential theory of (Katz and Fodor, 1964), a predicate associates a set of defining features with each argument, expressed within a restricted semantic vocabulary. Despite the persistence of this theory, however, there is widespread agreement about its empirical shortcomings (McCawley, 1968; Fodor, 1977). As an alternative, some critics of the Katz-Fodor theory (e.g. (Johnson-Laird, 1983)) have abandoned the treatment of selectional constraints as semantic, instead treating them as indistinguishable from inferences made on the basis of factual knowledge. This provides a better match for the empirical phenomena, but it opens up a different problem: if selectional constraints are the same as inferences in general, then accounting for them will require a much more complete understanding of knowledge representation and inference than we have at present.

The problem, then, is this: how can a theory of selectional constraints be elaborated without first having either an empirically adequate theory of defining features or a comprehensive theory of inference?

In this dissertation, I suggest that an answer to this question lies in the representation of conceptual knowledge. Following Miller (1990b), I adopt a “differential” approach to conceptual representation, in which a conceptual taxonomy is defined in terms of inferential relationships rather than definitional features. Crucially, however, the inferences underlying the stored knowledge are not made explicit. My hypothesis is that a theory of selectional constraints need make reference only to knowledge stored in such a taxonomy, without ever referring overtly to inferential processes. I propose such a theory, formalizing selectional relationships in probabilistic terms: the selectional behavior of a predicate is modeled as its distributional effect on the conceptual classes of its arguments. This is expressed using the information-theoretic measure of relative entropy (Kullback and Leibler, 1951), which leads to an illuminating interpretation of what selectional constraints are: the strength of a predicate’s selection for an argument is identified with the quantity of information it carries about that argument.

In addition to arguing that the model is empirically adequate, I explore its application to two problems. The first concerns a linguistic question: why some transitive verbs permit implicit direct objects (“John ate Ø”) and others do not (“*John brought Ø”). It has often been observed informally that the omission of objects is connected to the ease with which the object can be inferred. I have made this observation more formal by positing a relationship between selectional constraints and inferability. This predicts (i) that verbs permitting implicit objects select more strongly for (i.e. carry more information about) that argument than verbs that do not, and (ii) that strength of selection is a predictor of how often verbs omit their objects in naturally occurring utterances. Computational experiments confirm these predictions.

Second, I have explored the practical applications of the model in resolving syntactic ambiguity. A number of authors have recently begun investigating the use of corpus-based lexical statistics in automatic parsing; the results of computational experiments using the present model suggest that many lexical relationships are better viewed in terms of underlying conceptual relationships. Thus the information-theoretic measures proposed here can serve not only as components in a theory of selectional constraints, but also as tools for practical natural language processing.

TraumAID is a consultation system for the diagnosis and treatment of multiple trauma. It has been under development jointly at the University of Pennsylvania and the Medical College of Pennsylvania for the past eight years. TraumAID integrates diagnostic reasoning, planning and action. Its reasoner identifies diagnostic and therapeutic goals appropriate to the physician’s knowledge of the patient’s state, while its planner advises on beneficial actions to next perform. The physician’s lack of complete knowledge of the situation and the time limitations of emergency medicine constrain the ability of any planner to identify what would be the best thing to do. Nevertheless, TraumAID’s Progressive Horizon Planner has been designed to create a plan for patient care that is in keeping with the standards of managing trauma.

The data that support my claim are Purpose Clauses (PCs), infinitival constructions as in Do α to do β, and Negative Imperatives. I present a pragmatic analysis of both PCs and Negative Imperatives. Furthermore, I analyze the computational consequences of PCs, in terms of the relations between actions PCs express, and of the inferences an agent has to perform to understand PCs.

I propose an action representation formalism that provides the required flexibility. It has two components. The Terminological Box (TBox) encodes linguistic knowledge about actions, and is expressed by means of the hybrid system CLASSIC [Brachman et al., 1991].

To guarantee that the primitives of the representation are linguistically motivated, I derive them from Jackendoff's work on Conceptual Structures [1983; 1990]. The Action Library encodes planning knowledge about actions. The action terms used in the plans are those defined in the TBox.

Finally, I present an algorithm that implements inferences necessary to understand Do α to do β, and supported by the formalism I propose. In particular, I show how the TBox classifier is used to infer whether α can be assumed to match one of the substeps in the plan for β, and how expectations necessary for the match to hold are computed.

I argue that IRUs can only be explained by a processing model of dialogue that reflects agents' autonomy and limited attentional and inferential capacity. The central thesis is that the communicative function of IRUs is related to the cognitive properties of resource-limited agents. In order to investigate this interaction the thesis relies on two empirical methods: (1) distributional analysis of IRUs in a large corpus of naturally occurring dialogues, and (2) dialogue simulations in the Design-World environment which supports the parameterization of the dialogue situation, the task definition, and agents' cognitive properties.

The distributional analysis provides support for the claimed communicative functions by showing that IRUs demonstrate agents' autonomy and are used to support deliberation and inference. IRUs help autonomous agents coordinate on a collaborative task given their resource limits. The Design-World simulations show that discourse strategies that include IRUs are beneficial when agents are attention and inference limited or when the task is fault intolerant, inferentially complex, or requires a high degree of agreement. While some types of IRUs are beneficial simply as a rehearsal strategy, the general result is that IRUs are most beneficial when targeted at specific requirements of the communication situation.

A reasoning architecture is proposed in which specialized diagnostic reasoning and planning components are integrated in a cycle of reasoning and action/perception:

1. A Goal-Directed Diagnostic (GDD) reasoner which is predicated on the view that diagnosis is only worthwhile to the extent that it can affect repair decisions and that goals can be used to focus on such. Rather than focusing on a diagnosis object as the primary purpose of the diagnostic process, the GDD reasoner is tasked primarily with generating goals for the planner and with reasoning about whether these goals have been satisfied.

2. A Progressive Horizon Planner (PHP) which works by constructing intermediate plans via a combination of plan sketching and selection/optimization sub-processes, and then adapting these plans to reflect new information and goals. For the plan sketching sub-part, I propose a selection-and-ordering planning/scheduling paradigm, taking advantage of the limited interaction between goals.

I have implemented this architecture and reasoning components in TraumAID 2.0 - a consultation system for the trauma management domain. In a blinded comparison, out of 97 real trauma cases, three trauma surgeons have judged management plans proposed by TraumAID 2.0 preferable to the actual care by a ratio of 64:17 and to plans generated by its predecessor TraumAID 1.0 by a ratio of 62:9.

After a discussion of previous approaches to the informational articulation of the sentence, a hierarchical articulation is proposed: sentences are divided into the focus, which is the only information of the sentence, and the ground, which specifies how that information fits in the hearer's knowledge-store. The ground is further divided into the link, which denotes an address in the hearer's knowledge-store under which s/he is instructed to enter the information, and the tail, which provides further directions on how the information must be entered under a given address.

Empirical support for this representation of information packaging comes especially from the surface encoding of instructions in Catalan, which is then contrasted with that of English. Using a multistratal syntactic theory, it is then proposed that information packaging is structurally and purely represented at the abstract level of IS, which acts as an interface with informatics. Finally, in order to further argue for informatics as an autonomous linguistic component, some proposals that attempt to include informational notions under logical semantics are reviewed and countered.

This study is an effort to gain insight into one subdomain of pragmatics by integrating it into the larger process of language understanding. This is done by giving otherwise elusive informational notions a specific role in the component responsible for the entry of information into the hearer's knowledge-store.

Searching for things is a natural problem that arises when the blocks world assumptions (which have been the problem setting for most planning research) are modified by providing the agent only partial knowledge of his environment. Since the agent does not know the total world state, actions may appear to have nondeterministic effects. The significant aspects of the search problem which differ from previously studied planning problems are the acquisition of information and iteration of similar actions while exploring a search space.

Since introduction of the situation calculus [MH69], various systems have been proposed for representing and reasoning about actions which involve knowledge acquisition and iteration, including Moore's work on the interaction between knowledge and action [Moo80]. Such systems can be used to infer properties of plans which have already been constructed, but do not themselves construct plans for complex actions. My concern with searching has to do with a sense that Moore's knowledge preconditions are overly restrictive. Morgenstern [Mor88] examined ways to weaken knowledge preconditions for an individual agent by relying on the knowledge and abilities of other agents. Lesperance's research [Les91] on indexical knowledge is another way of weakening the knowledge preconditions. I am trying to reduce the amount of information an agent must know (provided he can search a known search space). If you dial the right combination to a safe it will open, whether or not you knew in advance that it was the right combination. Search is a way to guarantee you will eventually dial the right combination. So what I am exploring is how to systematically construct a search that will use available knowledge to accomplish something the agent does not currently know enough to do directly.

I claim it is possible for automated agents to engage in search behavior. Engaging in search behavior consists in recognizing the need for a search, constructing an effective plan, and then carrying out that plan. Expressing such a plan and reasoning about its effectiveness requires a representation language. I will select a representation language based on criteria derived from analyzing the search planning problem. Each of the three components of a system for engaging in search behavior will be designed and implemented to demonstrate that an automated agent can find things when he needs to.

Grammatical yet unprocessable sentences such as 1. are called 'garden-path sentences.' Their existence provides an opportunity to investigate the human sentence processing mechanism by studying how and when it fails. The aim of this thesis is to construct a computational model of language understanding which can predict processing difficulty. The data to be modeled are known examples of garden path and non-garden path sentences, and other results from psycholinguistics.

It is widely believed that there are two distinct loci of computation in sentence processing: syntactic parsing and semantic interpretation. One longstanding controversy is which of these two modules bears responsibility for the immediate resolution of ambiguity. My claim is that it is the latter, and that the syntactic processing module is a very simple device which blindly and faithfully constructs all possible analyses for the sentence up to the current point of processing. The interpretive module serves as a filter, occasionally discarding certain of these analyses which it deems less appropriate for the ongoing discourse than their competitors.

This document is divided into three parts. The first is introductory, and reviews a selection of proposals from the sentence processing literature. The second part explores a body of data which has been adduced in support of a theory of structural preferences - one that is inconsistent with the present claim. I show how the current proposal can be specified to account for the available data, and moreover to predict where structural preference theories will go wrong. The third part is a theoretical investigation of how well the proposed architecture can be realized using current conceptions of linguistic competence. In it, I present a parsing algorithm and a meaning-based ambiguity resolution method.