On the intersection of story understanding and learning

Abstract

Our theory of introspective multistrategy learning proposes that three transformations must occur to learn effectively from a performance failure in an intelligent system: failure explanation, learning goal specification, and learning-strategy construction. Likewise, our theory of story understanding proposes that the effective reader processes interesting input in three analogous phases when explaining an input: concept elaboration, question specification, and explanation-strategy construction. Moreover, in both learning and story understanding, the reasoner is more effective when using metacognitive knowledge. That is, an effective learner must be able to reason about reasoning failures, and an effective reader must be able to adequately monitor reading comprehension. We present a multistrategy framework whereby various reasoning methods can be applied to these cognitive tasks. The Meta-AQUA system is a multistrategy learner and story understanding system that operates in the domain of story understanding failures.

Postscript Version
PDF Version

Figure 1 shows a hierarchical decomposition of the relationships between problem solving, comprehension and learning. These reasoning processes share a number of intersecting characteristics. As indicated by the stripe-filled intersection on the left, learning can be thought of as a planning task. Cox and Ram (1995) discuss this analogy at length.^[1] This chapter examines the similarity between learning and story understanding as indicated by the filled intersection on the right in Figure 1.

Our theory of introspective multistrategy learning (IML) and story understanding is implemented in a computational system called Meta-AQUA (Cox, 1996b; Ram & Cox, 1994). Although Meta-AQUA is an integrated system, it is useful to distinguish between its performance task, the externally-observable task that the overall system carries out, and its learning task, the internal task that the system must carry out in order to improve its ability to execute the performance task. Meta-AQUA's performance task is story understanding. The task is to build a coherent conceptual interpretation of an input story in its foreground knowledge (FK). When the performance task fails, Meta-AQUA's learning task is to make changes to its background knowledge (BK)^[2] so that story understanding failures are not repeated when processing similar stories in the future. In IML theory, learning has three major subtasks:

• Failure Explanation (blame assignment): determining the underling cause of the failure,
• Learning Goal Specification: deciding what to learn in response to the failure, and
• Learning-Strategy Construction: deciding how to perform the necessary learning

As illustrated in Figure 2, blame assignment requires a system to circumscribe the source of reasoning failure. Deciding what to learn entails the explicit specification of desired changes to the BK in service of failure repair. Given such learning goals, the changes can be achieved by constructing a strategy or plan that achieves the learning specification. To generate the changes to the BK, then, the system need only execute the learning strategy.

Likewise, we view story understanding as the generation of changes to the story model in the FK in response to interesting input. In our theory, it also has three major subtasks:

• Concept Elaboration: determining the underling source of the interest,
• Question Specification: deciding what to ask in response to the interest, and
• Explanation-Strategy Construction: deciding how to perform the necessary explanation in order to answer the question.

Section 2 begins to describe our theory by presenting a generalized process model for multistrategy reasoning that applies to both problem-solving and comprehension tasks. Section 3 refines the process model specifically to comprehension tasks and then specializes it further to account for the task of story understanding. Section 4 develops a process model of learning that parallels the model of understanding. Section 5 then compares the model of understanding from Section 3 with the learning model of Section 4. The chapter concludes with a discussion in Section 6.

Although the cognitive science community has almost universally recognized representation as crucial to intelligent behavior, the issues of strategy selection and construction has received much less attention. The research that does exist often scopes the issue much narrower than did Newell and Simon (e.g., Brigham & Pressley, 1988; McDermott, 1988; Puerta, Egar, Tu, & Musen, 1992; Punch, Goel, and Brown 1996; Reder, 1987). An operational definition of the generalized reasoning task that subsumes both understanding and problem solving, however, can be cast in a multistrategy framework, assuming problem-solving goals and comprehension goals. Problem solving goals are typically specified as states in the world desired by the reasoner; whereas, comprehension goals are desires to understand an input (i.e., to relate the input to the knowledge the reasoner already possesses). Given such goals, both problem solving and comprehension can be operationalized as follows:

Given some input from the world (e.g., preprocessed perceptual input or text from a story) and a current context (including contextual goals and knowledge), if the input is anomalous, or otherwise interesting^[4], choose or construct a reasoning strategy with which to explain the input while, at the same time, furthering the goals.

The outermost level of computation focuses upon the choice (or construction) of a reasoning strategy, rather than the choice of a domain-specific solution operator. The outermost control is thus a second-order (executive) process at the meta level; the first-order explanation process is at the object level. This multi-level reasoning approach is reminiscent of the MOLGEN system (Stefik, 1981), in which a plane of reasoning exists in both the design plane (the reasoning task in MOLGEN's domain) and the meta-plane (the task of choosing an operator in the design plane). As a result of this division, to choose a reasoning strategy the system should understand and model its own first-order algorithms.

In our formulation, reasoning at the object level is a variant of a heuristic generate-and-test paradigm (Newell & Simon, 1976), with the enhancement of a front-end identification process to filter interesting input (see Figure 3). If no unusual input to the system exists, no significant resources will be expended on reasoning. Therefore, in the absence of interesting input, an understander will skim its data; a problem solver will simply act reactively or habitually. In such situations there is no great deliberation in pursuit of the contextual goals. With interesting input, however, a reasoner should construct and execute a strategy, thus generating some response that resolves the anomaly that sparked the interest. Subsequently, the result is verified by some means constructed by the reasoner. If the result is falsified, then the generation process begins anew.

Reasoning at the meta-level (i.e., multistrategy reasoning) concerns either choosing the right strategy from among alternatives or constructing a strategy by assembling a sequence of methods that together can accomplish a desired state. It does not matter whether the desired state is a solution to a problem-solving task, a state of understanding for a comprehension task, or a state of knowledge to be acquired or modified during a learned learning task. The framework persists in all three.

Consider the goal an art critic has when viewing a painting in a recently opened show. The critic wishes to achieve some internal mental state that relates the symbols and images in the painting to the current understanding of the genre, thus enabling an evaluation of the object. That is, the painting must be interpreted with respect to information already present in the critic's BK. A mental comparison is made between what the critic expects of such paintings with the images and emotions actually invoked by the current painting. Note that a surprising or unexpected image may be interpreted either as an exemplar of a new, creative category or as a discordant failure. Both judgements are with respect to what the critic has previously experienced, but in either case, unusual objects that violate the agent's expectations are the ones that garner the most attention because they are interesting.

Figure 4 provides a more detailed specification of this understanding process. Given some input and a current context (including a comprehension goal, the system's BK, and within the FK1, a current model of the previous input), if the input is interesting, choose or construct a strategy with which to explain the input, otherwise incorporate the input into FK1. Upon execution of the explanation strategy, output a new representation (FK2) of the input that has no anomaly and is coherent with respect to the BK. The input is understood given that it remains consistent and coherent in the face of future input. Also output a representational trace of the reasoning that produced the understanding.

The explanation of interesting input should further the overall goal of understanding the entire story. The explanation is a good one if it helps to incorporate the new input with previous input and it needs little or no re-explanation when given further input concerning the same topic. The explanation is also good if it addresses the particular features that made it interesting to begin with (Ram, 1989; Ram & Leake, 1991).

Although not all understanding goals of are as specific as those of the art critic (that is, the need for a critical judgement is not always present), the general process outlined above conforms to the constraints of many comprehension tasks, including the task of reading a story.

As an example of the story understanding task, Meta-AQUA might process a story about a polite, Memphis musician named Elvis boarding with a young, Southern family (see Figure 5^[5]). While processing the story, Meta-AQUA constructs a model of the characters and the actions involved in the story. When the story reveals that Elvis occasionally smokes ganja (marijuana) in the house, endangering his safety and freedom, as well as that of the family's with which he lives, the system detects an anomaly that must be explained to fully understand the story. The event is anomalous (and hence interesting) because the model of Elvis constructed before the point of his taking drugs was one of a law-abiding citizen. A conflict occurs as a result of trying to unify the picture of Elvis as a typical, adult male (assumed to be happy) with the picture of him as an individual likely to commit a crime (thus, apt to be desperate).

To explain the incongruity, the system must understand the anomaly. Meta-AQUA accomplishes this by consulting a decision model (Ram, 1990a) that describes the planning process an agent such as Elvis performs when considering a choice of actions in the world. The objective of the analysis is to refine the nature of the anomaly and to identify the parts of the story that bear on the anomaly, so as to more clearly ascertain what needs to be explained to resolve the anomaly. An analysis of the story yields the facts that Elvis is not desperate, yet at the same time he performs an act that threatens the loss of his liberty. This situation is certainly anomalous because the decision model asserts that people value the goal of preserving their own freedom above most other goals they possess, other than the goal of preserving their lives. A goal competition (Wilensky, 1983) therefore exists that Meta-AQUA must explain.

Subsequently, Meta-AQUA poses a series of questions about the anomaly and the context of the story surrounding the anomaly. In this case, the system asks what would cause a man to carry out an action he knew could result in his own arrest. If this question can be answered, then the anomaly would likely be resolved, and the story would be considered understood.

To explain events in a story, Meta-AQUA can generate two types of explanations.^[6] Physical explanations give a causal account of events according to a model of the way things work in the world, whereas volitional explanations give a causal account of why people perform the acts they do in the world (Ram, 1990a).^[7] The former class links physical events (such as the burning of flammable materials) with probable causes (such as the lighting of materials with combustible devices). The latter type of explanation links the actions of agents in a story to their goals and beliefs, thus providing a motivation for story characters. In the Elvis scenario, Meta-AQUA retrieves, instantiates, and adapts a cigarette-smoking explanation, which produces expectations in the story (e.g., that the smoking will relieve a nervous emotional state). It can either look for verification of the explanation by tying it into the story, or it can suspend the explanation until a later point in time. The explanation can be verified when subsequent sentences in the story confirm the hypothesis.

3.2.1. Interest identification

The first step the system performs is a simple interest detection. As previously mentioned, a concept is interesting if it is anomalous, intrinsically interesting, or if it is a concept about which something has recently been learned. In the first case, an anomaly is signaled when either the input conflicts with known facts in the BK, or when the system is otherwise unable to successfully incorporate the representation of the input into the current story model in the FK. This is often detected by a unification mismatch during story processing. When a new instance is input from a story, the conceptual frame is unified with a story template or schema from the BK. If unification fails, a mismatch has occurred, and a pointer to the location of the mismatch is returned as the paths value of the anomaly (see Figure 7). For example, the act of Elvis smoking pot does not unify with a pipe-smoking script because the value of ganja1 does not match the tobacco constraint on the object role-filler of the script (see Figure 8).^[10]

In the second case, an input is determined as interesting if it is inherently interesting; that is, it is interesting if it pertains to the intrinsic goals of the reasoner. Intrinsic, or innate, goals such as the desire to maintain a state of personal safety, are associated with loud noises and violent actions, for example. In a simple way, then, Meta-AQUA, categorizes as intrinsically interesting, any concept that inherits features from among the following: loud-action, violent-action, and sexual action.

Finally in the third case, when Meta-AQUA has performed some learning on a particular class of objects or actions, it assigns that conceptual type an ``interestingness marker.'' Therefore, when Meta-AQUA encounters new input pertaining to that concept it will again be considered interesting and receive closer processing. Such an approach allows the system to form hypotheses in one story and verify it in another. The interestingness marker is amortized across time, so that after repeated encounters with the concept, the reader will no longer exhibit interest in the subject.

3.2.2. Explanation generation

Once an input is determined to be interesting, an explanation process attempts to resolve the anomaly by constructing a causal account of the input with respect to both the story and the reader's knowledge. Given some anomalous state the reader encounters, if the reader is to fully understand the story, the following questions must be answered:

• How did the anomaly occur?
• What needs to be explained?
• How can I explain this?

Subsequently it will:
• Resolve the anomaly by generating an explanation.

Concept elaboration. The initial step is to elaborate the anomaly in order to provide a relevant context for determining what occurred within the story. The reasoner refines the anomaly in such a way that a specific question can be posed. Since the specification of the explanation process must be more precise than simply ``explain the anomaly,'' simply asking what the reason is for the anomaly adds little benefit. Although it may be clear that some representation for a character like Elvis indicates that he isa typical-person.0, that a later representation of him isa criminal-person.0, and that the two representations will not unify in the program internals, a better characterization of the anomaly provides specific circumstances (including motivations, states, goals, and beliefs) in terms of both a model of normative decisions and a model of the current story that point to possible locations of the anomaly. Moreover, by providing a story context, a system avoids much search, since the context should contain only the pertinent details known so far. A talented programmer can set up the anomalies that its system knows about in such a way that resolution is all but guaranteed. It is better to have some process that attempts to focus the anomaly so that conditions not envisioned by the programmer can also be addressed.

Question specification. Given the context provided by the previous step, the function of the next step is to provide a set of questions that represents gaps in the model of the story with respect to the anomaly. Any such question can be viewed as a knowledge goal (Cox & Ram, 1995; Ram, 1991; Ram & Hunter, 1992), since it specifies the knowledge states that, if achieved, would provide coherence to both the story and what the system knows (its BK). The function of such knowledge goals, is to focus the resources and processing of the reasoner so that the combinatoric explosion of inferences is mitigated. For example by asking the question ``Why did Elvis smoke ganja in the pipe?'' the reader of the Elvis episode will concentrate inference upon the problem that is most relevant in the story. Without a causal explanation to the question, the story will be only partially understood.

Explanation-strategy construction. Following this specification the system can pick an explanation method that will answer the focal questions (i.e., achieve the knowledge goals). Depending upon the given situation and the organization of the BK (i.e., how memory is indexed), a system may choose from case-based reasoning (CBR), analogy, explanation application, or any number of reasoning methods for generation. For example, if a reader is reminded of prior case, CBR may be used; whereas, if the reader is reminded of an old explanation pattern (XP), explanation application may be used. Figure 9 shows an example explanation strategy using XP Application (Ram, 1991; Schank, 1986).^[11] Once a strategy is determined, the program can generate the explanation by executing the strategy.

3.2.3. Hypothesis verification

The resulting hypothesis is then tested for degree of fit or believability. To verify the hypothesized explanation, the verification process makes a similar four-step analysis. The first step, however, that of finding the source of the hypothesis, is known to follow from the generation process.^[12] Step two is to determine whether to attempt to prove or disprove the hypothesis. Given a target approach, the system then needs to choose an algorithm best suited to achieving the goal. To perform a test of the resulting hypothesis, a reasoner may devise an experiment, ask someone, or simply wait, in the hope that the answer will be provided by future input. Once the algorithms have been selected and ordered, the hypothesis can then be evaluated.

Assuming such a model for the story-understanding performance task, traces of system performance can be specified and recorded at run-time in declarative structures. These knowledge structures are used by learning mechanisms to reason about processing failures, if and when failure occurs. A trace contains a decide-compute node (D-C-NODE) for each of the sub-processes of an understanding task; that is, it records the decision and the reasons behind each decision in every step of Figure 6. Both the generation and verification processes have four steps each of which correspond to a process field in a D-C-NODE. The four fields are input analysis, goal specification, strategy decision, and strategy execution. For each field, the record stores both the enabling conditions and the resulting state. For the first three fields, the D-C-NODE records the decision basis, and for the last field, it records the side-effects of the process.

If a failure occurs (as detected by the algorithm to be presented in the forthcoming section), the system suspends the understanding performance-task and invokes the learning task. When this happens, the trace of the reasoning along with a characterization of the failure (as determined by the failure detection algorithm) is passed to the learning process for introspective explanation. When learning abates, the system resumes the story-understanding performance task.

Simon (1983) defines learning as ``changes in the system that are adaptive in the sense that they enable the system to do the same task or tasks drawn from the same population more efficiently and more effectively the next time'' (p. 28). Thus, some performance task exists that receives an input and acts upon it given its knowledge dealing with that class of data. A measure of this performance is then passed to a learning task, whereupon it makes changes to the knowledge used by the performance system, depending on the success or failure of the performance. This general view of learning is diagrammed in Figure 10.

For instance, students often learn to program computers in LISP when previously knowing another language such as Pascal. But as LISP novices, the code that results from their problem solving is usually overly-extenuated, inefficient, buggy, and written in an imperative style with loops and block control-structures. As students learn to debug their programs better and acquire mastery of more LISP functions, the code becomes much more compact, efficient, bug-free, and written recursively within a functional programming style. The difference in performance is due to a change in the knowledge and skills used by the programmer both to understand and solve problems and to implement the resulting solutions. These conceptual changes come about from a removal of rigid, Pascal-like coding habits, an acquisition of new LISP techniques, and a reorganization of the applicability conditions for much of the knowledge relevant to the task of computer programming.

In contrast to Simon's definition, the Inferential Learning Theory of Michalski (1991, 1994)^[13] defines a learning task as consisting of three components: some input (information), the BK, and a learning goal. Even though this description does not explicitly refer to the performance of a reasoning system, and so differs from IML theory, the concept of a learning goal is central to both Michalski's model and the model of learning presented here. The learning goal determines the relevant pieces of the input, the knowledge to be acquired, and the criteria for evaluating the learning. The model of learning presented here is consistent with these constraints, and, as championed by Michalski, concentrates on a multistrategy approach to learning whereby more than one learning strategy can be brought to bear upon a given learning task. Because the multistrategy approach applies equally well to both reasoning (in the form of either problem-solving or understanding) and to learning, this framework is a natural one for integrating the learning and the performance tasks.

Approaches to multistrategy learning fall into three broad categories, which we call strategy selection models, toolbox models, and cascade models. The common element in all these approaches is the use of multiple learning methods to allow the reasoning system to learn in multiple types of learning situations. In strategy selection models, the reasoner has access to several learning strategies, each represented as a separate algorithm or method. Learning involves an explicit decision stage in which the appropriate learning strategy is identified, followed by a strategy application stage in which the corresponding algorithm is executed. Methods for strategy selection also differ. The Meta-AQUA system uses characterizations of reasoning failures to determine what to learn and, in turn, the learning strategies to use when building a learning plan. Toolbox models are similar to strategy selection models in that they too incorporate several learning strategies in a single system. The difference is that these strategies are viewed as tools that can be invoked by the user to perform different types of learning. The tools themselves are available for use by other tools; thus, learning strategies may be organized as co-routines. In cascade models, two or more learning strategies are cascaded sequentially, with the output of one strategy serving as the input to another. Clearly, these categories of models are not exclusive of each other (e.g., a strategy selection system may choose to cascade learning strategies in certain circumstances), but they serve to characterize the major ways in which learning strategies may be integrated.

Research into multistrategy learning is useful on pragmatic grounds when complex worlds are the domains of learning systems. Such approaches allow for maximal flexibility. Significant interactions are present in multistrategy systems, however, that are not apparent in isolated systems. For example, if two algorithms modify the domain knowledge of the system, and a dependency exists between the two, such that one strategy modifies a part of the domain knowledge that the second one uses, then an implied sequencing must be enforced; that is, the first strategy must be applied before the second. Such dependencies do not exist in single-strategy systems.

The general model of learning from Figure 10 can be refined to a multistrategy framework as seen in Figure 11. The problem generation module outputs a story to the story-understanding performance system with the initial goal to understand the input. The performance module uses schemas from the BK to explain the story and to build a representation for it in the FK. If this task fails, then a trace of the reasoning that preceded the failure is passed to the learning subsystem.

A CBR subsystem within the learner uses past cases of introspective reasoning from the BK to explain the comprehension failure and to generate a set of learning goals. These goals, along with the trace, are then passed to a nonlinear planner. The planner subsequently builds a learning strategy from its toolbox of learning methods. The learning plan is passed to an execution system that examines and changes items in the BK. These changes enable improved future performance on the performance task (i.e., story understanding). Although Meta-AQUA's algorithms and knowledge structures have been reported in detail elsewhere (e.g., Cox, 1994, 1996b; Cox & Ram, 1995; Ram & Cox, 1994; Ram, Cox & Narayanan, 1995), the following two sections provide a short example and an outline of the learning algorithm in order to provide context for the comparison between learning and story understanding.

Figure 12 illustrates a short story generated by Tale-Spin and input to the story-understanding module of Meta-AQUA. In the story, Meta-AQUA finds it unusual for Lynn to strike a ball because the program's conceptual definition of the ``hit'' predicate constrains the object attribute to animate objects. It tries to explain the action by presupposing that Lynn tried to hurt the ball (a volitional explanation pattern, or XP, retrieved from the BK instantiates this hypothesis). In a following sentence, however, the story provides an alternate explanation (i.e., the hit action is intended to move the ball to the opposing person). This input causes an expectation failure because the system had expected one explanation to be true, but another proved true instead.

When the Meta-AQUA system detects an explanation failure, the performance module passes a trace of the reasoning to the learning subsystem. At this time, the learner needs to explain why the failure occurred (assign blame) by applying an introspective explanation to the trace. A meta-explanation pattern (Meta-XP)^[15] is retrieved using the failure symptom as a probe into memory. Meta-AQUA instantiates the retrieved meta-explanation and binds it to the trace of reasoning that preceded the failure. The resulting structure is then checked for applicability. If the Meta-XP does not apply correctly, then another probe is attempted. An accepted Meta-XP either provides a set of learning goals (determines what to learn) that are designed to modify the system's BK or generates additional questions to be posed about the failure. Once a set of learning goals are posted, they are passed to the nonlinear planner for building a learning plan (strategy construction).

Figure 13 lists the major state transitions that the three learning processes produce. The learning plan is fully ordered to avoid interactions. For example, the abstraction step must precede the other steps because a knowledge dependency exists between the changes on the hit concept as a result of the abstraction step and the use of the hit concept by both the generalization and indexing steps.^[16] After the learning is executed and control returns to sentence processing, subsequent sentences concerning the hit predicate causes no anomaly. Instead, Meta-AQUA predicts the proper explanation when Elvis hit the ball.

The first process performs failure detection. Five types of failures can occur. Failure detection inputs two structures (an expected outcome, E, and the actual outcome, A) and the trace of the reasoning producing these knowledge structures. The algorithm for this process is shown in Figure 14.^[17] The detection process occurs either during the verification phase of the performance task of the system or during the generation phase after a resumption of a suspended generation goal. This second condition occurs after the performance system previously tried to generate a hypothesis, but could not. The generation phase suspends the goal and new input later provides the answer. See impasse condition in Figure 14. Along with the trace, the process outputs a determination of which of the failures exist (if any) to the next phase. While reading the story from Figure 12, the detection process returns a contradiction between the input instance of Lynn hitting the ball and the conceptual definition of hit in the BK, and another contradiction between the expected explanation for this event and the one provided by the story.

The second phase concerns the actual determination of the causes of failure and the construction of a learning strategy which is then executed. Figure 15 defines this learning task and shows the overall information flow to and from the learning process. The strategies from which it may construct a learning plan is dependent upon the Meta-XP structures in memory. Although this phase will be discussed in some detail by the next sections, alternate strategies that may result include combinations of fine-grained knowledge transmutations or more global approaches such as a student's strategy of re-reading instructions when all else fails. The output of the phase is an implicit hypothesis that the learning was correct along with an augmented trace. The changes to the BK from learning are attached to a set of D-C-Nodes and are indexed in memory where the changes occur.

The third phase concerns verification. Although beyond the scope of this chapter and more suitable for future research, verifying the learning could involve either of two strategies. The system could be reminded of a change to the BK (as associated with the D-C-Nodes and described above) at some future time when the changed knowledge is reused. The learning can then be checked as to whether it is effective. Alternatively, the system could actually make a deliberate test of the newly learned knowledge by trying to falsify the information. When either of these processes finish, the verification phase would output an evaluation of the quality of learning.

The most critical of the three phases above, and the one upon which we place the most emphasis, is the second phase that generates changes to the BK. The remainder of this section offers additional details concerning its decomposition. Ram and Cox (1994) have argued that three fundamental learning-processes must be performed if learning is to be effective in an open world where many sources of failure exist. The processes are referred to as blame assignment (Birnbaum, Collins, Freed, & Krulwich, 1990; Minsky, 1961/1963; Stroulia, Shankar, Goel & Penberthy, 1992; Weintraub, 1991), deciding what to learn (Cox & Ram, 1995; Hunter, 1989, 1990; Keller, 1986; Krulwich, 1991; Leake & Ram, 1993; Ram & Hunter, 1992; Ram & Leake, 1991, 1995), and learning-strategy construction (Cox & Ram, 1991; Ram & Cox, 1994; Michalski, 1991). In the event of a performance failure, these processes answer the following three questions:^[18]

• How did the failure occur?
• What must be learned?
• How can this be learned?

• Repair the background knowledge.

To justify our process decomposition that answers these three questions, we advance the following argument: To construct a strategy, a system needs to know what is supposed to be learned; to decide what needs to be learned, it must know the cause of failure; to determine the cause of the failure, it must perform blame assignment; and to perform complete blame assignment in many situations, it must reflect upon its own reasoning. The subsections to follow presents an overview of the algorithm that instantiates these processes and Figure 16 sketches it in brief. The system records a trace of the reasoning used in the performance task in a number of trace meta-explanation structures. Each trace is inspected to detect a failure. When the system detects a failure, it invokes learning. During learning, the system constructs a learning strategy via the three process steps: blame assignment, deciding what to learn, and strategy construction. Subsequently, the system executes the learning strategy to perform the necessary knowledge repairs.

4.3.1. Blame assignment (step 2a, Figure 16)

Take as input a trace of the mental and physical events that preceded a reasoning failure; produce as output an explanation of how and why the failure occurred, in terms of the causal factors responsible for the failure.

Blame assignment is a matter of determining what was responsible for a given failure. Thus, the function of blame assignment is to identify which causal factors could have led to the reasoning failure as determined from the output of the performance task and contained in the reasoning trace. That is, blame assignment is like troubleshooting; it is a mapping function from failure symptom to failure cause. The purpose is the same whether the troubleshooter is explaining a broken device or itself (Stroulia, 1994).

The input trace describes how results or conclusions were produced by specifying the prior causal chain (both of mental and physical states and events). The learner retrieves an abstract meta-explanation pattern, or Meta-XP, from memory and applies it to the trace in order to produce a specific description of why these conclusions were wrong or inappropriate. This instantiation specifies the causal links that would have been responsible for a correct conclusion, and enumerates the difference between the two chains and two conclusions (what was produced and what should have been produced). Finally, the learner outputs the instantiated explanation(s). The Meta-XP that explains the symptoms of our earlier story asserts that the expectation-failure between competing explanations is due to three factors: incorrect domain knowledge (overly restrictive definition of the hit predicate), novel situation (never encountered the explanation that people hit objects to make them change locations), and erroneous association (the hurt explanation was associated with the actor that performed the hitting rather than the object hit). See again Figure 13.

4.3.2. Deciding what to learn (step 2b, Figure 16)

Take as input a causal explanation of how and why failure occurred; generate as output a set of learning goals which, if achieved, can reduce the likelihood of the failure repeating. Include with the output, both tentative goal-dependencies and priority orderings on the goals.

The previously instantiated Meta-XP assists in this process by specifying points in the reasoning trace most likely to be responsible for the failure. The Meta-XP also specifies the suggested type of learning goal to be spawned by this stage. Because these goals are tentative, it may be necessary to retract, decompose, or otherwise adapt the learning goals dynamically during run-time. This stage of learning mediates between the case-based approach of blame assignment and the non-linear planning approach of strategy construction. The learner includes with the output learning goals both tentative goal-dependencies and priority orderings on the goals. The trace is passed as output as well.

In the case of the hit anomaly, the Meta-XP focuses the learner on two learning goals. The story instance of hitting needs to be reconciled with the hit definition and the two explanations need to be differentiated. These two learning goals determine the kinds of changes to the BK necessary for effective learning. They operationalize the effects that matter and localize the inferences required for instantiating the changes.

4.3.3. Learning-strategy construction (step 2c, Figure 16)

Take as input a trace of how and why a failure occurred and a set of learning goals along with their dependencies; produce as output an ordered set of learning strategies to apply that will accomplish the goals along with updated dependencies on the set of goals.

The final learning-strategies are organized as plans to accomplish the learning goals. The plans are sequences of steps representing calls to standard learning algorithms. The plans are created by a Common LISP version of Tate's (1976) Nonlin planner (Ghosh, Hendler, Kambhampati, & Kettler, 1992). In order to use the nonlinear planner, the learning module translates the learning goals and the relevant context of the program environment to a predicate representation. In this form, Nonlin assembles a learning plan just as if it were creating a plan to stack a series of labeled blocks. The only difference is that the planner is given a set of learning operators that describe actions that modify the mental world (i.e., the BK) instead of the blocks world (see Cox & Ram, 1995 for full details of the planning analogy and implementation used in strategy construction).

The learner instantiates the plan, translates it back into a frame representation, and, then executes the learning plans (in step 2d, Figure 16). In the case of Lynn and Elvis' game of ball, the reconciliation goal is achieved by performing abstraction on the concept of hit. Thus the constraint on the object slot is raised to physical object (the parent of animate and inanimate), rather than being limited to animate ones. The goal of differentiating the two explanations is solved by running EBG on the new explanation, indexing the new explanation, and then reindexing the two explanation with respect to each other to achieve discrimination. Nonlin generates a full order for the plan steps that achieves the conjunctive learning goals and avoids goal interactions. At the termination of the plan execution, control is returned to the performance system and story understanding is resumed.

In an empirical study of the effects of this learning method on explanation performance, Cox, (1996a) demonstrates that Meta-AQUA performs significantly better under a fully introspective mode than under a reflexive mode that ablates learning goals. Figure 17 shows Meta-AQUA's cumulative improvement in performance (as measured by an evaluation function rating question answering ability) given one 24-story sequence of problems randomly generated by the Tale-Spin program. The lower curve represents Meta-AQUA without learning, the middle curve represents the system without learning goals, and the upper curve includes all three transformations.

To serve as experimental trials and to minimize order effects, Tale-Spin generated six such random sequences of stories. On each of these runs, Meta-AQUA processed a sequence three times, once for each experimental manipulation. The system begins all runs with the same initial conditions. For a given experimental condition, it processed all of the stories in the sequence while maintaining the learned knowledge between stories. At the end of the sequence, the system resets the BK. The input size for a run varies in length, but averages 27.67 stories per run. The corpus for the six runs included 166 stories, comprising a total of 4,884 sentences. The stories varied in size depending on the actions of the story and Tale-Spin's randomness parameters, but averaged 29.42 sentences.

Across all six experimental runs, the expected gain in learning (i.e., the differential between the average ``learning goal'' improvement of 102.7 and the average ``no learning goal'' improvement of 65.7) is a 56.38 percent difference. That is, across a number of input conditions, the use of learning goals to order and combine learning choices should show about 1.5 times the improvement in performance than will a straight mapping of faults to repairs when interactions are present. In general, these data lead to the conclusion that the process which posts learning goals (deciding what to learn) is a necessary transformation if negative interactions between learning methods are to be avoided and if learning is to remain effective. Moreover, we showed that because learning algorithms can negatively interact, the arbitrary ordering of learning methods can actually lead to worse system performance than no learning at all.

Both the form and the function of the generation phases in learning and understanding are similar (see Figure 18). The structure of both is to take some unusual input (reasoning failure or incongruous story concept), elaborate the input, generate some goal that provides focus for the process, then change some knowledge base to achieve the function of the process. Changes during story understanding take place in the FK, whereas changes during learning take place in the BK.

A number of differences, however, exist between learning and understanding. For example, as understanding can be likened to recovery, so too, learning can be likened to repair. In the planning literature a number of researchers have made the distinction between recovery and repair (see for example, Owens, 1991 and Hammond, 1989). When a plan fails, the planner must recover from the error so additional progress can be made toward the goal. After recovery, the plan needs to be repaired and stored again in memory, so that the plan failure will not recur.

For example, if an autonomous robot vehicle finds an expected fuel cache missing and thereby runs out of gasoline, it must first recover from the potentially threatening situation by obtaining fuel (example taken from Owens, 1991). Therefore, the explanation of the failure will dictate the means of recovery. If the robot concludes that it cannot find the gasoline because it is lost, then it should recover by obtaining orientation information; whereas, if it explains the fuel's absence because of theft, then the recovery taken will involve turning back or calling for assistance. The repair (to adjust its plans and the information upon which the plan was based) also follows from the explanation of the failure. For instance, if the robot previously considered taking on extra fuel, but did not because it assumed that the fuel cache would be at the proper location and easy to find, then this explanation of its decision would lead the system to modify its knowledge concerning the persistence of fuel caches. This modification would bias it toward conservative decisions in the future, and thus make it less likely to repeat the failure.

The difference between recovery and repair can be applied to the processes of understanding and learning in an analogous manner. The understanding process requires a recovery phase when it fails. If some explanation does not work, first there is a need to create a new explanation or somehow to seek one out. Once the correct (or more useful) explanation has been derived, the system needs to learn from the experience by repairing its knowledge, so as not to repeat the failure. Thus, as seen in Figure 18, the understanding process operates on the FK to instill the change that removes the anomaly (thus constituting the recovery); whereas the learning process operates on the BK, producing a repaired knowledge base with which the failure will not be as likely in future similar situations. The recovery is a system's response to anomalous input from the outside world that its knowledge could not adequately understand, whereas the learning is a response to the inadequacy in the reasoner's world.

Functional reasons exist for having an explicit input analysis stage in both learning and understanding (i.e., concept elaboration, or more specifically anomaly elaboration, in understanding and failure explanation, or more specifically blame assignment, in learning). Most programs accept cleanly defined problems as input, such that little ambiguity and sharp distinctions concerning what needs to be done exist. In more practical systems, problem elaboration is necessary to clarify what may actually be ill-defined tasks. For example, planning tasks in artificial intelligence are often structured and circumscribed by the programmer or user, not the planner. A planner may be given operational goal specifications from which a state, such as one block being on top of another, may be achieved. The tasks of recognizing that a problem exists for which a plan is required and establishing the goal specifications, however, are not considered a part of the planner's reasoning process. In comprehension tasks such as story understanding, the problems are not usually so well defined. In learning, the problem of recovery is to modify the story representation in such a way that the anomaly is coherent with respect to the rest of the story and the system's BK. This specification is so broad that either the programmer must include the specifications implicitly, or the explanation must be somewhat trivial. To narrow the range of behaviors appropriate for recovery, then, is to elaborate the input anomaly, so as to identify what went wrong and why.^[19]

Although the relationship between text comprehension and metacognitive activities has been studied since the turn of the century albeit under the guise of differing technical terms (Brown, 1987), more recent research examining the relationship between metacognitive skills and educational instruction have made significant progress. For example, Forrest-Pressley, MacKinnon, and Waller (1985) and Garner (1987) report successful instruction procedures related to both problem solving and reading comprehension (see also Ram & Leake, 1995, for a related discussion). Chi (1995, Chi, Bassok, Lewis, Reimann, & Glasser, 1989) reports that improved learning is correlated with human subjects who generate their own questions and explain the answers themselves (see also Pressley & Forrest-Pressley, 1985). This is the so called self-explanation effect. Thus, the ability of a system to pose self-generated questions both indexes actual understanding and simultaneously reduces the probability of asking only the easy questions.

Consider the following quote from Gavelek & Raphael (1985).

One form of metacognition - metacomprehension - addresses the abilities of individuals to adjust their cognitive activity in order to promote more effective comprehension. We have been interested in a specific aspect of metacomprehension - namely, the manner in which questions generated by sources external to the learner (i.e., from the teacher or text), as well as those questions generated by the learners themselves, serve to promote their comprehension of text. (p. 104)

The ability to adjust cognition in order to improve comprehension is at the heart of the research presented here. Thus, simply being able to recognize that a gap exists in one's own knowledge, and to therefore ask the question ``Why don't I understand this?'' (Ram, 1991), is the first step to improving the understanding, rather than actually giving an answer. So, to evaluate the ability of the performance of the Meta-AQUA system, credit should be given for simply posing a question that deserves asking.

Garner (1987) has argued that metacognition and comprehension monitoring are important factors in the understanding of written text. Reading comprehension is therefore considered to be chiefly an interaction between a reader's expectations and the textual information.^[20] Psychological studies have also confirmed a positive correlation between meta-memory and memory performance in cognitive monitoring situations (Miner & Reder, 1994; Schneider, 1985; Wellman, 1983) and between the use of metacognitive abilities with standard measures of intelligence (Davidson, Deuser, and Sternberg, 1994). This evidence, along with results from the studies above linking problem-solving performance with metacognitive abilities, directly supports the conviction that there must be a second-order introspective process that reflects to some degree on the performance element in an intelligent system, especially a learning system involved in tasks such as story understanding.

1. Anderson, J. R. (1983). The architecture of cognition. Cambridge, MA: Harvard University Press.

2. Anderson, J. R. (1993). Rules of the mind. Hillsdale, NJ: Lawrence Erlbaum Associates.

3. Barsalou, L. W. (1995). Storage side effects: Studying processing to understand learning. In A. Ram & D. B. Leake (Eds.), Goal-driven learning (pp. 407-419). Cambridge, MA: MIT Press/Bradford Books.

4. Birnbaum, L. (1986). Integrated processing in planning and understanding (Tech. Rep. No. 489). Doctoral dissertation, Yale University, Department of Computer Science, New Haven, CT.

5. Birnbaum, L., Collins, G., Freed, M., & Krulwich, B. (1990). Model-based diagnosis of planning failures. In Proceedings of the Eighth National Conference on Artificial Intelligence (pp. 318-323). Menlo Park, CA: AAAI Press.

6. Brigham, M. C., & Pressley, M. (1988). Cognitive monitoring and strategy choice in younger and older adults. Psychology and Aging, 3(3), 249-257.

7. Brown, A. (1987). Metacognition, executive control, self-regulation, and other more mysterious mechanisms. In F. E. Weinert & R. H. Kluwe (Eds.), Metacognition, motivation, and understanding (pp. 65-116). Hillsdale, NJ: Lawrence Erlbaum Associates.

8. Carbonell, J. G., Knoblock, C. A., & Minton, S. (1991). PRODIGY: An integrated architecture for planning and learning. In K. VanLehn (Ed.), Architectures for intelligence: The 22nd Carnegie Mellon symposium on cognition (pp. 241-278). Hillsdale, NJ: Lawrence Erlbaum Associates.

9. Chi, M. T. H. (1995). Revising the mental model as one learns. Plenary address to the Seventeenth Annual Conference of the Cognitive Science Society. Pittsburgh (July 23).

10. Chi, M. T. H., Bassok, M., Lewis, M., Reimann, P., & Glasser, R. (1989). Self-explanations: How students study and use examples in learning to solve problems. Cognitive Science, 13, 145-182.

11. Cox, M. T. (1994). Machines that forget: Learning from retrieval failure of mis-indexed explanations. In A. Ram & K. Eiselt (Eds.), Proceedings of the Sixteenth Annual Conference of the Cognitive Science Society (pp. 225-230). Hillsdale, NJ: Lawrence Erlbaum Associates.

12. Cox, M. T. (1996a). An empirical study of computational introspection: Evaluating introspective multistrategy learning in the Meta-AQUA system. In R. S. Michalski & J. Wnek, (Eds.), Proceedings of the Third International Workshop on Multistrategy Learning (pp. 135-146). Menlo Park, CA: AAAI Press.

13. Cox, M. T. (1996b). Introspective Multistrategy Learning: Constructing a learning strategy under reasoning failure (Tech. Rep. No. GIT-CC-96-06). Doctoral dissertation, Georgia Institute of Technology, College of Computing, Atlanta.

14. Cox, M. T., & Ram, A. (1991). Using introspective reasoning to select learning strategies. In R. S. Michalski & G. Tecuci (Eds.), Proceedings of the First International Workshop on Multistrategy Learning (pp. 217-230). Washington, DC: George Mason University, Artificial Intelligence Center.

15. Cox, M. T., & Ram, A. (1992). An explicit representation of forgetting. In J. W. Brahan & G. E. Lasker (Eds.), Proceedings of the Sixth International Conference on Systems Research, Informatics and Cybernetics: Vol. 2. Advances in Artificial Intelligence - Theory and Application (pp. 115-120). Windsor, Ontario, Canada: The International Institute for Advanced Studies in Systems Research and Cybernetics.

16. Cox, M. T., & Ram, A. (1995). Interacting learning-goals: Treating learning as a planning task. In J.-P. Haton, M. Keane & M. Manago (Eds.), Advances in case-based reasoning (pp. 60-74). Berlin: Springer-Verlag.

17. Cullingford, R. (1978). Script application: Computer understanding of newspaper stories (Tech. Rep. No. 116). Doctoral dissertation, Yale University, Department of Computer Science, New Haven, CT.

18. Cullingford, R. (1981). Micro SAM. In R. C. Schank & C. Riesbeck (Eds.), Inside computer understanding: Five programs plus miniatures (pp. 120-135). Hillsdale, NJ: Lawrence Erlbaum Associates.

19. Forrest-Pressley, D. L., MacKinnon, G. E., & Waller, T. G. (Eds.) (1985). Metacognition, cognition and human performance (Vol. 2, Instructional practices). New York: Academic Press.

20. Garner, R. (1987). Metacognition and reading comprehension. Norwood, NJ: Ablex Publishing Corporation.

21. Gavelek, J. R., & Raphael, T. E. (1985). Metacognition, instruction, and the role of questioning activities. In D. L. Forrest-Pressley, G. E. MacKinnon, and T. G. Waller (Eds.), Metacognition, Cognition and Human Performance. Vol. 2 (Instructional Practices), Academic Press, Inc., New York, pp. 103-136.

22. Ghosh, S., Hendler, J., Kambhampati, S., & Kettler, B. (1992). UM Nonlin [a Common Lisp implementation of A. Tate's Nonlin planner]. Available FTP: Hostname: cs.umd.edu Directory: /pub/nonlin Files: nonlin-files.tar.Z

23. Gombert, J. E. (1992). Metalinguistic development. Chicago: University of Chicago Press.

24. Hammond, K. J. (1989). Case-based planning: Viewing planning as a memory task. Vol. 1. of Perspectives in artificial intelligence. San Diego, CA: Academic Press.

25. Hunter, L. E. (1989). Knowledge acquisition planning: Gaining experience through experience (Tech. Rep. No. 678). Doctoral dissertation, Yale University, Department of Computer Science, New Haven, CT.

26. Hunter, L. E. (1990). Planning to learn. In Proceedings of Twelfth Annual Conference of the Cognitive Science Society (pp. 261-276). Hillsdale, NJ: Lawrence Erlbaum Associates.

27. Kass, A., & Leake, D. (1987). Types of explanations (Tech. Rep. No. 523). New Haven, CT: Yale University, Department of Computer Science.

28. Keller, R. M. (1986). Deciding what to learn (Tech. Rep. No. ML-TR-6). Rutgers University, Department of Computer Science.

29. Klahr, D. & Dunbar, K. (1988). Dual space search during scientific reasoning. Cognitive Science, 12, 1-48.

30. Kolodner, J. L. (1993). Case-based reasoning. San Mateo, CA: Morgan Kaufmann Publishers.

31. Krulwich, B. (1991). Determining what to learn in a multi-component planning system. In Proceedings of the Thirteenth Annual Conference of the Cognitive Science Society. Chicago, IL, (August 7-10), pp. 102-107.

32. Laird J. E., Rosenbloom, P. S., & Newell, A. (1986). Chunking in Soar: The anatomy of a general learning mechanism, Machine Learning, 1, 11-46.

33. Leake, D., & Ram, A. (1993). Goal-driven learning: Fundamental issues and symposium report. AI Magazine, 14(4), 67-72.

34. McDermott, J. (1988). Preliminary steps toward a taxonomy of problem-solving methods. In S. Marcus (Ed.), Automating knowledge acquisition for expert systems (pp. 225-256). Boston: Kluwer Academic Publishers.

35. Medin, D. L., Lynch, E. B., Coley, J. D., & Atran, S. (1996). The basic level and privilege in relation to goals, theories, and similarity. In R. S. Michalski & J. Wnek, (Eds.), Proceedings of the Third International Workshop on Multistrategy Learning (pp. 71-83). Menlo Park, CA: AAAI Press.

36. Meehan, J. (1981). Talespin. In R. C. Schank & C. Riesbeck (Eds.), Inside computer understanding: Five programs plus miniatures (pp. 197-258). Hillsdale, NJ: Lawrence Erlbaum Associates.

37. Michalski, R. S. (1991). Inferential learning theory as a basis for multistrategy task-adaptive learning, In R. S. Michalski & G. Tecuci (Eds.), Proceedings of the First International Workshop on Multistrategy Learning (pp. 3-18). Washington, DC: George Mason University, Artificial Intelligence Center.

38. Michalski, R. S. (Ed.). (1993). Multistrategy learning [Special issue]. Machine Learning, 11 (2/3).

39. Michalski, R. S. (1994). Inferential theory of learning: Developing foundations for multistrategy learning. In R. S. Michalski & G. Tecuci (Eds.), Machine learning IV: A multistrategy approach (pp. 3-61). San Francisco: Morgan Kaufmann.

40. Michalski, R. S., & Ram, A. (1995). Learning as goal-driven inference. In A. Ram & D. Leake (Eds.), Goal-driven learning (pp. 479-490). Cambridge, MA: MIT Press/Bradford Books.

41. Michalski, R. S. & Tecuci, G. (Eds.). (1994). Machine learning IV: A multistrategy approach. San Francisco: Morgan Kaufmann.

42. Miner, A. C., & Reder, L. M. (1994). A new look at feeling of knowing: Its metacognitive role in regulating question answering. In J. Metcalfe & A. P. Shimamura (Eds.), Metacognition: Knowing about knowing (pp. 47-70). Cambridge, MA: MIT Press/Bradford Books.

43. Minsky, M. L. (1963). Steps towards artificial intelligence. In E. A. Feigenbaum & J. Feldman (Eds.), Computers and thought (pp. 406-450). New York: McGraw Hill. (Original work published 1961)

44. Newell, A. (1990). Unified theories of cognition. Cambridge, MA: Harvard University Press.

45. Newell, A., & Simon, H. A. (1972). Human problem solving. Englewood Cliffs, NJ: Prentice-Hall.

46. Newell, A., & Simon, H. A. (1976). Computer science as empirical inquiry: Symbols and search. (The 1976 ACM Turing Lecture). Communications of the ACM, 19, 113-126.

47. Owens, C. (1991). A functional taxonomy of abstract plan failures. In Proceedings of the Thirteenth Annual Conference of the Cognitive Science Society (pp. 167-172). Hillsdale, NJ: LEA.

48. Pressley, M., & Forrest-Pressley, D. (1985). Questions and children's cognitive processing. In A. C. Graesser & J. B. Black (Eds.), The psychology of questions (pp. 277-296). Hillsdale, NJ: Lawrence Erlbaum Associates.

49. Puerta, A., Egar, J., Tu, S., & Musen, M. (1992). A multiple-method knowledge-acquisition shell for the automatic generation of knowledge-acquisition tools. Knowledge Acquisition, 4, 171-196.

50. Punch, W. F., III, Goel, A. K., & Brown, D. C. (1996). A knowledge-based selection mechanism for strategic control with application in design, diagnosis and planning. International Journal of Artificial Intelligence Tools, 4(3), 323-348.

51. Pylyshyn, Z. W. (1991). The role of cognitive architecture in theories of cognition. In K. VanLehn (Ed.), Architectures of cognition: The 22nd Carnegie Mellon symposium on cognition (pp. 189-223). Hillsdale, NJ: Lawrence Erlbaum Associates.

52. Ram, A. (1989). Question-driven understanding: An integrated theory of story understanding, memory and learning (Tech. Rep. No. 710). Doctoral dissertation, Yale University, Department of Computer Science, New Haven, CT.

53. Ram, A. (1990a). Decision models: A theory of volitional explanation, In Proceedings of Twelfth Annual Conference of the Cognitive Science Society (pp. 198-205). Hillsdale, NJ: Lawrence Erlbaum Associates.

54. Ram, A. (1990b). Knowledge goals: A theory of interestingness. In Proceedings of Twelfth Annual Conference of the Cognitive Science Society (pp. 206-214). Hillsdale, NJ: Lawrence Erlbaum Associates.

55. Ram, A. (1991). A theory of questions and question asking. Journal of the Learning Sciences, 1, (3&4), 273-318.

56. Ram, A. (1993). Indexing, elaboration and refinement: Incremental learning of explanatory cases. Machine Learning, 10, 201-248.

57. Ram, A. (1994). AQUA: Questions that drive the understanding process. In R. C. Schank, A. Kass, & C. K. Riesbeck (Eds.), Inside case-based explanation (pp. 207-261). Hillsdale, NJ: LEA.

58. Ram, A., & Cox, M. T. (1994). Introspective reasoning using meta-explanations for multistrategy learning. In R. S. Michalski & G. Tecuci (Eds.), Machine learning IV: A multistrategy approach (pp. 349-377). San Francisco: Morgan Kaufmann.

59. Ram, A., Cox, M. T., & Narayanan, S. (1995). Goal-driven learning in multistrategy reasoning and learning systems. In A. Ram & D. Leake (Eds.), Goal-driven learning (pp. 421-437). Cambridge, MA: MIT Press/Bradford Books.

60. Ram, A., & Hunter, L. (1992). The use of explicit goals for knowledge to guide inference and learning. Applied Intelligence, 2(1), 47-73.

61. Ram, A., & Leake, D. (1991). Evaluation of explanatory hypotheses. In Proceedings of the Thirteenth Annual Conference of the Cognitive Science Society (pp. 867-871). Hillsdale, NJ: Lawrence Erlbaum Associates.

62. Ram, A., & Leake, D. (1995). Learning, goals, and learning goals. In A. Ram & D. Leake (Eds.), Goal-driven learning (pp. 1-37). Cambridge, MA: MIT Press/Bradford Books.

63. Reder, L. M. (1987). Strategy selection in question answering. Cognitive Psychology, 19, 90-138.

64. Rosenbloom, P. S., Laird, J. E., & Newell, A. (Eds.). (1993). The Soar Papers: Research on integrated intelligence. Cambridge, MA: MIT Press.

65. Schank, R. C. (1979). Interestingness: Controlling inferences. Artificial Intelligence, 12, 273-297.

66. Schank, R. C. (1986). Explanation patterns: Understanding mechanically and creatively. Hillsdale, NJ: Lawrence Erlbaum Associates.

67. Schneider, W. (1985). Developmental trends in the metamemory-memory behavior relationship: An integrative review. In D. L. Forrest-Pressley, G. E. MacKinnon, & T. G. Waller (Eds.), Metacognition, cognition and human performance (Vol. 1, Theoretical perspectives, pp. 57-109). New York: Academic Press.

68. Simon, H. A. (1979). What the knower knows: Alternative strategies for problem-solving tasks. In F. Klix (Ed.), Human and artificial intelligence (pp. 89-100). Amsterdam: North Holland.

69. Simon, H. A. (1983). Why should machines learn? In R. S. Michalski, J. G. Carbonell, & T. M. Mitchell (Eds.), Machine learning I: An artificial intelligence approach (pp. 25-37). Los Altos, CA: Morgan Kaufmann.

70. Stefik, M. (1981). Planning and metaplanning (MOLGEN: Part 2), Artificial Intelligence. 16, 141-169.

71. Steier, D. M., Laird, J. E., Newell, A., Rosenbloom, P. S., Flynn, R., Golding, A., Polk, T. A., Shivers, O. G., Unruh, A., & Yost, G. R. (1993). Varieties of learning in Soar: 1987. In P. S. Rosenbloom, J. E. Laird, & A. Newell (Eds.), The Soar papers: Research on integrated intelligence (Vol. 1, pp. 537-548). Cambridge, MA: MIT Press. (Original work published in 1987)

72. Davidson, J. E., Deuser, R., & Sternberg, R. J. (1994). The role of metacognition in problem solving. In J. Metcalfe & A. P. Shimamura (Eds.), Metacognition: Knowing about knowing (pp. 207-226). Cambridge, MA: MIT Press/Bradford Books.

73. Stroulia, E. (1994). Failure-driven learning as model-based self-redesign. Doctoral dissertation, Georgia Institute of Technology, College of Computing, Atlanta.

74. Stroulia, E., Shankar, M., Goel, A., & Penberthy, L. (1992). A model-based approach to blame assignment in design. In J. S. Gero (Ed.), Proceedings of AID'92: Second International Conference on AI in Design (pp. 519-537).

75. Tate, A. (1976). Project planning using a hierarchic non-linear planner (Tech. Rep. No. 25). Edinburgh, UK: University of Edinburgh, Department of Artificial Intelligence.

76. Weintraub, M. A. (1991). An explanation-based approach to assigning credit. Doctoral dissertation, Ohio State University, Columbus.

77. Wellman, H. M. (1983). Metamemory revisited. In M. T. H. Chi (Ed.), Contributions to Human Development. Vol. 9 (Trends in memory development research). S. Karger, AG, Basel, Switzerland.

78. Wilensky, R. (1983). Planning and understanding: A computational approach to human reasoning. Reading, MA: Addison-Wesley.

79. Wisniewski, E. J., & Medin, D. L. (1991). Harpoons and long sticks: The interaction of theory and similarity in rule induction. In D. H. Fisher, M. J. Pazzani, & P. Langley (Eds.), Concept formation: Knowledge and experience in unsupervised learning (pp. 237-278). San Mateo, CA: Morgan Kaufmann.

Footnotes

^[1]

Cox (1996b, pp. 294-299) discusses some of the intersections between problem solving and comprehension represented in the top-center of the figure. For example, a problem solver must be able to monitor the execution of a solution to confirm that it achieves its goal. If the comprehension process determines that the goal pursuit is not proceeding as planned, then the planning failure must be addressed and the plan changed.

^[2]

The system's background knowledge, or BK, contains more that just the domain theory of the performance task. It represents all long-term knowledge including metaknowledge, heuristic knowledge, associative knowledge, and knowledge of process. In contrast to the BK, the FK constitutes the current model of the input that has been constructed, and the memory of the reasoning with which such a model was built.

^[3]

See Simon (1979) for a discussion of alternative strategies given a single problem representation.

^[4]

Interesting input is either an anomalous conceptualization or something pertaining to the intrinsic goal of the reasoner. For example, sex, violence, and loud noises are intrinsically interesting (Schank, 1979). In addition, anything concerning a concept about which something has been learned recently will be categorized as interesting. For a more detailed set of interestingness heuristics see Ram (1990b).

^[5]

Figure 5 was produced by a modified version of the Tale-Spin story-generation program (Meehan, 1981). This program provides automatically generated, albeit stylized, input to the Meta-AQUA system.

^[6]

For a far more exhaustive taxonomy, see Kass & Leake (1987).

^[7]

To explain story-understanding events (e.g., the process of explanation itself), Meta-AQUA can generate a third type of explanation. Meta-explanations give a causal account of mental events according to our model of the way things work in the story-understanding process. This class will be briefly discussed in Section 4.

^[8]

The script applier understands a story by matching input sentences to stereotypical sequences of events (i.e., to scripts). For example, a simple pipe-smoking script consists of subscenes to get the pipe, put tobacco into it, smoke it, then clean it, and hierarchically, these scenes are composed of subscenes (see Figure 8). Although scripts omit many of the causal relations between events in a story, they can help an understander interpret a story by providing details not explicitly mentioned in the story.

^[9]

This is not unlike Klahr and Dunbar's (1988) model of scientific discovery, where there is a hypothesis generation phase followed by hypothesis verification and evidence testing phases. The major difference, though, is that IML theory assumes no explicit exploration of a hypothesis space via search. Instead a simple, indexed memory provides suggestions that constitute hypotheses.

^[10]

In the frame definition, =X is a variable binding to the outermost slot named X.

^[11]

An XP is a directed graph with nodes that are either states or processes and links that are either ENABLES links (connecting states with the processes for which they are preconditions), RESULTS links (connecting a process with a result), or INITIATE links (connecting two states). The XP provides a causal justification for a distinguished node called the EXPLAINS node by providing its causal antecedents.

^[12]

Yet, in instances where a hypothesis is not self-generated but provided to the reasoner as input, step one would indeed require significant computation.

^[13]

See also Michalski & Ram (1995) for a more detailed inspection of the relation between views presented here and those of Michalski.

^[14]

A more critical evaluation of the single-strategy approach is that learning is actually a melange of several mechanisms of the architecture (Pylyshyn, 1991). Learning can be obtained as a result of goal-driven problem solving (as is with the Soar framework), or by the passive exposure to experience or goal-orientations (for instance, see Barsalou, 1995), or by instruction, by trial and error, by perceptual reorganization or insight, or numerous other mechanisms. The position here is that learning is best modeled as a multistrategy process, even if different learning strategies are ultimately implemented by a single underlying mechanism.

^[15]

Unlike volitional or physical XPs that explain why persons perform particular actions and how object behave and function, a Meta-XP explains how and why mental actions (such as the explanation process itself) occur. For instance, the EXPLAINS node of IMXP-NOVEL-SITUATION-ALTERNATIVE-REFUTED points to an Expectation Failure (the reader expected one explanation to be true, while another explanation proved to be better), and the Meta-XP provides the causal antecedents that led to the failure (i.e., an erroneous association indexed the first explanation in the BK, whereas the second explanation was missing from the BK). Both Ram & Cox (1994) and Cox (1996b) provide representational details.

^[16]

During mutual re-indexing, the explanations are differentiated based on the object attribute-value of the hit. However, the abstraction transmutation changes this attribute. The generalization method applied to the new explanation also uses this attribute. See Cox & Ram (1995) for a more complete analysis.

^[17]

The notation X(outFK) means that the concept X is out of the set of beliefs with respect to the FK. The semantics of such notation is further explained in Cox (1996b) and Cox and Ram (1992).

^[18]

Note the similarity to the analogous questions pertaining to comprehension on page 12.

^[19]

Kolodner (1993) also speaks of situation assessment (or elaboration) of new input in preparation for case retrieval. The function is the same as input analysis above. In non-trivial systems, a significant part of the problem is to massage the input into a form that is most useful for both processing and retrieval.

^[20]

A special relation exists between metacognition, question asking and text understanding (see Gavelek & Raphael, 1985; Pressley & Forrest-Pressley, 1985). In effect, human learners use question-asking and question-answering strategies to provide an index into their level of comprehension of a given piece of text. This metacognitive feedback helps readers find areas where their understanding of the story is deficient, and thus where greater processing is necessary. Such a perspective supports our ancillary claim that question generation is a key activity in text comprehension and also that meta-level processing is important in such a learning context. As a final tangent, not only is metacognition important in language understanding, it is also important in language generation (i.e., in metalinguistic develompent; see Gombert, 1992).