The problems of access -- retrieving linguistic structure from some mental grammar -- and disambiguation -- choosing among these structures to correctly parse ambiguous linguistic input -- are fundamental to language understanding. The literature abounds with psychological results on lexical access, the access of idioms, syntactic rule access, parsing preferences, syntactic disambiguation, and the processing of garden-path sentences. Unfortunately, it has been difficult to combine models which account for these results to build a general, uniform model of access and disambiguation at the lexical, idiomatic, and syntactic levels. For example psycholinguistic theories of lexical access and idiom access and parsing theories of syntactic rule access have almost no commonality in methodology or coverage of psycholinguistic data. This paper presents a single probabilistic algorithm which models both the access and disambiguation of linguistic knowledge. The algorithm is based on a parallel parser which ranks constructions for access, and interpretations for disambiguation, by their conditional probability. Low-ranked constructions and interpretations are pruned through beam-search; this pruning accounts, among other things, for the garden-path effect. I show that this motivated probabilistic treatment accounts for a wide variety of psycholinguistic results, arguing for a more uniform representation of linguistic knowledge and for the use of probabilisticallyenriched grammars and interpreters as models of human knowledge of and processing of language.

In order to learn effectively, a reasoner must not only possess knowledge about the world and be able to improve that knowledge, but it also must introspectively reason about how it performs a given task and what particular pieces of knowledge it needs to improve its performance at the current task. Introspection requires declarative representations of meta-knowledge of the reasoning performed by the system during the performance task, of the system's knowledge, and of the organization of this knowledge. This paper presents a taxonomy of possible reasoning failures that can occur during a performance task, declarative representations of these failures, and associations between failures and particular learning strategies. The theory is based on Meta-XPs, which are explanation structures that help the system identify failure types, formulate learning goals, and choose appropriate learning strategies in order to avoid similar mistakes in the future. The theory is implemented in a ...

Abduction is a form of non-monotonic reasoning that has gained increasing interest in the last few years. The key idea behind it can be represented by the following inference rule $$O = \mathop C\limits_| - N = \mathop P\limits_|^| - O - \mathop C\limits_|^| - .$$ i.e., from an occurrence of ohgr and the rule ldquophiv implies ohgrrdquo, infer an occurrence of phiv as aplausible hypothesis or explanation for ohgr. Thus, in contrast to deduction, abduction is as well as induction a form of ldquodefeasiblerdquo inference, i.e., the formulae sanctioned are plausible and submitted to verification. In this paper, a formal description of current approaches is given. The underlying reasoning process is treated independently and divided into two parts. This includes a description of methods for hypotheses generation and methods for finding the best explanations among a set of possible ones. Furthermore, the complexity of the abductive task is surveyed in connection with its relationship to default reasoning. We conclude with the presentation of applications of the discussed approaches focusing on plan recognition and plan generation.

We describe a method of automatically abducing qualitative models from descriptions of behaviors. We generate, from either quan titative or qualitative data, models in the form of qualitative differen tial equations suitable for use by QSIM. Constraints are generated and filtered both by comparison with the input behaviors and by dimensional analysis. If the user provides complete information on the input behaviors and the dimensions of the input variables, the resulting model is un ique, maximally constrained, and guaranteed to reproduce the input behaviors. If the user provides incomplete information , our method will still generate a model which reproduces the input behaviors, but the model may no longer be un ique. Incompleteness can take several forms: missing dimensions, values of variables, or en tire variables.

Many abductive understanding systems explain novel situations by a chaining process that is neutral to explainer needs beyond generating some plausible explanation for the event being explained. This paper examines the relationship of standard models of abductive understanding to the case-based explanation model. In case-based explanation, construction and selection of abductive hypotheses are focused by specific explanations of prior episodes and by goal-based criteria reflecting current information needs. The case-based method is inspired by observations of human explanation of anomalous events during everyday understanding, and this paper focuses on the method's contributions to the problems of building good explanations in everyday domains. We identify five central issues, compare how those issues are addressed in traditional and case-based explanation models, and discuss motivations for using the case-based approach to facilitate generation of plausible and useful explanations in...

This paper is concerned with information-seeking dialogues in a restricted domain (we consider a consultation system for a Computer Science Department, delivering information about the various tasks that the users may want to perform: for example, how to access the library, get information about the courses of the Department, etc) and presents a framework where a plan recognition and a user modeling component are integrated to cooperate in the task of identifying the user's plans and goals. The focus of the paper is centered on the techniques used for building the user model and exploiting it in the determination of the user's intentions. For this task, we use stereotypes and we propose some inference rules for expanding the user model by inferring the user's beliefs from both the sentences s/he utters and the information stored in the plan library of the system, that describes the actions in the domain. Moreover, we introduce some disambiguation rules that are applied to the informati...

In this paper, we argue that the resolution of anaphoric expressions in an utterance is essentially an abductive task following [HSAM93] who use a weighted abduction scheme on horn clauses to deal with reference. We give a semantic representation for utterances containing anaphora that enables us to compute possible antecedents by abductive inference. We extend the disjunctive model construction procedure of hyper tableaux [BFN96, Kuh97] with a clause transformation turning the abductive task into a model generation problem and show the completeness of this transformation with respect to the computation of abuctive explanations. This abductive inference is applied to the resolution of anaphoric expressions in our general model constructing framework for incremental discourse representation [Kuh99] which we argue to be useful for computing information updates from natural language utterances [Vel96].

Abduction is often viewed as inference to the "best" explanation. However, the evaluation of the goodness of candidate hypotheses remains an open problem. Most artificial intelligence research addressing this problem has concentrated on syntactic criteria, applied uniformly regardless of the explainer's intended use for the explanation. We demonstrate that syntactic approaches are insufficient to capture important differences in explanations, and propose instead that choice of the "best" explanation should be based on explanations' utility for the explainer 's purpose. We describe two classes of goals motivating explanation: knowledge goals reflecting internal desires for information, and goals to accomplish tasks in the external world. We describe how these goals impose requirements on explanations, and discuss how we apply those requirements to evaluate hypotheses in two computer story understanding systems. In order to learn from experience, a reasoner must be able to explain what...

This paper presents an algorithm for first-order Horn-clause abduction that uses an ATMS to avoid redundant computation. This algorithm is either more efficient or more general than any other previous abduction algorithm. Since computing all minimal abductive explanations is intractable, we also present a heuristic version of the algorithm that uses beam search to compute a subset of the simplest explanations. We present empirical results on a broad range of abduction problems from text understanding, plan recognition, and device diagnosis which demonstrate that our algorithm is at least an order of magnitude faster than an alternative abduction algorithm that does not use an ATMS.

While it has been realized for quite some time within AI that abduction is a general model of explanation for a variety of tasks, there have been no empirical investigations into the practical feasibility of a general, logic-based abductive approach to explanation. In this paper we present extensive empirical results on applying a general abductive system, Accel, to moderately complex problems in plan recognition and diagnosis. In plan recognition, Accel has been tested on 50 short narrative texts, inferring characters ' plans from actions described in a text. In medical diagnosis, Accel has diagnosed 50 real-world patient cases involving brain damage due to stroke (previously addressed by set-covering methods). Accel also uses abduction to accomplish model-based diagnosis of logic circuits (a full adder) and continuous dynamic systems (a temperature controller and the water balance system of the human kidney). The results indicate that general purpose abduction is an e ective and e cient mechanism for solving problems in plan recognition and diagnosis. 1