Recent years have seen a renewed interest in cognitive systems with the ability to explain either external phenomena or their own internal reasoning processes while solving problems. Some successful models of explanationgeneration have made use of structured representations, reasoned over using analogical or deductive mechanisms. But before such models can be adapted for use in real-world situations, they need to incorporate additional features associated with explanation-generation. For example, generated explanations may differ qualitatively based on the explanandum's domain; e.g., explanations rooted in physical causality to explain physical phenomena vs. folk-psychology explanations that rely on propositional attitudes (believes, knows, intends, . . .). This may affect the generated explanations in both explicit and implicit ways. We tackle both the explicit and implicit effects of this cognitive feature and incorporate them into a comprehensive cognitive architecture: CLARION (especially its meta-cognitive and non-action-centered subsystems).
These two explananda refer to the same event at different levels of abstraction by invoking different concepts. The type of explanation (or alternately, explanans) that might be deemed an appropriate response to each of these explananda differ as well. An explanation whose language features many propositional attitudes of the chicken (e.g. “C believes,” “C knows,” “C wants,” etc.) may be appropriate for explaining E1, but may not constitute a satisfactory explanation in response to E2. An explanation rooted in physical causality (referring to the normal properties of muscle contractions, for example) may be the other way around: it would be appropriate for E2 but less so in response to E1. In short, the presentation of the explanandum affects the sort of explanation that is most appropriate.
The question hinted at in the above example, of whether to root an explanation in physical causality or propositional attitudes, reflects a parallel one faced by cognitive systems: What factors are used by agents to determine which qualitative features of an explanation are appropriate? In the present paper, we explore and model one answer to this question: that the concepts used in the presentation of the explanandum affect the explanans in both implicit and explicit ways. We model these ways using the MetaCognitive Subsystem (MCS) of the CLARION cognitive architecture.
We do not hope, nor do we attempt, to resolve any questions regarding whether
one type of explanation is better than another. Although discussion in the philosophical
literature of the so-called intentional stance [
The remainder of this paper proceeds as follows. After further motivating the features the modeling of which is our target, will first discuss related previous work in modeling explanation-generation, in order to set the stage for the communication of our own, and to introduce concepts we use in this paper such as metaknowledge, metacognition, and so on (§2). In section 3, we present the cognitive architecture CLARION, and briefly discuss recent developments in its representational capabilities which make it possible for us to do the work we present herein. We close with brief demonstrations in section 4, and section 5 concludes with final remarks. 1.1
The type of feature of explanation-generation we aim to model here, which we refer to as F effects for convenience, are the effects that the presentation of the explanandum has on the explanation generated. If the explanandum e is a simple fact about some world, let us define the full explanandum E as the explanandum plus all of the contextual facts required to understand the explanandum. For example, to return to the earlier example, the position of the chicken relative to the road, the position of the road relative to the four cardinal directions, and so on, are all examples of facts comprising E. The presentation of the full explanandum P(E) is a particular form of the full explanandum E. This distinction is important. E1 and E2 might be considered two partial presentations of the same full explanandum, but they differ in their presentations.
F effects, then, are those which the presentation of the full explanandum exhibits
on the explanations generated. We can further subdivide these into Fe effects, and Fi
effects; these are explicit and implicit effects, respectively. Examples of both in the
psychological literature are numerous, e.g. see [
Determinations of similarity based on simple featural overlap might be considered
an implicit process, or one that operates primarily using the representations on
CLARION’s lower level [
But explicit processes may play a large role in explanation as well. One way to
identify explicit processes, or those that operate primarily using the representations like
those on CLARION’s top level, is to perform experiments on human subjects that
require them to verbalize their thoughts in some way. In explanation, one example relates
to the so-called “self-explanation effect,” in which children who verbalize their
explanations seem to be able to improve the quality of their learning, and learn more [
Our basic hypothesis for the modeling of F effects in the present paper is that the knowledge structures used to construct explanations are selected based on parameters in the metacognitive system, which themselves may be influenced, either explicitly or implicitly, by the concepts used in the explanandum’s presentation. 2
In this section we provide an overview of some recent modeling of metacognition in order to give the reader a feel for the state of the art in the field, and to clarify the present paper’s contribution. Explanation, and in particular the modeling of explanation using analogy, has been tackled before. Thagard (2012) divides the computational models of explanation thus far into four types: probabilistic; those based on artificial neural networks; logical; and those based on schemas or analogy . The approach described in this paper falls in between the last two of these four types, since the template-matching system which we describe in the next section allows for both rule-based deductive reasoning and a form of analogical reasoning.
Hummel and Landy [
Friedman and Forbus [
Let us make two broad observations from the preceding summaries of literature.
First, we see a form of metacognition in the work by Friedman and Forbus [
Therefore, the work we propose in this paper distinguishes itself from the above approaches, on whose shoulders our work stands, in four key ways. First, our approach distinguishes between the full explanandum and its presentation. Second, we assume that this presentation affects a metacognitive system which in turn constrains the type of explanation that is generated. Third, we propose the use of specialized knowledge structures, such as templates and constraint chunks (both of which are described shortly), to allow such constraints to take the form of highly expressive knowledge structures.
Finally, we acknowledge both explicit and implicit effects of the explanandum on the explanation generation, and model both using the cognitive architecture CLARION, in such a way as to take advantage of the features it provides. We next summarize the aspects of CLARION we have used. 3
CLARION is an integrative cognitive architecture with a several key features that we
take advantage of here. These features include dual representation, a division of
cognitive subsystems in a way that has previously been demonstrated to be
psychologically plausible, and a flexible knowledge framework which can capture sub-conceptual,
unstructured-conceptual, and structured-conceptual knowledge simultaneously [
WHOLE COMPONENT
Chases Dog
Cat Dog
Cat
level division is reflected in each of CLARION’s subsystems: the Motivational,
ActionCentered, Non-Action-Centered, and Meta-Cognitive Subsystems (MS, ACS, NACS,
and MCS, respectively). A primary focus of CLARION has been psychological
plausibility, and much work has been devoted to defining mechanisms within its subsystems
that are tied to known psychological phenomena and processes [
The focus in the present paper is exclusively on an interaction between the NACS
and MCS. In particular, recent work by Licato et al. has demonstrated how structured
knowledge can be represented and reasoned over using no more than the
psychologically plausible mechanisms already defined in the NACS [
The NACS contains declarative knowledge, or general knowledge about the world
that is not action-centered, which is often used for making inferences on the basis of its
knowledge. The top level of the NACS contains localist chunks linked to units on the
bottom level called DV pairs (Dimension-Value) pairs. The DV pairs can be linked to
each other, and the chunks can also be linked to each other. However, the links between
chunks are a special type of directed link called Associative Rules (ARs), which are
represented pictorially using arrows. All of the links between top and bottom level units
have weights that can be changed over time. This unique structure allows CLARION
to define a directed similarity measure between two chunks [
The MCS [
Structured knowledge in the NACS is achieved by first allowing top-level chunks to differentiate into types: object chunks, proposition chunks, template chunks, etc. These chunks are then linked using ARs and specialized chunks called Cognitively Distinguished Chunks (CDCs). For example, the proposition Chases(Dog,Cat) can be represented as in Figure 1. 3.1
Analogical and deductive reasoning are carried out by defining special structures called
Templates. These are essentially NACS structures that have been grouped under a single
Template Chunk (TC) using a CDC defined for that purpose. In deductive reasoning, a
template can specify the antecedent and consequent portions of a rule separately, so that
when a structure sufficiently matches the antecedent portion, the consequent contains
information on how to transfer the matched knowledge structure to create a new
inference. Analogical reasoning can also be modeled by converting potential source analogs
into templates and relaxing the match requirements. Matching structures to templates
uses an Ant Colony Optimization algorithm inspired by [
Explanation-construction proceeds as follows. We assume that we are given a knowledge base of templates. Each template is either a single structure, in which case it is to be used as a source analog for analogical matching and inference, or the template consists of an antecedent and consequent portion, in which case it is to be used as a deductive inference rule (e.g. Figure 2).
Given some knowledge structure K and template T , if a match is found from K to T (using the minimum conditions for an acceptable match specified by T itself), then a new structure K0 is created from the elements of K and the instructions provided by T (these instructions are not explicitly stated by T in any way, rather they are implicit in the template’s structure itself).1
Each template is grouped under a single Template Chunk (TC). The chunks in each template may each be linked to DV pairs in the NACS bottom level, and the template’s TC is linked to a disjunction of all DV pairs linked to all non-CDC chunks in the template.
Algorithm 1 The Template Selection algorithm. This is used to filter out the template structures and select a subset of them based on how much they satisfy the constraints.
We can now introduce a new type of chunk, which we will call a Constraint Chunk (CC). A CC is a chunk that resides on the top level of the MCS, and is used to either bias the parameters of cognitive processes based in the other (non-MCS) subsystems, or to point to the TC of a template which serves as a inviolable rule to constrain cognitive processes. The precise way in which it performs this biasing function is described shortly in the present section.
Just as the NACS chunks are linked to distributed units on the NACS bottom level, CCs are also linked to distributed units on the bottom level of the MCS. However, unless a similarity measure is defined between elements on the bottom levels of the NACS and MCS, no similarity measure will exist between chunks on their top levels. At least for this project, then, the design decision was made to allow the NACS and MCS to draw from a common pool of bottom-level distributed units, so that the same similarity measures used between two chunks of the NACS could be used from NACS to MCS chunks.2
Explanation generation is a simple backward-chaining process that starts with a set of knowledge structures B = {Bi} corresponding to beliefs or knowledge that the agent holds, which are not part of the full explanandum, a set of templates T = {Ti}, a set of CCs C = {Ci}, and a full explanandum E.
The algorithm will start by selecting the relevant template structures. This requires
that we have a set of CCs which are currently created manually in order to allow external
users to set the qualitative features of the desired explanation, but the CCs are in such
2 This design decision is partially justified by CLARION’s view that meta-cognitive processes
are intermeshed with other processes, and although the MCS is treated as a separate subsystem,
it should really be viewed as closely integrated with the processes of the other subsystems [
Next, the algorithm selects Ti 2 T subject to the constraints set by the CCs. It does
this by activating all templates a fixed amount, and then activating excitatory CCs,
allowing the activation to propagate using similarity-based reasoning [
As a result, we have a degree of activation for each Ti 2 T which reflects the constraints defined by the CCs. We collect the top n template chunks with the highest activations. This resulting set of templates is called the active template set (TA). The pseudocode for the creation of TA is shown in Algorithm 1.
The backward-chaining process can now begin. The algorithm starts by defining S as the set of facts in the full explanandum E. The templates are momentarily reversed: If some fact s 2 S matches the conclusion portion of a template in TA, inference is performed on the antecedent portion of that template to produce a new set of facts, which replace s in S. If any of these newly added facts match beliefs in B, they are removed from S. This constitutes a single iteration of the backward-chaining process, which repeats until either S is empty, no more facts are found that can be added to S, or a preset time limit is reached. The remaining facts in S are then outputted as abductive assumptions.
We offer the pseudocode describing the general explanation-construction algorithm in Algorithm 2.
Require: Beliefs or knowledge the agent holds B = {Bi} Require: A set of active templates TA obtained through Algorithm 1. Require: Set of facts S = {si} in full explanandum E.
Let currAssumptions S while currAssumptions * B or timeout not yet reached do for all t 2 TA do if Consequent of t matches some f 2 currAssumptions and f 2 / B then Let A = The facts comprising the antecedent of t currAssumptions (currAssumptions { f }) S A end if end for end while return currAssumptions as the abductive explanation of E.
Our very brief proof-of-concept demonstrations will serve as examples for testing the model we describe in this paper. These examples attempt to construct explanations when given a small knowledge-base, using the analogical comparison and transfer mechanisms defined in the NACS and the constraints in the MCS.
Wants& FoxF3&
Grapes&
Not& Reach&
Decides&
Sour& In order to clarify how we model the implicit and explicit effects of full explanandum presentation on explanation, we present a simple example demonstration that generates explanations for why a chicken crossed the road. The two full explananda, presented here in English for readability, are:
Ei The chicken decided to cross the road; the chicken was heading East. Ep The chicken’s body moved, crossing the road; the chicken was heading East.
Note that there is a very subtle difference in presentation: Ei invokes the concept of “deciding” whereas Ep does not. The algorithm will construct a new CC by simply creating a new chunk whose connected DV pairs are the disjunction of the DV pairs connected to the chunks in P(E), the presentation of the full explanandum. This new CC bias the templates selected in the explanation-generation step, and thus will allow us to test Fi constraints. The templates provided to the system would include: – If there is wind blowing east, and that wind is blowing on an object o, then o will move east. – If c wants to achieve goal g, and g requires that action a happen, then c will decide to perform action a. – If there is an object o that is East of c, and c likes o, then c will want to achieve the goal of moving East. – If the chicken wants to achieve the goal of moving East, then the action of the chicken crossing the road must happen.
We now run the explanation-generation algorithm, and output the top explanation generated. When full explanandum Ei was used, the explanation generated the majority of the time (presented here again in English for readability) was:
Assume there is an object o that is East of the chicken. Assume the chicken likes o. The chicken will want to achieve the goal of moving East. The action of the chicken crossing the road must happen. The chicken will decide to cross the road.
Assume that there is wind blowing east. Assume that wind is blowing on the chicken. The chicken will move east.
Explicit effects are modeled by creating an inhibitory CC that is also the template chunk for a structure corresponding to the proposition p = “The wind is blowing east.” This will attempt to prevent any explanations that have p as one of its intermediate structures.
We ran the trial with Ep as the full explanandum, except this time the inhibitory CC corresponding to p is included. As expected, the explanations which require that the wind is blowing east are suppressed, and the explanation is generated as if Ei were provided instead. 5
It is increasingly important that cognitive systems be able to explain and justify their conclusions and choices to the humans they will inevitably work with. For such systems, generating qualitatively different types of explanations may be essential. Using the work we have presented in this paper, such a thing can be accomplished with a few parameter changes in a meta-cognitive system. These parameters may be changed autonomously according to contextual factors, or by normal processes rooted in CLARION’s subsystems, such as the ACS, MCS, or MS. We have presented a model that can explain produce explanations at different levels of abstraction, like E1 and E2 in this paper’s introduction.
The work here is certainly not complete; a much wider variety of explanations must eventually be addressed. For example, the ability to justify behaviors using a proof defined in a fully formalized logic is (for some domains) a glaring absence to be tackled soon, but the work in this paper can be used as a springboard for moving in that direction.
An obvious next step is to flesh out the proof-of-concept demonstration briefly described in this paper, and to examine how it performs when provided with a much larger knowledge-base. Furthermore, more sophisticated deductive reasoning is necessary to augment the part of our explanation-generation algorithm that uses inhibitory CCs corresponding to full structures. In the demonstration we presented herein, p =“The wind is blowing east.” was used to find and suppress templates that may have led to intermediate propositions equivalent to p. But if a template leads to a logically equivalent proposition such as “The wind is not not blowing east,” our algorithm would have failed.
Finally, our current system does not demonstrate learning. If the templates drawn on by the explanation generator are insufficient, then presumably a human would eventually learn a new set of templates, somehow; this is not modeled in the present work. Clearly, there is much to do.
This work was funded by grant N000141310342 from the Office of Naval Research.