From: AAAI-88 Proceedings. Copyright ©1988, AAAI (www.aaai.org). All rights reserved. eyornd HSA: Structures Paul R. Cohen Experimental Depaztment and Knowledge of computer University present a method for inference rules We describe base. edge these First, rules. automatically we derived jects to discover second study sibility they which tested of these obey cusses four ments, and dictions Introduction Can cough syrup that make people rules (e.g., has five toes) Rules increasingly plicitly X CONTAINS y, CAUSES z 2 CAUSES t two cause sitions to reduce that can, and alcohol is rule: and structure 1 obviously bases as property [Lenat and inheritance over the same purposes, that is, supportand knowledge engineering. We have describes smoke, so based on their them inverses. In the inference over 3000 specific to human permissible 3). We will begin Then subjects the plausibility 2 periment Cohen 4500 first rules and nodes study from inferences we these by re- from the GRANT (Sec. 2). which The sec- that the plausibility of these they obey a kind of transitiv- by describing we will discuss assessing rules. we de- to discover rules were plausible ond study tested the hypothesis rules can be predicted by whether that roughly in the rules with concepts and presented of these base et al., 1985; contains 300 plausible we generated studies knowledge [Cohen KB and their the variables syntactically large GRANT approximately Then ones two empirical project The by 9 relations placing implausible on a moderately 19871. relations. KB, ones from for the GRANT ity (Sec. fires cause Rule structure.” ses, and results. that we expect years. First, knowledge syrup at the system makes simply won’t and Feigenbaum, degradation knowledge (e.g., roles depends the effort engineers can be plausibly In this section support of our know won’t ex- offer a the question [Lenat Collins et al., 19751. knowledge, which inheritance these studies, hypothe- the role of knowledge in of inferences. I: we describe to produce Identifying 2.1 state inferred. knowledge explicitly Property bases, those Property inheritance befor *We are indebted to Carole Beal, David Day, and Adele Howe for their comments on drafts of this paper, to Carole Beal for her help with the statistical analysis, and Evan Smith for his assistance with this project. ‘This research is funded by the Office of Naval Research, under a University Research Initiative Grant, Contract #N0001486-K-0764 and by a gift from Tektronix. of these rules. Background over ISA links can be written we propo- inheritance, how to use the structure of property many other plausible inference rules, the plausibility and how we determined rules such as Rule 1, to make Second, we expect plausible of building needn’t doesn’t you drunk answer 1987; to become they boundaries that on general formulate as as plausible inference up for a lack of specific knowledge. inference paper depend veloped brand inference Y in coming cough plausible answer-it et al., 1986; Lenat Graceful alcohol plausible of performance A brittle whether Our favorite contains abduction have important knowledge. drunk? and knowledge look for a fire). these degradation gracehE “deep Kjeldsen, a general and causal like pre- of plausible inference include property inhave five toes, Ginger is a cat, so Ginger cats if you see smoke, This Both If you didn’t already know that you could infer it from two spe- syrup from familiar plausible underlying knowl- accurate inheritance “mega-frame” 1 has the same derived Rule P Other moderately 19871. ferentiate disjudg- little Feigenbaum, linked cific propositions-cough heritance in subjects’ property developed a simple method for deriving such rules from the relations in a knowledge base, and we have shown how to dif- by whether relatively like to build links, and can serve ing graceful degradation plau- paper Rules ISA The the Science are needed Rule spesub- judgments. because it contains alcohol. cough syrup is intoxicating, intoxicating-and that The of variance to achieve 1 plausible. can be predicted concludes of these were hypothesis Labora.tory relieves us from having to state explicitly that each of a class has each property of that class [Brachman, 19851. 300 over 3000 to human of transitivity. sources edge is needed rules the rules a kind deriving approximately Systems and Information example, member from relations in a knowltwo empirical studies of plausible inference rules, generated cific inferences, and presented them L. Eoiselle Massachusetts Abstract plausible Cynthia of Massachusetts Amherst, We antic Networks for PBausible where erty the relation of n2. For ISA n2, n2 R n3 n1 R 723 R between example, wings, HAS-COMPONENT and n2 and n3 is viewed if a canary then canary as a prop- is a bird and HAS-COMPONENT bird wings (Fig. 1.a). Here, R is HAS-COMPONENT and the inherited propMany plausible inference erty is “HAS-COMPONENT wings.” rules ISA. have this For example, syrup inherits TAINS structure, but inherit in the “cough syrup” the “CAUSES intoxication” over links inference, property other above, than cough over the CON- relation: Cohen and Loiselle 415 cough syrup alcohol cough Figure syrup 1.b shows premises HAS-COMPONENT alcohol, CAUSES intoxication base CAUSES intoxication pairs of relations (including relations and an equal number of rules. The two other but different cloud” COMPONENT storm”); the other (and, “rain “storm MECHANISM-OF They cloud”). rain” HAS- COMPONENT-OF rain” (and, equivnthe conclusions arc But and “rain Since the same is “storm “cloud MECHANISM-OF HAS-MECHANISM have One premise equivalently, is “cloud lently, examples. conclusions. and storm,” COMPONENT-OF each pair of relations constructed from nine from relations Experiment sible rules produce I over -- - * - component a. Property the inheritance over ISA component-of mechanism-of cloud - 300 so (18’ + 18)/2 to the thousands the effort then was to generate by chaining GRANT of first storm b. Figure This rain .Inferences that property 1: Inheritance illustrates : but, generation have large have the inheritance structure inferences Each and structurally-identical that each pair of relations plausible inference rules that have the same erty inheritance over ISA links. For relations rules knowledge can produce two are: method Rule 2 Rl ?xz, n2 ~2 nn n1 R2 n3 tllld nected R2 = 1.b shows these 1 introduces R2-INV R~, RI-INV n1 n3 RI-INV nj the drawn the conclusion and is always n2 = cloud, as solid lines. drawn we will The +first genernfion inference, “storm and n3 = rain. nse throughout. hypotenuse as a dashed represents line. MECHANISM-OF as the premise of a. second generation inference, runoff.” conclusion “storm HAS-PRODUCT mechanism-of cloud erain rain has-product .- * .---has .- -* product .- storm storm. First rain,” which .fl .-- .-- . .- generation inference Figure Knowledge Representation storm, cloud, rain” serves has the runoff test inference of relations triples. We to be first pair, say HAS- the GRANT for KB and rain instantiate l.b, and generation with yielding “rain test is repeated generation item during the previous the conclu- COMPONENT-OF items. (though test yield Procedure Items in the a computer data set were program. Next, the whether set or else to indicate Each item sion with that was seen 20 items judged second were and approximately that of each was produced the of 3116 to of which test human asked concluitems, of subjects first subjects by to indicate were from Following was in the 700 items gen- one or both were unone or both premises. did not understand by two subjects. (none generation one premise or did not follow they the and half are second presented was shown it followed we add KB. not necessarily a data Subjects conclusion of conclu- item). whether both premises were acceptable, acceptable, or they did not understand judge dozens their However, that which roughly half are first generation eration items.2 2.3 yield and to generate condition search rules KB five, must be a conclusion sion of a first generation the 315 GRANT select to the GRANT the added second In all, in the randomly of al! the triples procedure items, subject Second generation inference 2: Second-generation we pro- and five second For each we search in Figure MECHANISM-OF pairs This Rules can be chained by letting the conclusion of one serve as a premise for another. Figure 2 shows how the conclusion of a 416 “storm conclusions Rules are represented as triangles formed from three concepts and three relations. The legs of the triangle represent premises, and are always For instance, and n3, respectively n2, sions, a11 d for ~1 = HAS-COMPONENT, notation on two relations. and MECHANISM-OF, and 3, above). Most nz n1 = storm, MECHANISM-OF, Figure For each KB.’ storm.” m alternatives to judge. to n3 by MECHANISM-OF). Each triple represents a pair from which tvro inferences can be drawn (see Rules 2 n1 ,n,z,nn Figure subjects conclusions, then present of premises siolls and Rule 3 to bases. triples of nodes nl,n2,n3 that are connected by these relations (i.e., nr is connected to n2 by HAS-COMPONENT, and n2 is con- nl, nl very or predict a powerful 315 rules from the GRANT rule is based COMPONENT structure as prop~1, Rz these rules and We expected duced 10 test items (five first generation items generation items) by the following method: of rules. we describe if we could discover we would of constructing to human We derived * all pos- Design them : “component of = 342 these (as KB inferences. to be plausible; ones, One To determine whether the rules produce plausible we first instantiate them with specific concepts, : + N)/2 with themselves) KB is constructed GRANT is affected rules second-generation plausible 2.2 &storm cloud rain conclusions. produced 200,000 reduce links (N2 and to determine which of them The other was to find out how KB of conclusions roughly few of these canary” GRANT plausible below), .---has a knowledge produce paired inverses, 1 had two goals. for the Applying 2 .- and their two rules, will were generated. the plausibility respectively. produces N relations the asked to premises, the conclusion. a practice data from the data ses- set), each set. This ‘Pruning duplicates reduces the original 342 rules to 315. 2We don’t have 3150 items because, for some rules, the GRANT KB yielded fewer than five first generation instances. took about five hours, distributed over three or four self-paced sessions. Results 2.4 Since the of the test items came from an existing COhWONENT-OF knowledge base we expected that most would be .judged acceptof first generation able. This is in fact the case: 82% percent premises and 63% of second generation premises were judged to be acceptable. The following results pertain only MECHANISM-OF items. Each ation rule is represented items and in the data five second set by five first generation items, and each item SUBFIELD-OF ranging from 0 to 10, are equal to the sum of t,he number judged plausible plausibility score, first generation is 4.18 ing statistic Both items for second are significantly most chaining This inferences eration ones. The fact that to produce to have plausibility judged judged (var. = 4.88). both this, one would expect increasingly-implausible evidence that less plausible approximately between 3 and Table t h h SUBJECIh -4 HAS-SUBFIELD relations + h and corresponding deep relations nJ in one rule and n3 to nl in the other). In contrast, the intransitive structures do not require any ordering on nodes in the con- second than 50% conclusion. gen- first gen- of the rules 7 (of a possible 10); of the rules to be predominantly In one, ing between to both. tween in the temporally-prior them and, indicate that no hierarchical other intransitive rule, to 7~2, but no ordering thus, required order- n2 is hierarchically-superior in the and n1 no is implied conclusion. The bemean h plausible “2 -“3 h 1 Discussion "1 While these results indicate that many rules generate and many others are inantly plausible conclusions, the premises nr and n3, only Similarly, are both or implausible. 2.5 t * SETTING gh Surface 1: ) and the correspondis 3.17 Given by the scores the rest of for and from each other are below chance means are significantly Subjects generation 315 rules, from chance is supported eration they items rules are not plausible. of inferences clusions. = 6.92), generation different at the p < .Ol level. that (var. for each over the HAS-PURPOSE h SETTING-OF a rule, +v I I SUBJECT-OF -+T HAS-PRODUCT + was seen by two subjects. Thus, 10 judgments are made of the items in each generation of each rule. Two plausibility scores for of items that subjects each rule. The mean HAS-MECHANISM PURPOSE-OF gener- h HAS-FOCUS + PRODUCT-OF h HAscoMPoNFiNT h FOCUS-OF to those CAUSED-BY h * Deep structure relation t CAUSES premises Surface Deep structure Surface relation br-’ -- : .-- : h .- .-. h 1 .: .- “1 h predonlpredomi. nantly implausible, they do not tell us how to predict which will be plausible and which will not. We wanted to find a small set of common predictions. depend only characteristics of rules on which to base these Furthermore, we wanted these characteristics on the relations in the rules, not on the nodes any exogenous to or a. factors. We discovered two common aspects of relations. Some Example transitive deep structures rela- Figure tions, such as HAS-COMPONENT have a hierarchical interpretation. Others, such as CAUSES, can be interpreted as temporalre. lations. Lastly, hierarchical ~2 may OF n2 ," mechanism is required that relations be a process Expressing rules rules importantly, structures, called rules were plausible and TZ~, then and others represent ordering particular the nr CAUSES structures between ordering n3, then between reduces ones. nj 3 shows two the premises to be preserved, the ~2." requires nr and n3 in the conclusion (nr surface evidence deep relations where rules that relations are For example, that when structures t and are the structure We call structures for ambiguous have transitive interpretations, to all the and corresponding 4.a like this ones. map to deep were of interpretations, reBut PURPOSE-OF have one may directions. interpretation under the other. be transitive or ambiguous. transitivity structures, predicts the the results All ambiguous structures but we knew from our data rules but provide one direction. in opposite under our data suggested that of rules with unambiguous sized a characteristic in just intransitive in Figure and is a factor. relations h point transitive were less clear not point necessarily \;t'~' a transitivity surface implausible, (p < .Ol), HAS-MECHANISM Although plausibility an are judged different h and t elements relations Therefore, rules n!, n2, and nl imply whose of these of these is clear intransitive. transitive COMPONENT-OF because n3 that, the of deep “If n1 CAUSES and intransitive are significantly post-hoc lations why some The values strong deep to explain Figure preponderance these Transitivity or both. a characteristic implausible. the relatiorc ns,” and “If n1 COMPONENT-OF transitive n.1 and deep relations seemed rules: the deep structures. and two intransitive n2 COMPONENT-OF call these deep that Each interpretation, we identified structures structures CAUSES of these subsumes Example intransitive deep structures plausibility score for transitive rules was 8.94 (out of 20; var. = 16.83), and for intransitive rules, 5.89 (var. = 14.46). Again, MECHANISM- or process that exists or 1 lists the deep relationa relations. to 95 unique transitivity, can have both in “nl hierarchically or t (temporal) in terms set of 315 surface deep that to all 18 surface has a h (hierarchical) transitive interpretations: nr, or nr may be an object prior to achieving nz. Table correspond More such as MECHANISM-OF and temporal 3: Transitive b. plausible. called We that hypothe- consistency, Cohen and Loiselle that 417 t n2 - t t n, -a. Consistent interpretation Figure b. Consistent interpretation h 4: Ambiguous Transitive % t deep structures might discriminate ones. A structure plausible ambiguous rules from implausible has a consistent interpretation when its deep relations the all have same interpretation, example, Figure 4.a has a consistent all its deep links can be interpreted interpretation is transitive. pretation, deep (t, but relations Figure either h or t. 4.b has a consistent it is intransitive; and the interpretations 4.b transitive Figure I 1 Figure 5: Single interpretation deep structures For interpretation in which as h. Moreover, this h that make lntransltlve h ..-fl : .- ...* .- t predicts the plausibility of these rules. A graph of the means t interof the 60 % - are inconsistent t, and h). inconsistent 3 At Experiment the 2: Exploring end of Experiment 1, we had that transitivity predicts termines the interpretation biguous structures. plausibility, (transitive Experiment Transitivity formed the hypotheses and that consistency deor intransitive) of am- 2 tests these 20% hypotheses. Experiment 2 1: COMPONENT-OF, CAUSES, deep relations edge base.) Since rules) first generation inverses. each of these test items other relations infrequently surface relations the 95 rules they From for each these, of the (Fig. replicate 6: Transitivity 6) suggests in the knowl- that has a unique sibility chance) generate we chose sample.3 Figure Experiment map to as in Experiment We generated 10 56 rules, just as we items. Items were the mean and inconsistent transitive each viewed all the test is that interpretation predict interpretation, the plausibility because of rules the mean plau- plausibility scores rules to chance items of transitive, performance; are collapsed intransitive, transitive into and in- one category. rules are composed deep interpretation COMPONENT-OF; Inconsistent Intransitive with single interpretations +20% transitivity, of the Eight deep as determined structure, of surface (CAUSES, see Fig. of transitivity effect predicts relations CAUSED-BY, 5). With and consistency interpretations. main we cannot inconsistent 1. ig .r .-ul 6= _m 8 a Results significant analysis score for these rules is roughly five out of 10 (i.e., at irrespective of whether the rule is transitive. Figure 7 compares 56 structures that have no consistent Transitive presented Our hypothesis consistency x PRODUCT-OF, Procedure subjects single from 1. Fourteen fects (The relatively as a representative did in Experiment tent relations MECHANISM-OF, deep relation, thus, 3.3 ten deep structures. (and 3.2 on and occurred corresponding 95 different focused and their PURPOSE-OF I ini;ym;tive transitive items Design 3.1 A two-way these that plausibility. have just analysis one Chance g 5 (50%) $ .+ a -10% -20% the ef- in rules with of variance found -30% a (p < .OOl) and a signifi- 3Rules generated from a single surface relation and its inverse always map to one transitive and two intransitive deep structures. Our sample included the transitive structure and one of the intransitive structures (chosen randomly). Pairs of non-identical relations and their inverses form four transitive and four intransitive rules. Our sample included two transitive and two intransitive rules from each of these sets. Knowledge Representation + 10% HAS-COMPONENT, we can analyze on plausibility of transitivity by the consis- Figure cant transitivity x consistency interaction (p < .OOl), but no main effect of consistency (p > .2), confirming that transitivity 418 - 7: Scores Analyzing that for rules all our rules in terms have consistent transitive transitive rules, transitive and intransitive togram presented 8 inconsistent for all rules in Figure “Unfortunately, of these categories interpretations, rules, consistent (including the yields 20 consistent and 4 rules that have both interpretations.4 eight analyzed 18 in- The earlier) hisis 8. the test items for the other six rules shared has-mechanism -weapon war +30% $2 .e e ‘Fjz g = mu ‘Ea 8 0” lf =r n lnconslstent rules consistent transitive & lntransltive rules battle a. . Surface +20% plausible sistent to infer 8: Scores interpretations. expected for all rules structure rule However, the mean one of our earlier of rules with con. plausibility score for tween GPI intransitive and relatively it was in Figure 7. Discussion While the predictive interpretations. transitive score chance, roughly of inconsistent was higher 8). rules, (61%); and the intransitive to be implausible, GPI with bility scores. that rules with conhave the scores a mean have higher-thanrules are non-GPI discernible transitivity Among intransitive intra.nsitive rules, classes-relatively rules. After a post-hoc have transitive tests difference removing items GPI GPI rules, the (Fig. 10). of why intran- higher-than-expected items, GPI dif- plausible rules decreases explanation on be- plausi- had no statistically effect. And since there was no interaction and GPI, we regard them as independent between factors. plau- to be at plausibility interpretations, and among rules Post-hoc for tran- the mean we expected mean of those transitive effect. of inconsistent provides diluted However, which score and inconsistent in rules that implausible plausibility sitive between (Fig. some a significant two statistically-distinct that interpretations halfway rules with consistent pected it becomes items means is high for rules It is not surprising and intransitive sibility of transitivity und intransitive score which Therefore, power one interpretation, plausibility (p < .05), mean 3.4 properties rules Since scores. .05), with no interaction (Newman-Keuls) found GPI (p < the means than multiple will have also GPI, we ran a post-hoc transitivity x GPI analysis of variance, and found main effects of transitivity (p < .OOl) and ferentiates sistent a concept plausibility inconsistent rules is higher than chance, and the mean plausibility score of consistent intransitive rules is much closer to chance only that concepts hierarchically-inferior to it. GPI explains why some intransitive Although less clear-cut, Figure 8 echoes results: transitivity predicts the plausibility rules Deep but plausible over ISA links, nr must be a subclass or instance of nr, that is, ISA must point “up.” We relax this for GPI because it is often Figure sitive 9: An intransitive + 10% Chance (50%) -10% have b. structure Figure score which of we ex- was not as low as we expected (43%). Transitive GPI Incon- Rules Intran- We hypothesize that both these effects are due to an unanticipated factor that is raising the plausibility of some but not all of these structure of the rules rule rules. deep have nl Whereas as property form. ble, as illustrated in Figure relation. high We expect plausibility Generalized deep structure, If nl relation n1 is related This definition “up” or “down” for the n3 t n1 HAS- is often plausi- battle inher- 9. In this instantiation, even this structure if they + 10% Chance (50%) -10% to yield relatively are intransitive, and powerful (GPI) comparable Figure plausible of with transitivity: to infer the direction from n1 to n2, whereas to n3 that of h; it can point in property revision with GPI of Figure 8 infer- is a characteristic it is plausible Post-hoc 10: because to n3 by i does not restrict gii d5& 2 2 nal -- from war over a COMPONENT-OF inheritance i, then the same structure to nz by h, and n2 is related is related by any the deep HAS-MECHANISM is a common property have and inconsistent but its conclusion rules with ratings property inheritance ence rule. a rule’s nz weapon” “HAS-MECHANISM the intransitive example, nz, n3 is intransitive, MECHANISM its For COMPONENT-OF rules over ISA links, some but not all of both structures this all our surface inheritance inheritance many common premises. This was an unavoidable consequence of our decision to generate test items randomlv. Four had consistent transitive interpretations, two had consistent intransitive interpretations. 4 Contributors In this section judgments sentation to Plausibility we will discuss of plausibility. is given in [Cohen our goal is to find plausible degradation and who uses these the plausibility bility help the (A more and that detailed Loiselle, inference knowledge factors 19881.) rules that engineers. contribute analysis Recall support Ideally, and to prethat graceful the agent rules should not need much knowledge to judge of their conclusions. For example, the plausi- of the conclusion of the rule n1 m CAUSES n2, CONTAINS nl n3 CAUSES n2 and Cohen and Loiselle 419 seems n3. not to depend In contrast, on the objects to judge that instantiate the plausibility ~1, nz, and of a conclusion of the plausibility 77% and 68% of the time. No knowledge is required to apply these criteria. Greater accuracy requires more knowledge, rule n1 CAUSES rately n.2 knowledge nl is given What elusions n1 and ~2 that about can tell us how likely n2. knowledge of the what factors contributes rules sibility (2’) among components: subject for the total of the con- 1 and 21 Said differently, variance of plau- in judgments T has four additive proportion in subjects’ item variance-the knowledge, experience, moti- between-rule that of T due only instantiate variance-the differences proportion in the surface deep structure variance-the deep of T due only structures structures are then one needs rule-the concepts Similarly, between-rule proportion transitive, to of T due onlv to intransitive, or structure For istics, and intransitive deep structure For our all test predictions of transitive items of implausibility Since variance transitive of T. these T must items with 68% numbers estimates [Hays, these will be p. 4851) frac- character- correct for 77? of intransitive the remaining items. variance in factors. variance are 16% (based on the for transitive rules, 27% for intransitive rules, and 52% for GPI rules. That is, if a rule is transitive, then knowing which rule it is provides little additional information about the plausibility of items. this knowledge accounts for much of the variance SCOTCS of intransitive and GPT items. Estimates ferences a rule’s than of these Conclusion This pa.per suggest,s rules which analyses experiments knowledge surface that from relied purpose. base, variance. our prediction subject Two factors item dation in knowledge and GPI) Knowledge Representation GRANT which KB, our results they in the GRANT are more among of was built are limited general, are common, KB to prove the generality and knowledge their methods engineering. inference to this because the h and because of our results. to support Clearly, rules require graceful these degra- purposes masses are of knowledge We are very encouraged conclusions. for implausible by the relaand than ones. References “I lied about the 19851 Ronald J. Brachman. or, defaults and definitions in knowledge represenAl dlagazine, 6(3):80-93, Fall 1985. [Brachman, trets” tation. [Cohen and Kjeldsen, 19871 Paul R. Cohen and Rick Kjeldsen. Information retrieval by constrained spreading activation in semantic agement, [Cohen derive bases 1988!. plausible and pre- of these correctly networks. and Loiselle, 19881 Paul in for Large Rules 88-54, Processing and Man- 1987. Explorations ference port Information 23(4):255-268, Loiselle. R. Cohen KnowZedge University and Cynthia L. of Plausible Structure the In- Technical Bases. of Massachusetts, Re- Amherst, MA, 1988. [Cohen et al., 19851 Paul R. Cohen, Alvah Davis, David Greenberg, Rick Kjeldsen, Sue Lander, Representativeness and uncertainty Loiselle. Cindy classification rules. predict, systems. AI 6(3):136-149, Magazine, S. and in Fall 1985. [Collins et al., 19751 Miller. and 19731 Willim On of the Tenth Intelliaence. ., [Lenat Using 19871 D. International ,.-pages D. B. Lenat, common ness and knowledge Winter sense Joint Un- Lenat and E. TJl Conference Sciences. 1973. Feigen- Pror~d~r/q on AI lzf~c~ul Milan, , Italv,“. 1987. M. Prakash, knowledge acquisition 1986. B. edition, of knowled~r. 1173-1182, and 1975. for the Social second and M. In D. G. Representation Statistics thrrsholds N. Aiello, knowledge. New York, and Winston, the et al., 19861 CYC: editors, Press, L. Hays. and Feigenbaum, baum. E. Warnock, incomplete A. Collins, Holt , Rinehart, [Lenat from Academic derstanding, [Hays, A. Collins, Reasoning Bobrow dif- the judgement. and Loiselle, for the conclusions (transitivity differences on the accuracy and t are general semantic components, and because transitivity and GPI are common structural characteristics. But further Knowing of its plausibilitv is making we can automatically the relations on the Although we believe relations 6(4):65-85, of plausibility 420 Our for a different However, in plausibility show that are given in ICohen dict judgments structural variance of the remaining improves knowing 5 inference and subject for most instantiation far more Details of item account the Day, Michael and our prediction and subject of between-rule 1973, extent for a large structural for 62% lOO%, of our and to a lesser of GPT items; are not of knowto predict the limit accounts of plausibilitv and will be correct Preliminary represents rules, of a its plausibility. the effect on one’s ability variance be due to the rule, item, w2 statistic predict of test items Subject instantiation represents plausibility. rules, the specific item-to variance to predict for GPI tion in the test plausibility. a.bility to know of individual and not accu- of rules. Ideally, deep structure variance should account for the largest component of T. If 100% of T was due to deep structure variance, then transitivity and GPI would be perfect predictors of plausibility. In contrast, if a large fraction of T is due to item their the contribution rules we could tively high accuracy of criteria that require no knowledge, by the fact that our accuracy is higher for plausible rules structures. variance, specific our predictions. to judge to differences n2, nn in the rule. 7x1, the but because to T, we do not know the limit not met if plausible proportion in the concepts ing the surface estimate about them; Our goal was to develop of T due only to indivitl- and so on. whether knowledge instantiate work is required ual differences GPI plausibility 7 Wt believe our subjects. variance-the vation, to the in Experiments account that our subjects 77,I we need particularly concepts and m,, and M. Shepherd. to overcome bottlenecks. brittle- AI Magazine,