Using textual clues to improve metaphor processing 
St6phane Ferrari 
LIMSI-CNRS 
PO Box 133 
F-91403 Orsay cSdex, FRANCE 
ferrari@limsi.fr 
Abstract 
In this paper, we propose a textual clue ap- 
proach to help metaphor detection, in order 
to improve the semantic processing of this 
figure. The previous works in the domain 
studied the semantic regularities only, over- 
looking an obvious set of regularities. A 
corpus-based analysis shows the existence 
of surface regularities related to metaphors. 
These clues can be characterized by syn- 
tactic structures and lexical markers. We 
present an object oriented model for repre- 
senting the textual clues that were found. 
This representation is designed to help the 
choice of a semantic processing, in terms of 
possible non-literal meanings. A prototype 
implementing this model is currently un- 
der development, within an incremental ap- 
proach allowing step-by-step evaluations. 1 
1 Introduction 
Metaphor is a frequently used figure of speech, re- 
flecting common cognitive processes. Most of the 
previous works in Natural Language Understanding 
(NLU) looked for regularities only on the semantic 
side of this figure, as shown in a brief overview in 
section 2. This resulted in complex semantic pro- 
cessings, not based on any previous robust detec- 
tion, or requiring large and exhaustive knowledge 
bases. Our aim is to provide NLU systems with a set 
of heuristics for choosing the most adequate seman- 
tic processing, as well as to give some probabilistic 
clues for disambiguating the possibly multiple mean- 
ing representations. 
A corpus-based analysis we made showed the exis- 
tence of textual clues in relation with the metaphors. 
These clues, mostly lexical markers combined with 
syntactic structures, are easy to spot, and can pro- 
vide a first set of detection heuristics. We propose, in 
1This work takes part in a research project sponsored 
by the AUPELF-UREF (Francophone Agency For Edu- 
cation and Research) 
section 3, an object oriented model for representing 
these clues and their properties, in order to integrate 
them in a NLU system. For each class, attributes 
give information for spoting the clues, and, when 
possible, the source and the target of the metaphor, 
using the results of a syntactic parsing. A prototype, 
STK, partially implementing the model, is currently 
under development, within an incremental approach. 
It is Mready used to evaluate the clues relevance. 
In conclusion, we will discuss how the model can 
help chosing the adequate semantic analysis to pro- 
cess at the sentence level or disambiguating multiple 
meaning representations, providing probabilities for 
non-literal meanings. 
2 Classical methods: a brief 
overview 
The classical NLU points of view of metaphor have 
pointed out the multiple kinds of relations between 
what is called the source and the target of the 
metaphor, but rarely discuss the problem of detect- 
ing the figure that bears the metaphor. For our pur- 
pose, we choose to present these approaches in two 
main groups, depending on how they initiate the se- 
mantic processing. 
The previous works led to a classification intro- 
duced by Dan Fass (Fass, 1991). In the compari- 
son view, the metaphor corresponds to an analogy 
between the structures representing the source and 
the target of the figure, as in Gentner's works (Gen- 
tner, 1988) and their implementation (Falkenhainer 
et al., 1989). The interaction view, as in Hobbs 
(Hobbs, 1991), points at the novelty brought by the 
metaphor. Fass also distinguishes a selection restric- 
tions violations view presenting the metaphor as a 
kind of anomaly. We would argue that the two pre- 
vious views already considered metaphor as a kind 
of anomaly. Indeed, the semantic anMysis proposed 
for dealing with metaphors were processed depend- 
ing on the results of another, say a "classical" one 2. 
2We prefer to call it a classical rather than literal 
meanings processing because it can deal with some con- 
ventional metaphors, even if not explicitly mentioned. 
351 
Thereby, detecting a metaphor meant detecting an 
anomaly in the meaning representation issued from 
such a classical analysis. 
Fass proposed a method for discriminating literal 
meanings, metaphors, metonymies and "anomalies", 
merging different points of view (Fass, 1991). In 
this approach, multiple semantic analysis can be pro- 
cessed, resulting in possibly multiple meaning repre- 
sentations. In (Prince and Sabah, 1992), a method 
to overcome similar kinds of ambiguities reveal the 
difficulties encountered if no previous detection is 
made. James Martin's approach (Martin, 1992), 
called the conventional view by Fass, is based on 
Lakoff's theory on cognitive metaphors (Lakoff and 
Johnson, 1980). It requires a specific knowledge rep- 
resentation base and also results in multiple repre- 
sentation meanings. Detecting a metaphor is mean- 
ingless here, and conventional metaphoric meanings 
can be viewed as polysemies. Martin revealed at 
least that the heuristic of the ill-formness of mean- 
ing representations issued from classical analysis is 
not sufficient at all to deal with all the possible 
metaphors. 
In our point of view, all the previous approaches 
were founded. The main remaining problem, how- 
ever, is to choose an adequate processing when con- 
fronted with a metaphor, and thus, to detect the 
metaphors before trying to build their meaning rep- 
resentation. This can be partially solved using tex- 
tual clues. 
3 Textual clues: object oriented 
description 
If the classical views of the metaphor overlook the 
textual clues, in other domains, especially those 
concerning explanation, they have been wisely re- 
introduced. In (Pery-Woodley, 1990), Pery-Woodley 
shows the existence of such clues related to the 
explanatory discourse. They can help in generat- 
ing explanations in natural language as well as in 
modelling the student in a intelligent tutoring sys- 
tem (Daniel et al., 1992). A corpus of 26 explana- 
tory texts in French, of about 200 words each, has 
been collected under a shared research project be- 
tween psychologists and computer scientists, in or- 
der to study metaphors and analogies in teaching. 
The analysis we made showed the existence of tex- 
tual clues in relation with metaphoric contexts and 
analogies (e.g. "like", "such as", "illustrated by"). 
They can be characterized by syntactic regularities 
(e.g. the comparative is used in structures such as 
"less than", "more than"; the identification is made 
through attributes or appositions, ...). They also 
involve lexical markers (e.g. "literMy", "illustrat- 
ing", "metaphorically" ,). These properties, already 
found in the previous works, can help detecting the 
clues themselves. Studying the relation between the 
syntactic regularities and the lexical markers, one 
can observe that the first build the ground where to 
find the second. We thus propose an object-oriented 
model for representing these clues. A generic textual 
clue can thereby be described by the two following 
attributes: 
• the Surface Syntactic Pattern representing the 
syntactic regularity, with a label on the item 
where to find the lexical marker 
• the Lexical Marker itself 
Typically, the word "metaphor" itself can be used 
as a lexical marker in expressions such as '~to ex- 
tend the conventional metaphor, pruning such a 
tree means to generalize". On the other hand, 
"metaphor" will not be a marker if used as the 
subject of the sentence, like in this one. Thus, 
describing the syntactic regularities surrounding a 
lexical marker improves its relevance as a marker. 
We propose to represent this relevance for proba- 
bilistic purposes. Each clue that was found is cur- 
rently evaluated on a large corpus (about 450,000 
words). The frequencies of use of the lexical mark- 
ers in metaphoric contexts are represented in the 
relevance attribute (see example below). 
The syntactic structures may also give infor- 
mation about the source and the target of the 
metaphor. For instance, in the sentence "Yesterday, 
at home, Peter threw himself on the dessert like a 
lion.", the subject inherits the properties of speed 
and voracity of a lion attacking its victim. It is here 
possible to spot the source and the target of the 
metaphor using the syntactic properties of the com- 
parison. Two attributes are added to textual clues 
related to metaphors, corresponding to the elements 
of the sentence bearing the source and the target. 
Example of textual clue representations 
type metaphor-analogy 
name B.2.2.2 
comment comparison involving the meaning of a 
marker, adjective, attribute of the object, object 
before the verb 
SSP GNo GN1 Vx Adjo \[prep\] GN2 
LM Adjo: pareil (meaning "similar") 
target GN1 
source GN2 
LM relevance (15/28) 
number of occurrences 28 
conventional metaphors 3 
new metaphors 2 
metaphomc contexts 12 
total 15 
Notations: GN and GV stand for nominal or verbal 
groups, Adj and Adv for adjectives and adverbs, and 
prep for prepositions. 
The model has been partially implemented in a 
tool, STK, for detecting the textual clues related to 
352 
metaphors and adding specific marks when found. 
In its current version, STK allows us to tokenize, 
tag, and search for lexical markers on large corpora. 
The tagger we use is the one developped by Eric 
Brill (Brill, 1992) with a set of tags indicating the 
grammatical categories as well as other information 
such as the number and the gender for nouns and 
adjectives. It is evaluated under GRACE 3 protocol 
for corpus-oriented tools assigning grammatical cat- 
egories. It is currently used for the evaluation of 
the textual clues that were found. The latter can 
be easily retrieved using STK, avoiding lexical am- 
biguities. They are then analyzed by hand, in order 
to determine their relevance attribute. In the previ- 
ous example of textual clue, the relevance values are 
issued from this corpus-based analysis. 
4 Conclusion, perspectives 
Classical approaches to the metaphor in NLU re- 
vealed multiple underlying processes. We there- 
fore focussed our study on how to help detecting 
metaphors in order to chose the most adequate se- 
mantic processing. Textual clues can give informa- 
tion about the figures that bear the metaphor, which 
are easy to spot. Indeed, they can be found using 
the results of syntactic parsing. We proposed an 
object-oriented model to represent these clues and 
their multiple properties. 
If textual clues give information about possible 
non-literal meanings, metaphors and analogies, one 
may argue they do not allow for a robust detection. 
Indeed, a textual clue is not sufficient to prove the 
presence of such figures of speech. The relevance of 
each clue can be used to help disambiguating mul- 
tiple meaning representation when it occurs. This 
must not be the only disambiguation tool, but when 
no other is avalaible, it provides NLU systems with 
a probabilistic method. 
Our future works will focuss on the study of the 
relation between the metaphors introduced by a clue 
and others that are not conventional. The guideline 
is that novel metaphors not introduced by a clue at 
the sentence level may have been introduced previ- 
ously in the text. 

References 
Brill, E. (1992). A simple rule-based part of speech 
tagger. In Proceedings of the Third Conference 
on Applied Natural Language Processing, Trento. 
ACL. 
Daniel, M., Nicaud, L., Prince, V., and Pery- 
Woodley, M. (1992). Apport du style Linguistique la Mod~lisation Cognitive de l'\]~l~ve. Lec- 
ture Notes in Computer Sciences, 608:252-260. 
Proceedings of the International Conference on In- 
telligent Tutoring Systems (ITS-92), MontrEal. 
Falkenhainer, B., Forbus, K., and Gentner, D. 
(1989). The Structure-Mapping Engine: Algo- 
rithm and Examples. Artificial Intelligence, 41:1- 
63. 
Fass, D. (1991). met : A Method for Discriminating 
Metonymy and Metaphor by Computer. Compu- 
tational Linguistics, 17(1):49-90. 
Fass, D., Hinkelman, E., and Martin, J., editors. 
Proceedings of the IJCAI Workshop on Computa- 
tional Approaches to Non-Literal Language, Syd- 
ney, Australia. 1991. 
Gentner, D. (1988). Analogical Inference and Ana- 
logical Access, In: Analogica, chapter 3, pages 
63-88. Edited by Prieditis A., Pitman Publish- 
ing, London, Morgan Kaufmann Publishers, Inc., 
Los Altos, California. 
Hobbs, J. (1991). Metaphor and abduction. In (Fass 
et al., ), pages 52-61. 
Lakoff, G. and Johnson, M. (1980). Metaphors we 
live by. University of Chicago Press, Chicago, 
U.S.A. 
Martin, J. (1992). Computer Understanding of Con- 
ventional Metaphoric Language. Cognitive Sci- 
ence, 16:233-270. 
Pery-Woodley, M. (1990). Textual clues for user 
modeling in an intelligent tutoring system. Mas- 
ter's thesis, University of Manchester, England, 
Great-Britain. 
Prince, V. and Sabah, G. (1992). Coping with 
Vague and Fuzzy Words : A Multi-Expert Natural 
Language System which Overcomes Ambiguities. 
In Acts of PRICAI'92, Seoul, Corea. September, 
1992. 
