Collative Semantics 
Dan Fass 
Computing Research Laboratory 
New Mexico State University 
Las Cruces, NM, USA 88003 
O. Abstract 
This paper introduces Collativc Semantics (CS), a new 
domain-independent semantics for n~tural language processing 
(NLP) which addresses the problems of lexieal ambiguity, met(> 
nymy, various semantic relations (conventional relations, 
redundant relations, contradictory relations, metaphorical rela- 
tions and severely anomalous relations) aud the introduction of 
new information. We explain the two techniques CS uses for 
matching together knowledge structures (KSs) and why seman- 
tic vectors, which record the results of such matches, arc infor- 
mative enough to tell apart semantic relations and be tile basis 
for lcxical disa.mbiguation. 
1. Introduction 
Collative Semantics is a new domain-independent sem,~n- 
tics for NI,P. This paper foeusses on CS, describing tile main 
issues it addresses (lexical ambiguity, mctonymy, semantic rela- 
tions, introduction of new information) and general details of 
its knowledge representation, knowledge structures, techniques 
for matching together knowledge structures, and the way it 
records the results of matching in semantic vectors. 
CS has l)een implemented in a natural language program 
eaiicd recta5 which has been described in detail clsewherc (Fass, 
1986). Briefly, the program produces semantic representations 
of single sentences requiring lexical disambiguation and met(> 
nymic infereneing. While processing such sentences recta5 com- 
putes semantic vectors for the matches between pairs of word- 
senses which are rich enough to discriminate between instances 
of conventional relations, semantically redundant relations, con- 
tradictory relations, metaphoricM relations and severely 
anomalous relations between word-sense pairs. Semantic vectors 
also record the assertion of new information. Meta5 can process 
all the example sentences given in the next section. 
2. Semantic Relations, Novel Information and Meto- 
nymy 
In this section we provide brief descriptions of those 
semantic phcnomcna considered by CS, except lexical 
ambiguity, starting with semantic relations. Conventional, 
metaphorical and severely anomalous relations can all be 
described using the terminology of Richards (1936). The subject 
term is the "topic," the term it is compared to is the "vehicle," 
the similarity or resemblance between them is the "ground," 
and any difference or dissimilarity is tile "tension." We also 
adopt Perrine's (1971) four-fold classification of metaphors into 
combinations of explicit and implicit topics and vehicles. 
In a metaphorical relation there is tension between the 
topic and vehicle because the topic term is not a type of vehicle 
term. What is salient (Ortony, 1979) is given by the context, a 
salient property of the vehicle is found and an analogical match 
discovered between it and u property from the topic. The 
remaining properties of the topic and vehicle have similarities 
and differences and, although these are not central to recognis- 
ing the metaphorical relation, the higher the proportion of 
differences to similarities the "better" the metaphor. 
For example, in the metaphorical relation between 'cal" 
and 'drink' ill 
(1) "The car drank gasoline." 
the topic is 'ear' and the implicit vehicle is 'animal', the agent 
preference Wilks, 1975) of 'drink'. The tension is caused by a 
car (the topic) not being a type of animal (the vehicle). What is 
salient in tiffs context is the action of drinking, given by the 
main sentence verb. The salient property of the vehicle is one 
referring to the salient action, i.e. that an animal dl'inks pot- 
able liquids. An analogical match is found between tile salient 
property of the vehicle and a property of the topic : animals 
drink potable liquids as cars use gasoline. The ground is the 
expending of liquid. Matching the remaining properties of the 
topic and vehicle, some pairs of properties are the same 
(animals and ears are both bounded, three-dimensional and 
solid) but other pairs express differences (animals are living, 
cars are nonliving; animals are composed of flesh, cars are made 
of metal). 
Ill a conventional rclation there is no tension because the 
topic term is a type of vehicle term. The sMient property of tile 
vehicle matches identically with a property from tile topic.. 
There is a high proportion of similarities to differences amongst 
matches of other pairs of properties of the topic and vehicle. 
For example, ill the conventional relation between 'man' 
and 'drink' in 
(2) "The mall drank beer." 
the topic is 'man' and the implicit vehicle is again 'animal', the 
preferred agent of 'drink'. A man (the topic) is type of animal 
(the vehicle). What is salient is drinking so the salient property 
of animals is again that they drink potable liquids. An identical 
match is found between animals drinking potable liquids and 
men drinking potable liquids. Of the remaining properties of 
animal and man matched together, a large proportion are simi- 
lar and very few are different. 
A severely anomalon~'~ relation has the same tension 
between topic and vehicle a~ a metaphorical relation : the topic 
term is not a type of vehicle term. A salient property is found 
in the vehicle but it does not find an identical or analogical. 
match with any property from the topic. A high proportion of 
the remaining properties of the topic and vehicle matched 
together are different and few are the same. 
Our description of semantically redundant and contradic- 
tory relations is based on Katz's (1964) definitions. Whether the 
assertion of a particular property onto a subject term is a 
redundant relation, contradictory relation or new information is 
a function of tile knowledge already in the term's dictionary 
entry. 
A contradictory relation is one in which a modifier asserts 
a property onto a subject term which is incompatible with a 
property already in the subject term, e.g. 
(3) "John McEnroe is female." 
where the assertion \[sexl, femalel I from the adjective clashes 
341 
with the property \[sexl, malel\] already in the dictionary entry 
for John McEnroe. 
A semantically redundant relation is one in which a 
modifier asserts a property onto a subject term which is identi- 
cal to a property already in that term, e.g. 
(4) "John McEnroe is male." 
is redundant if the dictionary entry for John McEnroe already 
contains \[sex1, male1\], the same as the property asserted by the 
adjective. 
If for sentence (4), the dictionary entry of John MeEnroe 
does not previously contain \[sex\], male1\], then the property 
asserted by the adjective is recognised as new information. 
Metonymy is a figure of speech in which the name of one 
thing is substituted for that of another related to it (Lakoff ~: 
Johnson, 1980), for example artist for artform as in 
(5) "John played Bach." 
(-- the music of Bach) and 
(6) "Ted reads Steinbeck." 
(~ the writings of Steinbeck). Other types of metonymy 
include container for contained as in 
(7) "Mary drank the bottle." 
(~ the liquid in the bottle), and co-agent for activity in 
(8) "Lcndl played Wilunder." 
(~ tennis with Wilunder), and part for whole (also known as 
synechdoche). 
3. Knowledge Representation and Knowledge Struc- 
tures 
The knowledge representatiou (KR.) of Collative Semantics 
is a hierarchically structured semantic network with word- 
senses as the nodes. Every node is also a sense-frame, the KS of 
Collative Semantics, so-called because sense-frames are frame- 
like structures representing individual word-senses. This 
arrangement of a semantic network with frame-like structures 
as nodes is common to many frame- and scmantic network- 
based systems (e.g. Bobrow ~ Winograd, 1977; Roberts 2z 
Goldstein, 1977; Brachman, 1979; IIirst, 1983). 
In developing our set of arc labels we have taken note of 
research at the epistemological level of semantic networks 
(Brachman, 1979) so, for example, we distinguish between 
'superinstance' which denotes membership of ~n individual 
within a class of individuals from 'supertypc' which denotes 
membership of a class of individuals by a class of individuals. 
Much like Quillian's (1968) planes, sense-frames are com- 
posed of other word-senses which have their own sense-fl'ames. 
There are no semantic primitives in the sense of Schank's 
(1973) Conceptual Dependency or Wilks' (1975) Preference 
Semantics. In CS word-senses perform the function of semantic 
primitives. They are used to capture broad gcneralisations 
among groups of words or word-senses; to represent the mean- 
ings of individual word-senses; to represent the underlying 
meaning of sentences and their parts; to support lexical disam- 
biguation; to support inferencing of various kinds; and to sup- 
port the discrimination of semantic relations. 
In the next section we explain how semantic primitives arc 
used in Preference Semantics to do lexical disambiguation and 
how word-scnses arc used in CS to do the same. 
4. Techniques for Matching Together Knowledge 
Structures 
Section 4 explains the two techniques used for matching 
together KSs in CS : computing a graph relation and cell- 
342 
matching. 
Computing a graph relation is developed from Wilks' 
(1975) technique for computing satisfied and violated prefer- 
ences in the Preference Semantics System. The 80-100 semantic 
primitives are organised into a directed graph (see Wilks 1977, 
Appendix A for details) in which they appear as nodes. The 
arcs are unlabelled but appear to denote set inclusion, e.g. 
MAN, the class of human beings, belongs to *ANI, the class of 
animate entities. A path searching algorithm operates over tile 
directed graph. Its source and destination nodes are two seman- 
tic primitives. Satisfied preferences are paths describing set 
inclusion; violated preferences are paths describing set exclu- 
sion. 
Computing a graph relation is a path search algorithm 
which operates over CS's hierarchically structured semantic 
network. Its source and destination nodes are two word-senses 
in that network. Five kinds of path (or graph relation) are 
sought between those nodes. Two of the paths describe types of 
set inclusion and hence are the equivalent of a path for a 
satisfied preference. The remaining three kinds of path are 
equivalent to a violated preference and describe types of set 
exclusion. 
Computing satisfied and violated preferences is the basic 
mechanism for doing lexical disambiguation in Preference 
Semantics. Collativc Semantics has at least ~he disambiguation 
power of Preference Semantics because computing a graph rela- 
tion produces paths describing set inclusion and exclusion over 
directed graphs with practically the same hierarchical organisa- 
tion, but whereas the nodes in Wilks' digraph are a restricted 
set of semantic primitives, the nodes in CS's digraph are an 
unrestricted set of word-senses. 
The result of computing a graph relation is recorded in a 
five-column graph relation array where each column 
corresponds to one of the five kinds of set inclusion or exclu- 
sion. An initialised array contains all zeroes. If computing a 
graph relation is used and a particular path is found then a 1 is 
added to the appropriate column of the array. This is a very 
different way of recording the result of a path search from 
Preference Semantics (the importance of recording results is 
considered in the next section). 
The second matching technique in CS, cell matching, is u 
type of multiple comparison algorithm for matching together 
KSs. A multiple comparison algorit\]am is a technique used to 
match together the elements from two sets (often within two 
KSs) by isolating a pair of elements from the two sets and 
matching those elements together in some way. 
Theories of metaphor in linguistics and psychology which 
compare two terms in a metaphorical relation presuppose some 
kind of multiple comparison between their elements. Those ele- 
ments have been termed "features," "properties" and "attri- 
butes" (amongst others) and usually have been contained 
within KSs, normally some type of frame. For example in 
linguistic semantics, Levin (1977) attempts to discriminate 
metaphors from other semantic relations by matching together 
sets of semantic features and in psychology, Ortony (1979) 
discusses metaphor as matches between attributes within two 
schemata. Different kinds of match between elements are con- 
sidered important. Ortony differentiates "identical" matches of 
attributes from "similar" matches based on some "structural 
isomorphism" between the knowledge associated with the two 
attributes. Similar matches are seen as especially important to 
metaphor. 
Cell matching takes as input two lists of elements called 
cells. A pair of cells are isolated from the two lists using a set 
of structural constraints and matched together. As cells are 
composed of word-senses, CS's first matching technique, corn- 
putinga graph relation, can be used to match cells together. 
Five different kinds of match are tried between cells, 
corresponding to the two types of set inclusion and three types 
of set exclusion. These include, to use Ortony's terms, identical 
and similar matches. 
The results of cell matching are recorded in a seven 
column cell-match array, where each column correspond to a 
different kind of cell relation. The first five columns 
correspond to the five types of set inclusion and exclusion. 
Some cells from the two lists fail the structural constraints and 
cannot be matched at all. The last two columns record such 
cases. An initialised cell-match array contains all zeroes. When 
matching two lists of cells together, each occurrence of a cell 
relation between pail's of cells adds 1 to the appropriate column 
of the array until all cells have been accountad for. A cell- 
match array, then, records the complete match of two lists of 
cells. 
5. Recording the Results of Matching Together 
Knowledge Structures 
Section 5 explains how the results of matching together 
sense-frames using computing a graph relation and cell match- 
ing are recorded in semantic vcctors. Semantic vectors arc 
informative enough to tell apart semantic relations between 
word-senses and 1)e the basis for lcxical disambiguation. 
In other approaches to lcxical disambiguation their record 
of the results of matching together KSz cannot be uscd to ade- 
quately discriminate semantic relations. In Preference Seman- 
tics, the record of the result of the path search algorithm is 
binary : either a satisfied or violated preference. A binary 
choice is inadequate because a single satisfied preference could 
be a literal, redundant, contradictory or contrary semantic rela- 
tion and a single violated prcference could be a metaphorical or 
severely anomalous one (see Fuss ~ Wilks, 1983). The same cri- 
ticism can bc made of marker passing (e.g. Quillian, 1968; 
Charniak, 1981; Iiirst, 1983), in which lexical disambiguation is 
done by an intersection search algorithm which finds the shor- 
test path between two nodes in a directed graph. The record is 
a single path, but a single path cannot be used for discriminat- 
ing among semantic relations. And in numerical weighting 
schemes (e.g. Waltz & Pollack, 1985) which "takes place on ,~ 
weighted network of associations, where 'activation energy' is 
distributed over tile network based on some mathematical 
function of the strength of connections" (Ibid, p.54), the 
numerical weights on nodes and edges cannot be used to distin- 
guish semantic relations. 
Semantic vectors are informative enough to be the basis 
for lexical disambiguation and tell apart semantic relations 
between word-senses because they contain graph relation arrays 
and cell match arrays. 
Graph relation arrays are used heavily for lexical disambi- 
guation because a graph relation array records two types of set 
inclusion, equivalent to a satisfied preference in Preference 
Semantics, and three types of set exclusion, equivalent to a 
violated preference. It can bc used for lexieal disambiguation 
just as Preference Semantics uses satisfied and violated prefer- 
ences. 
Cell match arrays record the multiple comparison of ele- 
ments in two KSs in terms of seven different kinds of match. A 
semantic vector usually contains two or more of these arrays. 
Cell match arrays can be used for differentiating semantic rela- 
tions because each relation produces a distinctive pattern of 
nnmbers in the columns of the arrays. See (Fuss, 1986) for 
examples. 
6. Summary 
The knowledge representation and knowledge structure 
Collative Semantics is similar to that of many existing frame- 
or semantic-network based systems. What makes CS different 
and able to address the issues it does (le×ical ambiguity, meto- 
nymy, semantic relations, introduction of new information) is 
the techniques it uses for matching together knowledge struc- 
tures and the recording of those matches into semantic vectors. 

References

Bobrow, D., & Winograd, T. (1977) "An Overview of KRL, A 
Knowledge Representation Language." Cognitive Sci~ 
enee, 1, pp. 3-46. 

Braehman, R.J. (1979) "On The Epistcmological Status of 
Semantic Networks." In N.V. Findler (Ed.) Associative 
Networks : Representation and Use of Knowledge 
By Computers, New York : Academic Press, pp. 3-50. 

Charniak, E. (1981) "Passing Markers : A Theory of Contex- 
tual Influence in Language Comprehension." Technical 
Report CS.-80, Dept. of Computer Science, Brown Univ., 
Providence, Rhode Island. 

Fass, D.C. (1986) "Collative Semantics : A Description of the 
Meta5 Program." Memorandum MCCS-86-23, Computing 
Research Lab., New Mexico State Univ., New Mexico. 

Fass, D.C. & Wilks, Y.A. (1983) "Preference Semantics, Ill-Formedness and Metaphor." American Journal of Computational Linguistics, 9, pp. 178-187. 

Hirst, Graeme (1983) "Semantic Interpretation against Ambi- 
guity." Technical Report CS-83-25, Dept. of Computer 
Science, Brown University. 

Katz, J.,l. (1964) "Analyticity and Contradiction in Natur.~l 
Language." In J.A. I"odor ~ J.J. Katz (Eds.) The Struc- 
ture of Language : P~eadings in the Philosophy of 
Language, Englewood Cliffs, N.J. : Prentice IIall. 

Lakoff, (\]., & Johnson, M. (1980) Metaphors We Live By, 
London : Chicago University Press. 

Levis, S.R. (1977) The Semantics of Metaphor, Baltimore : 
John Ilopkins University Press. 

Ortony, A. (1979) "Beyond Literal Similarity." Psychological 
Review, 86, pp. 161-180. 

Perriue, L. (1971) "Psychological Forms of Metaphor." College 
English, 33, pp. 125-138. 

Quillian, M.R. (1968) "Semantic Memory." In M. Minsky (Ed.) 
Semantic Information Processing, Cambridge, Mass : 
MIT Press, pp. 2\]6-270. 

Richards, I.A. (1936) The Philosophy of Rhetoric, London : 
Oxford University Press. 

Roberts, B.R., & Goldstein, I.P. (1977) "The FRL Manual," 
MIT AI Memo 409. 

Schank, R.C. (1973) "Identification of Conceptualizations 
Underlying Natural Language." In R.C. Schank & K.M. 
Colby (Eds.) Computer Models of Thought and 
Language, San Franscisco : W.Il. Freeman, pp. 187-247. 

Waltz, D.L., & Pollack, J.B. (1985) "Massively Parallel Parsing 
: A Strongly Interactive Model of Natural language 
Interpretation." Cognitive Science, 9, pp. 51-74. 

Wilks, Y.A. (1975) "A Preferential Pattern-Seeking Semantics 
for Natural Language Inference." Artificial Intelli- 
gence, 6, pp. 53-74. 

Wilks, Y.A. (1977) "Good and Bad Arguments about Semantic 
Primitives." Communication and Cognition, 10, pp. 181-221. 
