Use of Thematic Information to Speed Search of Semantic Nets

USE OF THEMATIC INFORMATION TO SPEED SEARCH OF SEMANTIC NETS

Lynne M Reder Department of Psychology Carnegie-Mellon University Pittsburgh, PA 15213

John R. Anderson Department of PsychoiOIY Carnegie-Mellon Unlverslt) Pittsburgh, PA 15213

ABSTRACT

A model Is proposed for the representation of thematically lntegratable Information In semantic memory and a process for searching through this memory. We aS$Ume that related facts about an individual are not directly attached to the Individual node, but rather, are' indirectly attached via a thematic sub-node. Thematic sub-nodes allow much·faster search of memory under certain conditions, but not others. We tested these predictions with human subjects and found :o:upport for these notions.

There is little doubt that the time to sear~n a computer data base depends in part on the number of facts stored in it. Similarly, the time for humans to retrieve a specific fact from memory has been shown to depend on the number of other, related facts. A large number of reaction time stu~ies (e.g., Anderson, 1974; Anderson & Bower, 1973; Hayes-Roth, 1977) have found that subjects are slower to recognize a fact the greater the number of facts learned to the same concept. For example, subjects are slower to recognize the fact "The psychologist wanted to go to Japan" if they have also studied "The psychologist thought the research was good" and "Everyone loves Japan" than if no other experimental facts share concept~o .vith the probe. This particular result has been named the /o.n. el/ect because latencies are found to be a function of the number of propositions fanning out of a given concept node in a propositional network representation (see Anderson, 1976). (Subjects are also slower to reject unstudied probes as a function of the fan of their concepts.)

There is a paradoxical aspect to the result that the more facts a person knows about a concept the slower he or she is to verify any one of them. It seems intuitive that experts (a) know lots of facts about certain topics, and (b) can answer more questions on those topics and answer them faster than anyone else. However, certain situations can be pointed out where experts seem to experience a type of interference that results from too much similar informatior.. Learning a new text editor in many ways is much easie,. if one has used a similar editor

This research was supported in part by NSF grants BNS76-00959 and 8NS78-17463 to John R. Anderson and by an NIMH post-doctoral fellowship to Lynne M Reder. We would like to thank T. lwasawa for assistance in programminp and in running subjects.

before. However, the more editors a person uses, the harder it is to keep the commands straight.

Even though experts may sometimes suffer more interference than non-experts, It also seems that if their Information can be i.ntegro.tecl Into various themes, then somehow the information should be less interfering. Consider the following knowledge representation for facts about an individual and process model for searching this knowledge: all facts studied about the individual are attached to that Individual's concept node. However, propositions that can be considered to be thematically related are not attached directly to the person node1 rather, they are attached to a theme node that Is attached to the person node. Figure 1 gives a representation for the following four sentences: Melvin croued his skis. Melvin shcused down the sk>pe. Melvin wo.shed his co.r. Melvin lfOt on the cha.ir li.ft. Facts that can be considered as having to do with Melvin-skiing form a sub-node e~ttached to Melvin. Facts that are not so related are attached to Melvin directly. The /o.n out of the Melvin concept, then, is two rather than four.

The search mechanism works as follows: The probe sentence is parsed. Each concept of the probe is activated In memory and activation spreads out along each arc attached to the concept. Seorch energ~ is assumed to be a constant for each activated concept. How much activation or search energy a particular arc receives Is a function of the total number of arcs. emanating from the concept (the amount of fan).

ln earlier experiments on the fan effect, the facts associated to a person were unrelated. In those situations, the search mechanism predits a simple effect of number of facts on time to recognize a probe. However, our predictions are more complex for the representation in Figure 1. We would predict that the time to verify Melvin wo.shed his cor would be the same 41s if we had learned only two unrelated facts about Melvin. Our predictions about Melvin cro1sed hi.· 1lcis are somewhat more complex.

The fan out of the Melvin concept is two. The fan out of the Melvin-skiing sub-node is three. Whether there is en effect of sub-node fan when attempting to recognize one of several thematically related facts depends on the natur .. of the foi.l.3 or false probes. If, for example, the probe to be rejected is Melvi.n bo.lced o. co.lce, search need not continue past the Melvin-skiing sub-node: it is clear to the subject that cake-baking is not part of the skiing theme. If, on the other hand, the probe to be rejected is Melvin ,.rfected his stem-clariltiel, search must proceed beyond the sub-node to inspect each proposition attached to It until the correct one Is found or ell stored propositions are exhausted.

EXPERIMENT

To verity this model* we conducted an experiment where subjects learned facts about various characters. These facts belonged to various themes, e.g., buying a car, jogging, going to a circus. A particular character could have one or two themes a~ciated with him or her and one or three facts associated with a given theme. For example a character with six associated facts would have three facts about each of two themes. A character with· four facts would have one fact from one theme and three facts from the otl1er. Ench condition in the experiment can be described by two digits, representing the number of facts in each ineme associated with a given character. The conditions are (l,Q),(l,l),(3,0),(3,1),(3,3). The O's denote conditions where there was only one theme.

Subjects studied these character-predicate pairings until they knew the facts perfectly (could recall all the facts about each character several times in succession). After this phase was completed, subjects went into the critical phase in which reaction times were collected to answer questions about the material studied. One sentence was. displayed on the screen at a time. The subject was told to press the yes or no response key depending upon whether the subject thought it had been studied during the acquisition phase. Test probes were presented in random order. The foils (sentences subjects had not studied) were of two types and were separated into different test blocks. One type was the themo.tico.Uy rel4ted foil (e.g., Melvin enjo~ed deep powder), and the other type was the unrelo.ted foil (e.g., Melvin bo.lced o. cake). Both types of foils have the. property that their predicates (verb-phrase) were studied with another

(b)

T\-\EM~140e

1303~

\INRE\.~ 116 !5

1076 _.:;.;..---

I

1

2

I

NUMBER OF THEMES

Fi&ure 2: Mean reaction time to judge that a probe had been studied as a function of number of themattc.lly related facts (graph a) and as a function of number of themes (~raph b).

·we ectuelly t .. tect two ellernelive modele in lhie ··periment, but epace conelreinle require thie eimplificelion.

character during the experiment. This ensured that subjects could not reject a probe simply on the basis of lexical familiarity.

We had two principal predictions for this experiment. First, when the foils are unrelated, there should be a much smaller effect of number of facts studied. This is because subjects should be able lo make their judgments just by inspecting the theme subnodes.** They would not have to search the facts attached to these nodes. Our second prediction was that there should be an effect of themes (more themes means more arcs out of the person node), but that it should be small since subjects can quickly reject the wrong theme sub-node, and not be affected by type of foils since the extra sub-node is rejected in the same fashion regardless of foil-type.

RESULTS

Figure 2 displays the data (latencies in millis~conds to decide whether the probe had been studied or not) partitioned two ways: on the left, the data are analyzed as a function of the amount of fan relevant to the probe that is attached to the character in question. For example, the relevant fan for a Melvin skiing probe would be three because the subject studied three Melvin skiing facts. If the probe had concerned Melvin and his car, the relevant fan would have been one. On the right, the data are displayed as a function the number of themes a!Ssociated with the character in the test probe, which can be one or two. Melvin has ·two themes, skiing and car washing. (There is no sub-node for car-washing because only one such fact was studied.)

Each graph plots two functions, one labelled unrelated and one labelled thematic. As explained above, the unrelated-thematic factor reflects the type of foils from which the subject had to discriminate the studied probes. The left panel plots the data as a function of the number of facts involving the same theme as the test probe. Subjects take longer to verify that a fact was studied the more other related facts were also studied. However, this fan effect is severely attenuated when the foils are not thematically related. Moreover, as predicted, subjects respond much faster in general when the foils are unrelated. We tako this to mean that subjects have a less stringent criteorion of saying "yes, this sentence is old" in these blocks. Subjects only search as far as the sub-node before responding when foils are not related to the sub-node. Smith, Adams, and Schnorr ( 1978) have also found that the fan effect is reduced with unrelatec foils. (However, they did not have the other conditions of this experiment viz., testing with thematic foils, and manipulating the number of themes associated with a character.)

··Jf there ·· only o..., theme feel, we .. aume aiAijecta do not bother to form 1 aubnode. lnat .. cti"1 .,..., propoailion ia problbly equiv·lent to lnepectl"1 · theme aub-·node.

The data on the right shows the effect of number of themes. Here too there is an effect of fan for themes, however this effect is smaller and is not affected by the related/unrelated foil block distinction. The lack of interaction of theme-fan with foil-block type was predicted. Regardless of type of foil , the subject must at least find the relevant sub-node before responding "old". If there is only one theme associated with a character, the fan off the character node will be less. Therefore, search should be faster than with two sub-nodes or a fan of two off the character node.

CONCLUSION

The data support the view put forth in the introduction that facts are not necessarily attached directly to the concepts in their propositions, but rather can be organized into thematically consistent subnodes. It seems to us that, in general, sub-nodes in memory serve several functions. First, ~ub-nodes make plausibility judgments easy. Typically people are not asked to discriminatfl thematic foils from related statements; normally one would accept as plausible other related skiing facts about Melvin (see Reder, 1979, for a fuller discussion). Second, by having many sub-nodes in memory, the number of propositions that must be inspected is smaller. If a subject is searching for a fact about Melvin baking a cake, she or he need not search all of the skiing facts to find it--search stops on a wrong path as soon as the sub-node is reached.