Non-Literal Word Sense Identification Through Semantic Network 
Path Schemata 
Eric lverson, Stephen Helmreich 
Computing Research Lab and Computer Science I~panment 
Box 30001/3CRL 
New Mexico State Unive~ty 
Las Cruc~, NM 88003-0001 
When computer programs disambiguate words 
in a sentence, they often encounter non-literal or 
novel usages not included in their lexicon. In a 
recent study, Georgia Green (personal communica- 
tion) estimated that 17% to 20% of the content word 
senses encountered in various types of normal 
English text are not fisted in the dictionary. While 
these novel word senses are generally valid, they 
occur in such great numbers, and with such little 
individual frequency that it is impractical to expli- 
city include them all within the lexicon. Instead, 
mechanisms are needed which can derive novel 
senses from existing ones; thus allowing a program 
to recognize a significant set of potential word senses 
while keeping its lexicon within a reasonable size. 
Spreading activation is a mechanism that 
allows us to do this. Here the program follows paths 
from existing word senses stored in a semantic net- 
work to other closely associated word senses. By 
examining the shape of the resultant path, we can 
determine the relationship between the senses con- 
~ned in the path; thus deriving novel composite 
meanings not contained within any of the original 
lexical entries. This process is similar to the spread- 
ing activation and marker passing techniques of Hirst 
\[1988\], Charniak \[1986\], and Norvig \[1989\] and is 
embodied in the Prolog program metallel based on 
Fass' program meta5 (Fass \[1988\]). 
Metallel's lexicon is written as a series of 
sense frames, each containing information about a 
particular word sense. A sense frame can he broken 
into two main parts: genera and differentiae. Gen- 
era are the genus terms that function as the ancestors 
of a word sense. Differentiae denote the qualities 
that distinguish a particular sense from other senses 
of the same genus. Differentiae can be broken down 
into source and target which hold, respectively, the 
preferences t and properties of a sense. Source con- 
=dns differentiae mform~on concen~g another 
word sense. Target infocma~on concerns the sense 
itself. 
Connections can be found to other word senses 
in one of two ways: through an ancestor relationship 
(genus) er through a preference or property relation- 
ship (differentia). In the case of differentiae, it is 
necessary to extract the word senses from a higher 
order structure. For example, \[it (n, z), 
contain (v, l), n~asic (n, Z) \] is not a word sens¢ 
~at is LL~ted in the lexicon, while ~asic (n, i) is 
Us~L It is therefore necessary to ex~act 
rausic (n,Z) from the larger dfffereada s~ucmre 
which it occurs and add it to the path. 
Not all paths are valid, indicating that some 
criteria of acceptability are needed during analysis. 
In addition, paths that are superficially different often 
end up being quite similar upon further analysis. 
Keeping this in mind, we have attempted to identify 
path schemata and associate them wkh types of non- 
literal usage. Specifically, we have concentrated on 
identifying instances of metaphor and metonymy. 
A metaphorical path schema is one in which 
the preference of a verb and the actual target of the 
preference both reference different 3 place differen- 
tiae 2 which can be said to be related. Two 3 place 
z Pn:f=mce* indicate the zema~dc category c~ the word 
=ca== dug fill= • specific u~umfic teL= with ~ w the 
word =ca== being de£u~L For ¢xamp~. d~ mm~v¢ ~mse 
of d~ verb e~ pmfen Cm normal u~ge) == =n~m=¢ ~bje~ 
and -- e~b~= objoc~ Vk~uiom of ~=~ pmfcnmc= =m m- 
dicmiom ~ aou-\[kcnd mmg~ (See Wflk= and Fus \[1990\].) 
z A 3 ,,~=_~_- diff=~m6= ~ a li= of tomes following a 
\[Subject, Verb, Object\] foemat in which ei~h= the Subject or 
the Objc~o0asbt= ofd~~mkm it (n, 1). 
343 
differentiae are related if both their respective rob- 
jeers and objects are identical or form a "sister" rela- 
tionship 3. Additictmlly, the two verbs of the dif- 
ferentiae as well as the verb which generated the 
preference must have a similar relationship 
The ship ploughed the waves. 
ship (n, 1) -anc-> 
watercraft (n, 1) -prop-> 
\[it (n, i), sail (v, 2), water (n, 2) \] -link-> 
water (n, 2) -anc-> 
environment (n, I) <-anc- 
soil (n, I) <-link- 
\[it (n, 1), plough (v, 2), soil (n, 1) \] <-prop- 
plough (n, 1) <-inst- 
plough (v, i) -ohj-> 
soil(n, 1) -ant-> 
environment (n, I) <-ant- 
water (n, 2) <-part- 
wave (n, I) 
For example in the path for the senw.nce The 
ship ploughed the waves, \[it (n, 1), sail (v, 2), 
water (n, i) \] and \[it (n, 1), plough (v, 2), 
soil (n, 1) \] are related ~ plough (v, 1), 
plough(v, 2) and sail(v, 2) a~ ch~dlP~ of 
transfer (v, i), and water (n, I) and 
soil (n, I) ai~ ch~dlP~ of environment (n, I). 
A/so, the pivot nodes 4 for the insmuneat and object 
p~ferences of plough (v, i) ~ b~h 
environment (n, l) , thereby indicating an even 
monger relationship between the insmmaent and the 
object of the senwnce. Thus, an analogy exists 
between ploughing soil and sailing water;, suggesting 
a new sense of plough that combines aspects of beth. 
Denise drank the bottle. 
denise (n, 1) -anc-> 
woman (n, 1} -prop-> 
\[sex (n, i), \[female (aj# I) \] \] -link-> 
female (aJ, i) -obj-> 
animal (n, I) <-agent- 
drink (v, i) -obj-> 
drink (n # 1 } -ant-> 
liquid(n, 1) <-link~ 
lit (n, 1 ), contain (v, I), liquid (n, I) \] <-prop- 
bottle (n, 1} 
A metonymic path is indicated when a path is 
found from a target sense through one of its inherited 
differentiae; thus linking the original sense to a 
related sense through a property or preference rela. 
tionship. For example in the sen~nce Denise drank 
the bottle, one of the properties of bottle (n, 1) is 
\[it (n, 1), contain (v, 1), liquid (n, 1) 1. 
This differealia allows us to derive a novel meto- 
nymic word sense for bottle in which the bottle's 
conwmts are denoted rather than the boule itself. 
Under memUel, any differentia can act as a conduit 
for a memnymy; thus facilitating the generation of 
novel metonymies as well as novel word senses. 
By using semantic network path schemata to 
identify instances of non-literal usage, we have 
expanded the power of our program without doing so 
at the expense of a larger lexicon. In addition, by 
keeping our semantic relationship and path schema 
criteria at a general level, we hope to be able to 
cover a wide variety of different semantic taxo- 
nomies. 

References 
Clmmi~, E 1986. A neat theory of marker pass- 
ing. Procs. AAAI-86. Philadelphia, PA. 
Fass, D. 1988. Collafive Semantics: A Semantics 
for Natural Language Processing. Memoranda 
in Computer and Cognitive Science, MCCS- 
88-118. Computing Research Laboratory, New 
Mexico State University. 
Hirst, G. 1988. Resolving lexical ambiguity compu- 
rationally with spreading activation and 
polaroid words. In Small and Cottrell (eds.), 
Lexical Ambiguity Resolution pp. 73-107. Mor- 
gan Ica-fmann: San Ma~o. 
Norvig, P. 1989. Marker passing as a weak method 
for text inferencing. Cognitive Science 
13(4)'..569-620. 
Wilks, Y., and D. Fass. 1990. Preference Semantics. 
Memoranda in Computer and Cognitive Sci- 
ence, MCCS-90-194. Computing Research 
l~_borato~, New Mexico State University. 
