Massachusetts Institute of Technology
20 Ames St .
Cambridge , MA 02139
We present a preliminary collection of ‘structural critics' that recognize problems with the form of items in a commonsense knowledgebase—where they are not expressed at the right level of detail, suitably contextualized, completely expressed, expressed in a uniform enough vocabulary, or sufficiently unambiguous. We describe these critics formally and discuss how they might be incorporated into a knowledge acquisition system to suggest repairs to existing items of knowledge.
It is difficult to build high quality knowledgebases. It is even more difficult when the people contributing the knowledge are untrained in knowledge engineering! For several years now the Open Mind Commonsense project has been collecting items of commonsense knowledge from the general public over the web [Singh, 2002]. While the collected knowledge consists largely of items most people would agree with (see the evaluation done in [Singh et al. 2002]), unfortunately the knowledge is often not
and suffers from other such ‘structural' problems that lead to difficulties during reasoning. While the collected knowledge has proven useful in a number of applications despite these flaws (e.g. [Barry & Davenport, 2002] and [Liu et al. , 2003]), we are nonetheless exploring ways for it to be repaired.
In this paper we present a preliminary collection of structural critics that can recognize the different kinds of errors that show up within our knowledgebase. Structural critics notice problems with the syntactic form in which knowledge is expressed, as opposed to problems with the content of the knowledge itself. We hope to incorporate these critics into the next generation of the Open Mind knowledge acquisition system, so that the users of the site can repair and improve the expression of the contributed knowledge.
In this section we describe five structural critics that notice different types of problems with a knowledgebase. To represent these critics we make use of a small ontology of ‘meta-structural' predicates that let us describe and compare the syntactic structures of assertions. For the examples below we will use the predicates shown in Table 1.
Table 1 . Meta-structural predicates
Statement(s)— s is a item of knowledge
Subsumes(s1, s2)— The syntactic form of s1 subsumes that of s2
HasModifier(s, pos, m)— Statement s contains a term that is modified by modifier m at position p
RefPronoun(s, p, r)— Statement s contains a referential pronoun r at position p
Contradicts(s1, s2)— The statements s1 and s2 assign contradictory properties to a term
Missing context. This critic notices that a given item of knowledge has implicit contextual assumptions that could be made explicit. For example:
The bride wears white in a wedding.
While this may be true by default, it is more likely to be true if we are talking about people who have received adequate health care throughout their lives. One way to formulate such a critic is to notice that there exist similar items in the knowledgebase that are similar in form to the original knowledge item, except that they add a few additional contextual modifiers. For example:
A wedding necklace is used in a Hindu wedding.
Using the predicates of Table 1, we might express this critic as follows:
MissingContext(s1) ← ∃ s2 [ Statement(s2) ^ Subsumes(Structure(s1), Structure(s2)) ^ HasModifier(s2, p, m) ^ ¬HasModifier(s1, p, m) ]
In the context of an interactive knowledge acquisition system, such a critic could suggest that the user elaborate the given knowledge item to employ a similar contextual modifier, for example, ‘The bride wears white in a Christian wedding.'
Incomplete variable bindings. This critic notices that a given item of knowledge does not fully specify all of the important individuals involved. For example:
Kicking someone causes pain.
This statement is true but vague—who feels the pain? One way to formulate such a critic is to notice that other similar knowledge items make use of more referential pronouns. For example:
Yelling at someone causes them to feel bad.
This item more clearly articulates the causal roles of the individuals involved. While knowledge expressed generally can still be useful, adding this kind of detail often makes it even more useful. We might express this critic as follows:
IncompleteVariableBindings(s) ← ∃ s2 [ Statement(s) ^ Subsumes(Structure(s1), Structure(s2)) ^ RefPronoun(s2, p, r) ^ ¬RefPronoun(s1, p, r) ]
The knowledge acquisition system could pose the analogous formulation ‘Kicking at someone causes them to feel pain', which the user could then reject or approve.
Too general. This critic notices that a given item of knowledge seems to be making a very broad generalization. For example:
People enjoy eating food.
This critic is a generalization of the previous two critics, which both recognize that the form of a given item subsumes other items in the knowledgebase, suggesting that the given item could have been expressed more specifically. We might express this critic as follows:
TooGeneral(s2) ← ∃ s2 [ Statement(s2) ^ Subsumes(Structure(s1), Structure(s2)) ]
The knowledge acquisition system could then suggest to the user various specializations of the given item based on the forms of other items subsumed by the given item.
Varying expressions. This critic notices that a concept or relation is expressed in different ways in different parts of the knowledgebase. For example:
The effect of not sleeping is that you feel tired.
The effect of exhaustion is that you think poorly.
Here there are two different language forms for expressing the concept of fatigue, and so these knowledge items cannot be unified. One way to formulate such a critic is to take advantage of redundantly expressed knowledge within the knowledgebase. If the knowledgebase contains both the assertions ‘Not sleeping makes you tired' and ‘Not sleeping leads to exhaustion', this suggests that those two terms could mean the same thing. We might express this critic as follows:
VaryingExpression(expr) ← ∃ s1, s2 [ Statement(s1) ^ Statement(s2) ^ Difference(s1, s2, pos) ^ ExpressionAt(s1, pos, expr), ExpressionAt(s2, pos, expr2) ^ SemanticallyRelated(expr, expr2) ]
The knowledge acquisition system could then pose to the user more standardized ways to formulate items of knowledge.
Symbol used ambiguously. This critic notices that a given symbol is used ambiguously in different parts of the knowledgebase. For example:
Showers happen outdoors.
Showers are taken indoors.
One way to resolve this contradiction is to posit that there are two senses of the term ‘shower' used here—rain showers, and the kinds of showers ones takes to get clean. One way to formulate such a critic is to notice precisely the type of contradiction shown in the example above:
AmbiguousSymbol(sym) ← ∃ s2 [ Statement(s1) ^ Statement(s2) ^ Contradicts(s1, s2) ]
The knowledge acquisition system could respond to such a critic by suggesting that the user expand the symbol to a less ambiguous form, for example, ‘Rain showers happen outdoors'.
We have presented a few types of structural critics reflecting some of the common types of problems that show up in the Open Mind Commonsense knowledgebase, and have shown some simple ways they could be expressed. While we expect that these initial formulations will be not spot problems very accurately, we see them as preliminary steps towards a richer classification of the kinds of defects that knowledge bases may contain.
[Barry & Davenport, 2002] Barry, B. & Davenport. G. (2002). Why Common Sense for Video Production? (Interactive Cinema Technical Report #02-01). Media Lab, MIT.
[Liu et al., 2003] Liu, H., Lieberman, H., Selker, T. (2003). A Model of Textual Affect Sensing using Real-World Knowledge. In Proceedings of IUI 2003. Miami , Florida .
[Singh, 2002] Singh, P. (2002). The public acquisition of commonsense knowledge. In Proceedings of AAAI Spring Symposium: Acquiring (and Using) Linguistic (and World) Knowledge for Information Access. Palo Alto , CA , AAAI.
[Singh et al., 2002] Singh, Push, Lin, Thomas, Mueller, Erik T., Lim, Grace, Perkins, Travell, & Zhu, Wan Li (2002). Open Mind Common Sense: Knowledge acquisition from the general public. In Proceedings of ODBASE'02 . Lecture Notes in Computer Science. Heidelberg : Springer-Verlag.