Representing Regularities in the Metaphoric Lexicon 
James H. MARTIN 
Berkeley Artificial Intelligence Research 
Computer Science Division 
University of California 
Berkeley, California 
94720 
Abstract 
This paper describes a system for representing knowledge 
about conventional metaphors for use by natural language 
analysis, generation and acquisition systems. A system of 
hierarchically related structured associations is used. These 
associations are implemented as a part of the KODIAK 
representation language. Particular attention is paid in this 
paper to representational mechanisms that can capture general- 
izations over the system of conventional metaphors as a whole. 
1. Introduction 
The most commonly occurring words in English display a 
wide range of metaphorically motivated meanings. A 
knowledge representation system capable of representing these 
meanings must satisfy two competing requirements. The first 
requirement is that it must be capable of representing the 
specific details of each lexicalized metaphor. More precisely it 
must provide enough knowledge to permit the easy recognition 
and interpretation of a particular use when it is encountered. 
The second requirement is that the representation must capture 
the relevant generalizations manifested by the system of meta- 
phors underlying the lexicon taken as a whole. The system 
therefore must represent both highly specific facts and capture 
systematic generalizations. 
2. Conventional Metaphors 
A conventional metaphor (Lakoff and Johnson 1980) con- 
sists of source and target concepts. The target concept is the 
concept that is actually under consideration. The source 
specifies the concepts in terms of which the target concept is 
being viewed. When we refer to metaphoric knowledge we 
mean explicit knowledge of the conventional associations 
between various source and target domains. These 
associations are a conventional part of the language. 
The first requirement therefore is that the representation 
must be able to capture individual metaphors as concepts con- 
sisting of sets of associations between source and target con- 
cepts. Consider example (1), 
(1) How can I kill a process? 
This example, from the UNIX domain, involves the conven- 
tional metaphor that to kill an ongoing process means to ter- 
minate it. The target concepts involve computer processes and 
the actions that terminate them. The source concept is that of 
the action of causing a living thing to die. The metaphor con- 
sists of the source, target and the set of associations linking 
them. The associations specify the correspondences between 
the source and target concepts. In this example they must 
specify that: the concept kill corresponds to terminate, the vic- 
tim of the killing is the process terminated and that the killer is 
the agent of the termination. 
3. Metaphorical Regularities 
It is not enough for the representation to simply account 
for each individual conventional metaphor in the lexicon. The 
set of metaphorical word senses in the language taken as a 
* "Ihil research was sponsored in part by the Defense Advance Research Plojeets Agency 
(DOD), A1pa Order No. 4871, Monitored by Space and Naval Warfare Systems Command 
under Contract No. N00039-84-C-0089. 
whole exhibits a number of regularities that must be accounted 
for. This paper will consider two of the most important regu- 
larities. The first is the metaphoric use of different words 
motivated by a single underlying metaphor. The second regu- 
larity has to do with the use of the single words with metaphors 
that exhibit various degrees of similarity. 
Consider the following examples. 
(2) Mary has a cold. 
(3) John gave Mary a cold. 
(4) Mary got a cold from John. 
The various uses of the verbs have, give and get in these exam- 
ples are not instances of isolated metaphors. They all ,arise 
from a common conventional metaphor centered around the 
idea of state as possession. This metaphor entails to be in an 
infected state is to possess a disease. The transfer metaphor 
underlying (3) and (4) is a natural extension to the core posses- 
sion metaphor. In the source domain, the concepts giving and 
getting are events that result in a having. In the same way the 
concept infecting leads to an infected state. Therefore the 
metaphor infection-is-transfer follows coherently from the core 
possession metaphor and the structure of the source and target 
concepts. The representation should efficiently capture the fact 
that coherent underlying metaphors give rise to surface 
metaphors that may involve several closely connected words. 
Now consider these examples. 
(5) John gave Mary a cold. 
(6) John gave Mary the flu. 
(7) Mary gave John a good idea. 
(8) Mary gave John a cold. 
The similarity of examples (5) and (6) is clearly a result of the 
application of the same metaphor to two closely connected tar- 
get concepts (infection with either a cold or the flu). The 
transfer metaphor in this case is at a level of abstraction that 
subsumes both of these examples. The representation must 
therefore permit the representation of metaphors at an 
, appropriate level of abstraction. 
Examples (7) and (8) are however not in the same kind of 
close similarity relationship. The metaphors underlying these 
examples have completely dissimilar target concepts. 
Nevertheless they do exhibit an underlying structural similar- 
ity. In both cases an abstract effect upon a person is viewed as 
a transfer of an abstract concept to the effected person by the 
agent responsible for the effect. A fine grained representation 
involving the hierarchical sharing of subparts will be used to 
capture this more distant similarity and still permit the 
representation of the distinct target concepts. 
The following sections will introduce the various 
mechanisms that will be used to both represent the specific 
details of individual metaphors and capture the necessary gen- 
eralizations and regularities. 
4. Representational Details 
The representation language used to capture conventional 
metaphoric knowledge is KODIAK (Wilensky 1986). 
KODIAK is an extended semantic net language in the spirit of 
KL-ONE (Braehman 1985). The use of structured associations 
396 
as described here for representing metaphoric knowledge has 
been successfnlly applied to the following tasks: natural 
language generation (Jacobs 1985), text understanding (Norvig 
1986, Jacobs 1987) and the acquisition of new word senses 
(Martin 198'7). 
In the following sections, pieces of a KODIAK 
knowledge base are shown with a graphic representation. For 
the purposes of this paper it is sufficient to understand that rec- 
tangles represent coherent concepts that can stand on their 
own. These are called absolutes. Concepts in circles are called 
asoectuals. Asnectuals are contents that are meaningful only 
relative to a related absolute. Absolutes are connected to thelr 
aspectuals via an S (Slot) link. Aspectuals are given type 
information via the C (Constrains) link that indicate the kind of 
concept tha~ can fill this aspectual. Arcs labelled D (Dominate) 
are a kind of ISA oi" inheritance link. For a more in-depth 
description of KODIAK and this graphical representation see 
Norvig (1987). 
4.1. Representing Single Metaphors 
Figure 1 shows the KODIAK representation of tile source 
domain frora (1). It states that a killing is a kind of action 
with a result that is a death-event which is in turn an 
event. The kili-victlm of the killing is an inherited role 
from action indicating that the kill-victim is effected by 
the action. The kill-victim is constrained to be a Living- 
thing and the killer must be an anlmate-agent. Finally 
the equate links require that the kill-victlm must be the 
same as the tiler participant of file death-event. 
< / 
Figure h Killing 
Figure 2 shows the corresponding concepts from tile tar- 
get domain. It states that a terminate-action-process is a 
terminate-process which is a kind of action. The 
termlnated-proeess role is an inherited role specifying the 
patient of the action. Tile result of the action is a terminate- 
process-effect which is a kind of terminate-event. 
Finally the terminated-process is equated to the 
terminated-process-event of the terminate-process- 
effect, This is analogous to the relationship between the 
kill-victlm and the diet shown in Figure 1. 
What is needed is a way of associating the appropriate 
source and larget concepts. Such an association is realized in 
KODIAK by using a relation called a metaphor-map. A 
metaphor-map is a kind of VIEW relation whose aspectuals 
specify con'esponding source and target concepts. A VIEW is 
a primitive KODIAK relation that permits a limited inheritance 
between concepts without requiring a strict ISA relationship. 
Metaphor maps are needed to link all the core source con- 
cepts in Fignre 1 to their counterparts in the target domain. In 
particular the killing maps to the terminate-action, the 
kill-victim maps to the terminated-process, the killer 
maps to the actor of the terminate-action and the result 
of the killing maps to the result of the terminating. It 
is the co-occurrence of all these maps that constitutes the con- 
ventional metaphor that terminating a process can be viewed as 
\[ Terrmn!)Acuoa S 
\[ Process \] 
Figure 2: Terminating 
a killing. 
This conventional co-occurrence of a set of metaphorical 
associations is called a metaphor-sense. A metaphor-sense is 
simply a KODIAK concept that ties together a coherent set of 
metaphorical maps. A metaphor-sense can participate in an 
unspecified number of component-map relations. A 
component-map relation is a kind of component-association 
relation that holds between a metaphor-sense and a metaphor- 
map. Figure 3 shows the metaphor-sense kJ_3.1-term±nate- 
metaphor that ties together all the metaphor-maps underlying 
example (1). 
To a significant extent metaphor-senses are the minimal 
meaning-bearing unit of conventional metaphors. Metaphor- 
maps represent the building blocks out of which meaningful 
metaphor-senses are constructed. The metaphor-sense 
represents the level at which one would say that there is a con- 
ventional metaphor that to terminate something is to kill it. 
This level of representation will frequently con:espond to a sin- 
gle metaphorical word sense. 
4.2. Representing Extended Metaphors 
As illustrated above with examples (2) through (4) it is 
frequently the case that a single core metaphor is the basis for 
the extended metaphorical use of a number of related words. 
Consider the following exmnples. 
(9) John gave Mary a cold. 
(10) Mary has a cold. 
A metaphor-sense accounting for (9) with the concepts under- 
lying tile word give is shown in Figure 4. This metaphor-sense 
represents the conventional metaphor that to infect can be 
viewed as a transfer. The relevant maps to this discussion are 
the givee-infected, given-cold, and the give-res-inf- 
result mappings. These mappings represent the ideas that 
recipient of the giving is viewed as the newly infected person, 
' the given thing is viewed as the cold and the result of tile giv- 
ing is the result of the infecting action. 
Now consider the schema underlying (7) as shown it, 
Figure 5. In this case the haver is the person with all infection, 
the cold is the th!ng had and the having corresponds to the state 
of being infected. 
Figure 6 shows the combined source concepts from the 
have-a-cold and give-a-cold schemas. This diagram is 
adapted from Norvig (1987). In it we can see how the source 
concepts from these two schemas are closely connected. In 
particular: the giving-result is constrained to be a kind of 
597 
Figure 3: Kill-Terminate.-Sense 
having (one that results from a giving), the givee of the giv- 
ing is equated to the hayer of the givee-has-given hav- 
ing. Finally the given-thing is equated to the had of this 
same having. Therefore all of the source concepts of the 
have-cold schema are either equated to or constrain concepts 
in the give-cold schema. A similar analysis follows for the 
target domain of infection. 
"~ ....... i ..... Id /\[ infec!-wilh-cotd 
~ give-infect-cold ~// ~ 
,,vor ,o0o  ?" 
givc-res-inf-res 
Figure 4: Give-A-Cold 
Having }~ H.ve.C.1,L_.\]~/~ B¢ing-lnf~ed \] 
~ "~s 
Figure 5: Have-A-Cold 
A given metaphol~cal use must meet the following 
requirements to be considered an extended use of a more tim- 
damental core metaphor. 
• The extended sense must contain all the core metaphori- 
cal associations as a subset of its own associations. 
398 
• The source and target concepts of the additional associa- 
tions must be closely connected to the source and target 
concepts specified by the core associations. 
• The new source and target concepts must be connected to 
the core source and targets in a coherent manner. 
Figure 7 shows the extension relationship between the 
core metaphor have-cold and tile extended sense give-cold. 
In this diagram we see that the component metaphor maps 
from the core sense are shared and specialized by the extended 
sense. This section will consider exactly how this instance of 
an extended sense meets each of the given extension require.. 
ments. This example will be used to illustrate in general the 
mechanisms that are used to accomplish metaphor extension. 
Figure 6: Giving and Having 
The rfietaphor-map representing the fact that a haver plays 
the role of the infected, the hayer-infected map, appears 
twice in the extended sense, once as the giver-infector and 
again as the givee-infected. The giver of the giving maps 
to the infector; this matches the core map since the giver is 
equated to be a kind of haver. This core map is "also used to 
represent the fact that the givee is also this kind of hayer as a 
result of the giving. The next map to consider from the core is 
the having-infected map. This map also appears twice in 
the extended sense. The first use is in the give-res-inf-res 
map. This map indicates that the result of the giving, a having 
by the givee, plays the role of the result of the infecting, an 
infection of the infected person. It is therefore a specialized 
map from a having to an infected state. It appears again in the 
....... _ J .... _G~-c~m____ ....... Ii~ze~Jl d_ .... 
'- ............ \[ J 
"-<-__ -- 
Givell-Cold 1 
Figure 7: Flaving Extending to Giving 
.qive-pre--\[nf-pre metaphor map. This map indicates that 
yea have to have a cold before you can give it to anyone. This 
is again basically a map from a having to an infected state. 
The final map to consider is tile core had-cold metaphor-map. 
This map appears once in the extended sense as the given-- 
co.l.d metaphor. 
The links between the shared maps in the extended sense 
and the malls in the core sense are shown simply as individaal 
D links. In fact the specialization of a metaphor-map is a more 
complex suucturcd association involving the specialization of 
each of the components. The tbllowing section on metaphor 
hiermvhies will go into the details of how this is actually 
accomplished. At this point it is sufficient to say that the first 
extension requi0ement is satisfied if all of the metaphor-maps 
in the core ¢xist in the extension as specializations. 
The second requireruent tbr extension is that tile source 
and target components of the extended sense must be closely 
connected to die concepts specified by the core. Figure 6 shows 
that this is clearly fulfilled in this case ill the source domain. 
The concepts in the source domain of tile extended sense are 
all either exlnated to, imruediately dominated by or related by a 
single relation to the concepts in the source of the core domain. 
The same is true in the target domain. In general the second 
reqnirement will be satisfied if the extended concepts are 
related to the core by these types of close association. 
The final reqnirement speeifies that file extended concepts 
be consistent extensions from both the source and target 
domain of the core metaphor. This is met in this case by rite 
result relations in both the source attd tat'get. On the source 
side the giving is related to the having fi'orn the core by a result 
relation. This is shown in the abbreviated slot form in Figure 6 
by the cj:~_vinw,-result slot. This relation is inherited from 
the more general category of action and indicates that a resnlt 
of a giving is a having. The satne analysis holds for the target 
domMn. The result of an infecting is a state called being- 
infected. These states are shown as the target concepts ill 
Figures 4 arm 5. The last requh'ement is satisfied if the same 
relationship holds between the core m~d extended concepts in 
both the source atld target domains. 
4.3, Metal~horlcM Hierarchies 
Metaphor-maps and ruetaphor-senses are both fifll- 
fledged KODIAK concepts and can therefbre be arranged in 
abs~action hierarchies. The use of hierarchies will be the pri- 
mary mechanism nsed to aeeom,t for the similarities and differ- 
ences among conventional ruelaphors. This use of abstraction 
hierarchies to capture regularities in the lexicon is similar to 
that described in Fass (1988). The following section will show 
how the various similarities; and differences among conven- 
tional metaphors follow directly froru the relations among 
metaphor-.m!:,ps. 
Consider again the k.t.tl-proeess map from (i). This is 
a mapping from a concept constrained to be a living thing to a 
tat'get concept that is not a living thing. This is a manifestation 
of a more general metaphor that allows us to view non-living 
things in W, rms of living things for the purpose of explaining or 
tmderstanding fl|eir behavior in terms of living things. Exam- 
ples (11) through (14) from Lakoff and Johnson (1980) all con- 
tain specialized instances of this general metaphor. 
(11) Inflation is eating up our savings. 
(12) Those ideas died a long time ago. 
(13) He is the father of modern quatltum mechanics. 
(14) Those ideas won't yield any fruit. 
Exatnple (11) is motivated by the metaphor that tile reduction 
in savings caused by inflation can be viewed as inflation con- 
suming the savings. Inflation is therefore viewed as an animal 
that can consume things. Example (12) contains a metaphor 
dealing with the duration of ideas. When an idea is no longer 
held or believed it has died. At a more specific level we have 
an idea filling the role of the diet of a death-ev~.~nt as in 
Figure 3. This role has the similar constraint as the kill- 
victim of (1) that it be a liv.t.ns-thinq. In a similat" way as 
(1) a set of interrelated rnctaphor-maps link the source domain 
of death to the target domain of ideas and their duration. In 
particular there is a metaphor-map from the dier to a concept 
that is not alive. 
Example (13) contains the notion that the creation of an 
idea is a birth-event attd that the originator of the idea plays the 
role of the father in the birth event with the clvated idea play- 
ing role of the child. Once again in this metaphor there is a 
metaphor-map from a concept that is not a living thing (the 
created idea) to a role that must be one (the child being born). 
This metaphor-map, however, is more specific since tile con- 
straint is not only to be a living thing but to be truman. Finally 
example (14) contains the notion that an idea can produce new 
ideas. This is metaphorically structured as a plant producing 
new fi'uit. In this case an idea is viewed as a specific kind of 
living thing, a phmt, 
What (1) and (11) through (14) till have in common i,; the 
idea that an abstract concept like a process or idea can be 
viewed as a living thing to explain some aspect of its nature. 
They differ in the particular kind of living-.thing that is used 
and in the role that it plays. These similarities and differences 
result in specific metaphor-maps in each particular case. What 
is needed is a mechattism that can captm'e the commonalitics 
and differences atnong these various metaphor-maps. 
This mechanism is provided by tile general inheritance 
mechanisms provided by KODIAK. Metaphor-mal~s that share 
properties are dominated by more absu'act parent maps thai 
capture the commonalities among the children. The source and 
target aspeetuals of the parent map are constrained by concepts 
that are more abstract than and dominate the constrainers on 
the children's source and target aspecmals. Figure 8 illustrates 
this situation with the maps from examples from (1) and (11) 
through (14). 
The top half of Fignre 8 shows the hierarchical relation- 
ships among the maps underlying the above examples. They 
all converge on the abstract ruetaphor-map representing the 
idea of viewing a non-living-thing as a living-thing. The two 
metaphor-maps in the dotted box are expanded in the bottom 
half of the diagram to show the exact details of the inherita nee 
links. In this expansion we can see that the idea-a~'~-- 
living-thing metaphor-map dominates the forgotten- 
idea-as-diet map. In the parent map the idea-as~target 
aspeetua\] is constrained to he an idea. The living-thing- 
as-source aspeetual is constrained to be a llving-thing. 
In the forgotten-idea-as-dier we see that the inherited 
source aspectual is specialized by being consu'ained by the 
diet aspectual of the death-event. The inherited target 
aspectual is fitrther specialized by being constrained by the 
399 
Uon \[ 
I , . ~ . , , ^c , , As , (6) 
IFor.otten Ideal ~ ? ~ -- l~rmlnated-l'roc~tss 
• ,~s (4) I~~~~ "~ L--s~a _J 
Idea \[ 
~"NC Idea-As -Livln g-Thing_ 
D 
K~ For ~ot ten-Idea-As-Diei. 
~ Thing 
Y 
Figure 8: Metaphor-map Hierarchy 
forgotten-idea aspectual. The constraining aspectuals, 
forgotten-idea and diet satisfy the appropriate inheritance 
restrictions since they in turn are constrained by idea and 
living-thing respectively. It is typically the case that the 
source and target aspectuals of metaphor-maps at the bottom of 
the metaphor hierarchy are constrained by aspectuals of some 
specific context. More abstract metaphors tend to have their 
aspectuals constrained by less specific absolutes. 
Careful examination of Figure 8 reveals that the 
metaphor-maps are actually being specialized in different 
ways. Consider the process-as-living-thing specializa- 
tion of the non-living-thing-as-living-thing. In this 
case the metaphor-map is specialized on the basis of the target 
concept. The generic category non-living-thing is replaced by 
process. In the case of the non-living-thing-as-plant 
mapping the specialization is done on the basis of the source 
concept. The metaphor-map hierarchy may therefore contain 
specializations based either the source or target aspectuals or 
both. 
Metaphor-maps may also participate in multiple inheri- 
tance. Consider the idea-as-aniraal and idea-as-plant 
maps, These maps inherit their respective specializations from 
multiple parents. The details for the idea-as-animal map 
are shown in Figure 9. In this diagram we can see that this 
map has two parents the idea-as-living-thing map and the 
non-living-thing-as-animal map. Its source and target 
aspectuals are each dominated by the both of the corresponding 
aspectuals from each parent. The multiple-inheritance 
mechanism assigns only the most specific constrainer to the 
lower child aspectuals. In this case the source aspectual is con- 
strained to be an animal, which is the more specific than 
living-thing. In the same way the target is constrained to be 
an idea rather than the more abstract non-living-thing. In 
this way the idea-as-animal map is created by specializing 
different parts of the two parent maps. 
Figure 9: Multiple Inheritance 
4.4. Representing Metaphorical Similarities 
Consider the relationship between the metaphors underly- 
ing (15) and (16). 
(15) Mary gave John a good idea. 
(16) Mary gave John a cold. 
These examples do not involve metaphors with overlap- 
ping scope nor are they related in a strictly hierarchical 
fashion. Nevertheless they do to have a structural similarity 
that should be reflected in the representation. 
In particular note that both contain a map that represents 
400 
the notion that something that is not an object can be treated as 
one, (a disease and an idea.) Both contain the notion that 
transfer represents some change of state with respect to the 
metaphorical object. The point of these examples is that the 
metaphor schemas underlying examples (15) and (16) contain 
individual component maps that are similar. This similarity is 
accounted for by the hierarchical relationship among the com- 
ponent maps in each schema. Figure 10 shows some of the 
relevant relationships between these schemas that accounts for 
their similarity. (Note that for the sake of readability most of 
the maps in each schema are omitted from this figure.) 
This figure illustrates the building block relationship 
between metaphor-maps and schemas. In the give-an-idea 
schema two maps are shown. The give-tell map is dom- 
inated by the communicate-as-transfer map which in turn 
is dominated by the act-upon-as-transfer (Jacobs 1985) 
map. The act-upon-as-transfer map is a map representing 
a general metaphor that allows abstract events with participants 
to be viewed as transfer-events where some participant 
receives or donates some metaphorical object. The given- 
idea map is dominated by the idea-as-object map which is 
in turn donfinated by the non-object-as-object map. This 
metaphor map reflects the widespread use of reification in our 
conceptual system. 
Similarly in the give-a-disease schema two maps are 
shown. The given-disease map is dominated by the 
disease-a:3-objeet map. This in turn is a sibling of the 
idea-as-ebject map since they both have the non-object- 
as-object mapping as a parent. Finally the qlve-infoct 
map is dominated by the infect-as-transfer map which 
like the co~nmunicate-as-transfe~: map is dominated by the 
act-upon-as-t r~ns fer mapping. 
Figure 10 illustrates the somewhat more distant relation- 
ship between the schemas underlying (15) and (16). They 
involve the same source domain with completely separate tar- 
get domains. The domains of communication and infections 
are not related in any meaningful way in the hierarchy. They 
do not sham any close meaningful parent categories. For this 
reason there: can be no abstract metaphor schema dominating 
both-of the schemas in question. However the similarity of the 
two schemas is accounted for by the hierarchical structure of 
the component maps and the sharing of maps with common 
parents. 
5. Summary 
An adequate representation of knowledge about the con- 
ventional metaphors in a language must meet two require- 
ments. It must capture the exact details of individual meta- 
phors to permit efficient recognition and interpretation. In 
addition it should reflect the regularities of the system of con- 
ventional metaphors as a whole. In particular the extension of 
core metaphors to extended senses should arise naturally from 
the structure of the metaphor and the structure of the source 
and target domains. This paper has described a system of 
representing conventional metaphors that efficiently meets 
these requirements. 
o,j0o.l I Ao, o .A, Tr   or\] 
Figure' t0: Giving Metaphors

References 

Brachman, J. & Schmohe, J. G. (1985) An overview of tile 
KL-ONE representation system. Cognitive Science, 9, 
171--216. 

Fass, D. (1988). Collative Semantics: A semantics for natural 
language processing., Report No. MCCS-88-118, Com- 
puting Research Laboratory, New Mexico State Univer- 
sity, Las Cruces. 

Jacobs, P. (1985). A knowledge-based approach to language 
production. Doctoral Dissertation, Computer Science 
Division Report No. UCB/CSD 86/254, University of 
California, Berkeley, CA. 

Jacohs, P. (1987). A knowledge framework for natural 
language analysis. Proceedings of the Tenth lnterp, a~ 
tional Joint Conference on Art~cial Intelligence. Milan, 
Italy: Morgan Kaufmann. 

Lakoff, G. & Johnson, M. (1980). Metaphors We Live By. Chi- 
cago, IL: University of Chicago Press. 

Martin, J. (1987). The acquisition of polysemy. Proceedings 
of the Fourth International Workshop on Machine Learn- 
ing. Irvine, CA: Morgan Kaufmann 

Martin, J. (1987). Understanding new metaphors. Proceedings 
of the Tenth International Joint Conference on Artificial 
Intelligence. Milan, Italy: Morgan Kaufmann. 

Norvig, P. (1986). A unified theory of inference for text under- 
standing. Doctoral Dissertation, Computer Science Divi- 
sion Report No. UCB/CSD 87/339, University of 
California, Berkeley, CA. 

Wilensky, R. (1986). Some problems and proposals for 
knowledge representation. Technical Report No. 
UCB/CSD 86/294. Berkeley, CA: University of Califor- 
nia, Berkeley, Computer Science Division. 
