Proceedings of the Twenty-Seventh AAAI Conference on Artificial Intelligence

Automatic Identification of Conceptual Metaphors with Limited Knowledge

Lisa Gandy1 Nadji Allan2 Mark Atallah2 Ophir Frieder3 Newton Howard4
Sergey Kanareykin5 Moshe Koppel6 Mark Last7 Yair Neuman7 Shlomo Argamon1∗
1
Linguistic Cognition Laboratory, Illinois Institute of Technology, Chicago, IL
2
Center for Advanced Defense Studies, Washington, DC
3
Department of Computer Science, Georgetown University, Washington, DC
4
Massachusetts Institute of Technology, Cambridge, MA
5
Brain Sciences Foundation, Providence, RI
6
Department of Computer Science, Bar-Ilan University, Ramat Gan, Israel
7
Departments of Education and Computer Science, Ben-Gurion University, Beersheva, Israel

Abstract
Full natural language understanding requires identifying and
analyzing the meanings of metaphors, which are ubiquitous
in both text and speech. Over the last thirty years, linguistic
metaphors have been shown to be based on more general con-
ceptual metaphors, partial semantic mappings between dis-
parate conceptual domains. Though some achievements have
been made in identifying linguistic metaphors over the last
decade or so, little work has been done to date on automati-
cally identifying conceptual metaphors. This paper describes
research on identifying conceptual metaphors based on cor-
pus data. Our method uses as little background knowledge Figure 1: Three levels of metaphor processing.
as possible, to ease transfer to new languages and to mini-
mize any bias introduced by the knowledge base construction
process. The method relies on general heuristics for identi-
fying linguistic metaphors and statistical clustering (guided In this paper, we describe a novel system for discovering
by Wordnet) to form conceptual metaphor candidates. Hu- conceptual metaphors using natural language corpus data.
man experiments show the system effectively finds meaning- The system identifies linguistic metaphors automatically and
ful conceptual metaphors. uses them to find meaningful conceptual metaphors. A key
design principle is to rely as little as possible on domain and
Introduction language specific knowledge bases or ontologies, but rather
Metaphor is ubiquitous in human language. Thus, identi- more on analyzing statistical patterns of language use. We
fying and understanding metaphorical language is a key limit such use to generic lexical semantics resources such
factor in understanding natural language text. Metaphor- as Wordnet. Such a knowledge-limited, corpus-centric ap-
ical expressions in language, called linguistic metaphors, proach promises to be more cost effective in transferring
have been shown to be derived from and justified by un- to new domains and languages, and to potentially perform
derlying conceptual metaphors (Lakoff and Johnson 1980), better when the underlying conceptual space is not easily
which are cognitive structures that map aspects of one con- known in advance, as, e.g., in the study of metaphor varia-
ceptual domain (the source concept) to another, more ab- tion over time or in different cultures.
stract, domain (the target concept). A large and varied sci- The system works at three interrelated levels of metaphor
entific literature has grown over the last few decades study- analysis (Figure 1): (i) linguistic metaphors (individual
ing linguistic and conceptual metaphors, in fields includ- metaphoric expressions), (ii) nominal analogies (analogi-
ing linguistics (e.g., (Johnson, Lakoff, and others 2002; cal mappings between specific nouns), and (iii) conceptual
Boers 2003)), psychology (e.g., (Gibbs Jr 1992; Allbritton, metaphors (mappings between concepts). For instance, the
McKoon, and Gerrig 1995; Wickman et al. 1999)), cogni- system, when given a text, might recognize the expression
tive science (e.g., (Rohrer 2007; Thibodeau and Borodit- ‘open government’ as a linguistic metaphor. The system may
sky 2011)), literary studies (e.g., (Goguen and Harrell 2010; then find the nominal analogy “government ∼ door” based
Freeman 2003)), and more. on that linguistic metaphor and others. Finally, by clustering
∗
Corresponding author. Email: [email protected] this and other nominal analogies found in the data, the sys-
Copyright  c 2013, Association for the Advancement of Artificial tem may discover an underlying conceptual metaphor like
Intelligence (www.aaai.org). All rights reserved. “AN ORGANIZATION IS A CONNECTION”.

328
 Related Work
There has been a fair amount of computational work on iden-
tifying linguistic metaphors, but considerably less on finding
conceptual metaphors. Most such work relies strongly on
predefined semantic and domain knowledge. We give here
a short summary of the most relevant previous work.
Birke and Sarkar (2006) approach the problem as a clas-
sical word sense disambiguation (WSD) task. They reduce
nonliteral language recognition to WSD by considering lit-
eral and nonliteral usages to be different senses of a single
word, and adapting an existing similarity-based word-sense
disambiguation method to the task of separating verbs oc-
currences into literal and nonliteral usages.
Figure 2: System Architecture.
Turney et al. (2011) identify metaphorical phrases by as-
suming that these phrases will consist of both a ”concrete”
term and an ”abstract” term. In their work they derive an al-
gorithm to define the abstractness of a term, and then use Note that the only manually-created background knowl-
this algorithm to contrast the abstractness of adjective-noun edge sources used are Wordnet and the number of defini-
phrases. The phrases where the abstractness of the adjective tions of words in Wiktionary; these are used in very circum-
differs from the abstractness of the noun by a predetermined scribed ways in the system, as described below. Other than
threshold are judged as metaphorical. This method achieved those resources, the only language-dependence in the system
better results than Birke and Sarkar on the same data. is the use of a parser to find expressions of certain forms as
For interpreting metaphors researchers have used both metaphor candidates, such as verb-object pairs.
rule and corpus based approaches. In one of the most
successful recent corpus-based approaches, Shutova et Linguistic Metaphors
al. (2012) find and paraphrase metaphors by co-clustering The system uses a heuristic algorithm to identify linguistic
nouns and verbs. The CorMet system (Mason 2004) finds metaphors. The method classifies expressions with specific
metaphorical mappings, given specific datasets in disparate syntactic forms as being either metaphorical or not based
conceptual domains. That is, if given a domain focused on on consideration of the two focal words in the expression.
lab work and one focused on finance it may find a conceptual The types of expressions we consider in this work are those
metaphor saying that LIQUID is related to MONEY. delineated by (Krishnakumaran and Zhu 2007):
Nayak & Mukerjee (2012) have created a system which
learns Container, Object and Subject Metaphors from ob- Type 1 The subject and complement of an identifying
serving a video demonstrating these metaphors visually clause e.g., “My lawyerT is a sharkF .”
and also observing a written commentary based on these Type 2 A lexical verb together with its direct object, e.g.,
metaphors. The system is a grounded cognitive model which “He enteredF politicsT .”
uses a rule based approach to learn metaphor. The system Type 3 An adjective and the noun it describes, e.g.,
works quite well, however it is apparently only currently ca- “sweetF childT ,” or “The bookT is deadF .”
pable of learning a small number conceptual metaphors.
We term the descriptive, possibly “metaphorical,” term in
The Metaphor Analysis System each such expression the facet (marked F in the examples
above), and the noun being described the target (marked T ).
As discussed above, our metaphor analysis system works at
Note that the facet expresses one aspect of a source domain
three interrelated levels of metaphor analysis: finding lin-
used to understand the target metaphorically.
guistic metaphors, discovering nominal analogies and then
clustering them to find meaningful conceptual metaphors. Our algorithm for identifying linguistic metaphors, an ex-
As shown in Figure 2, each level of metaphor analysis feeds pansion of our earlier work (Assaf et al. 2013), is based on
into the next; all intermediate results are cached to avoid the key insight in Turney et al.’s (2011) Concrete-Abstract
recomputing already-computed results. All levels of analy- algorithm, i.e., the assumption that a metaphor usually in-
sis rely on a large, preparsed corpus for statistical analysis volves a mapping from a relatively concrete domain to a
(the experiments described below use the COCA n-grams relatively abstract domain. However, we also take into ac-
database1 ). The system also currently uses a small amount count the specific conceptual domains involved. Literal use
of predefined semantic knowledge, including lexical seman- of a concrete facet will tend to be more salient for certain
tics via Wordnet, dictionary senses via Wiktionary, and Tur- categories of concrete objects and not others. For example,
ney’s ratings of the abstractness of various terms (this last in its literal use, the adjective “dark” may be associated with
is derived automatically from corpus statistics). Input texts certain semantic categories such as Substance (e.g. “wood”)
are also parsed to extract simple expressions as linguistic or Body Part (e.g. “skin”).
metaphor candidates to be classified by the system. To illustrate the idea, consider the case of Type 3
metaphors, consisting of an adjective-noun pair. The algo-
1
http://corpus.byu.edu/coca/ rithm assumes that if the modified noun belongs to one of the

329
concrete categories associated with the literal use of the ad-
jective then the phrase is probably non-metaphorical. Con-
versely, if the noun does not belong to one of the concrete
categories associated with the literal use of the adjective then
it is probably metaphorical. In a sense, this algorithm com-
bines the notion of measuring concreteness and that of using
selectional preferences, as has been well-explored in previ-
ous work on metaphor identification (Fass and Wilks 1983).
To illustrate the approach, consider analyzing an
adjective-noun expression A, N  such as “open govern-
ment”. First, if the adjective A has a single dictionary defi-
nition then the phrase is labeled as non-metaphorical, since
metaphorical usage cannot exist for one sense only. Then,
the algorithm verifies that the noun N belongs to at least
one high-level semantic category (using the WordStat dic-
tionary of semantic categories based on Wordnet2 ). If not, Figure 3: A schematic partial view of the nominal analogy
the algorithm cannot make a decision and stops. Otherwise, “government ∼ door.” Facets in the middle are associated
it identifies the n nouns most frequently collocated with A, with the target noun on the left metaphorically, and with the
and chooses the k most concrete nouns (using the abstract- source noun on the right literally.
ness scale of (Turney et al. 2011)). The high-level semantic
categories represented in this list by at least i nouns each are
selected; if N does not belong to the any of them, the phrase ated (in the appropriate syntactic relationships) with each of
is labeled as metaphorical, otherwise it is labeled as non- those facets with positive PMI.
metaphorical. The same method applies to other expression For example, consider finding candidate source terms for
types, by varying how collocations are computed. the target term “government.” We first find all facets in lin-
Based on analysis of a development dataset, disjoint from guistic metaphors identified by the system that are associ-
our test set, we set n = 1000; k = 100; and i = 16. ated with “government” with a positive PMI. These include
such terms as “better”, “big”, “small”, “divided”, “open”,
Nominal Analogies “closed”, and “limited.” Each of these facets also are asso-
Once a set of metaphorical expressions SM = {f, nt } ciated with other nouns in non-metaphorical senses. For in-
(each a pair of a facet and a target noun) is identified, we stance, the terms “open” and “closed” are associated with
seek nominal analogies which relate two specific nouns, a the words “door”, “table”, “sky”, “arms”, and “house”.
source and a target. For example, if the system finds many Identification. To identify which pairs of candidate source
different linguistic metaphors which can be interpreted as terms in SC and target terms are likely nominal analo-
viewing governments as doors, it may identify a nominal gies, the system computes a measurement of the similarity
analogy “government ∼ door.” of the non-metaphorical facets of each candidate with the
The basic idea behind the nominal analogy finding al- metaphorical facets of each target.
gorithm is that, since the metaphorical facets of a target To begin, define for each pair of a facet fi and noun
noun are to be understood in reference to the source noun, (source or target) ni , its association score aij to be its PMI,
we must find source/target pairs such that many of the and the pair’s metaphoricity score, mij such that mij = 1
metaphorical facets of the target are associated in literal if the pair is judged by the system to be metaphorical and
senses with the source (see Figure 3). The stronger and more mij = 0 if it is judged to be non-metaphorical. Currently,
numerous these associations, the more likely the nominal all metaphoricity scores are 1 or 0, but we plan to generalize
analogy is to be real. the scheme to allow different levels of confidence in linguis-
Note that the system will not only find frequent associa- tic metaphor classification.
tions. Since the association strength of each facet-noun pair- We then define a metaphoric facet distribution (MFT) for
ing is measured by PMI, not its raw frequency, infrequent facets given target terms by normalizing the product of as-
pairings can (and do) pop up as significant. sociation and metaphoricity scores:
Candidate generation. We first define the candidate set aij mij
of source nouns SC as the set of all nouns (in a given cor- PM (fi |nj ) = 
ij aij mij
pus) with strong non-metaphoric associations with the facets
used in the linguistic metaphors in SM . Start with the set of as well as a literal facet distribution (LFT), similarly, as:
all facets used in SM which have positive point-wise mutual
information (PMI; cf. (Pantel and Lin 2002)) with at least aij (1 − mij )
PL (fi |nj ) = 
one target term in the set. SC is then defined as the set of ij aij (1 − mij )
all other (non-target) nouns (in a given large corpus) associ-
As noted above, we seek source term / target term pairs
2
See http://provalisresearch.com/. such that the metaphorical facets of the target term are likely

330
to be literal facets of the source term and vice versa. A nat- of a single Wordnet synset, by which synset we represent the
ural measure of this tendency is the Jensen-Shannon diver- concept3 .
gence between the LFT of ns and the MFT of nt , The rest of the conditions amount to the following hard
and soft constraints:
DJS (PL (·|ns )  PM (·|nt ))
• α(Ct ) > α(Cs );
The larger the J-S divergence, the less likely the pair is to be • The larger α(Ct ) − α(Cs ), the better;
a good nominal analogy. • The more possible facets associated with covered nominal
We then compute a heuristic estimation of the probability analogies, the better;
that a noun pair is a nominal analogy: 
• The higher ns ,nt ∈N P (CP ) PN A (ns , nt ), the better;
2
PN A (ns , nt ) = e−α(ns )DJS (PL (·|ns )PM (·|nt )) • The lower α(Cs ), the better; and
• The more elements in N P (CP ), the better.
where α(ns ) is the abstraction score of ns . Empirically, we
have found it useful to multiply the J-S divergence by the We thus define the overall score for each pair of concepts
abstractness score of the candidate source noun, since good Cs , Ct  as:
source nouns tend to be more concrete (since metaphors typ-

ically map a concrete source domain to an abstract target (α(Ct ) − α(Cs ))φ(CP ) log ns ,nt ∈N P (CP ) PN A (ns , nt )
domain).
α(Cs )2
The output of this process is a set of potential nominal (1)
analogies, SA = {ns , nt }, each a pair of a source noun where
and a target noun, with an associated confidence score given |N A(Cs , Ct )|
by PN A . We note that at this stage recall is more important φ(Cs , Ct ) =
max(|Cs |, |Cs |)2
than precision, since many nominal analogies will be filtered
out when nominal analogies are clustered to find conceptual such that N A(Cs , Ct ) = {ns , nt |PN A (ns , nt ) > 12 } is
metaphors. the set of all likely nominal analogies for the concept pair.
The function φ thus gives the fraction of all possible such
Conceptual Metaphors nominal analogies that are actually found.
The system then retains just the candidate concept pairs
After computing nominal analogies SA , the system seeks a that satisfy the hard constraints above and have at least a
set of plausible conceptual metaphors that parsimoniously given number of shared facets (in our experiments 5 worked
covers the most likely ones nominal analogies. A conceptual well), ranking them by the scoring function in equation
metaphor (CM) can be represented as a pair Cs , Ct  of a (1). It then removes any concept pairs subsumed by higher-
source concept Cs and a target concept Ct . scoring pairs, where Cs , Ct  subsumes Cs , Ct  iff:
Which concept pairs most likely represent conceptual Cs is a hyponym of Cs and
metaphors? We consider three key conditions indicating a (Ct is a hyponym of Ct or Ct = Ct )
likely conceptual metaphor: (i) consistency of the concept
pair with linguistic metaphors (as identified by the system), or
(ii) the semantic relationship between the two concepts, and Ct is a hyponym of Ct and
(iii) properties of each of the concepts on their own. (Cs is a hyponym of Cs or Cs = Cs )
To address consistency with the corpus, consider a con- The result is a ranked list of concept pairs, each with a list
cept as a set of nouns, representing the set of possible in- of facets that connect relevant source terms to target terms.
stances of the given concept. A concept pair CP = Cs , Ct  Each pair of concepts constitutes a conceptual metaphor as
thus may be viewed as the set of noun pairs N P (CP ) = found by the system, and its distribution of facets is a rep-
{ns , nt |ns ∈ Cs , nt ∈ Ct }, each of which is a poten- resentation of the conceptual frame, i.e., the aspects of the
tial nominal analogy. The more such pairs are high-scoring source concept that are mapped to the target concept.
NAs, and the higher their scores, the more likely CP is to
be a conceptual metaphor. Furthermore, consider the facets Evaluation
that join these NAs; if the same facets are shared among
many of the NAs for CP , it is more likely to be a concep- Linguistic Metaphors
tual metaphor. We evaluated linguistic metaphor identification based on a
For the second condition, the semantic relationship of the selection of metaphors hand-annotated in the Reuters RCV1
two concepts, we require that the target concept Ct be more dataset (Lewis et al. 2004).
abstract than the source concept Cs ; the larger the abstract- For feasibility of annotation, we considered all expres-
ness gap, the more likely CP is to be a conceptual metaphor. sions of Types 1, 2, and 3 for a small number of tar-
Finally, the concepts should be semantically coherent, and get terms, “God”, “governance”, “government”, “father”,
the source concept should be comparatively concrete. 3
Our approach here is similar to that of (Ahrens, Chung, and
Currently, to ensure that concepts are semantically coher- Huang 2004), though they rely much more strongly on Wordnet to
ent we consider only sets of nouns which are all hyponyms determine conceptual mappings.

331
 Table 1: Linguistic metaphor detection results for RCV1 Table 3: Conceptual metaphor validation results. See text for
explanation.
Metaphor % metaphors Precision Recall
Type Condition CMs validated Trials validated
Type 1 66.7% 83.9% 97.5% by majority
Type 2 51.4% 76.1% 82% All CMs 0.65 0.60
Type 3 24.2% 54.4% 43.5% High-ranking CMs 0.71 0.64
Foils 0.53 0.60

and “mother”. From each of the 342,000 texts we identi-

fied the first sentence (if any) including one of the target Table 4: Framing results. Implied LMs are facet/target pairs
words. Each such sentence was parsed by the Stanford De- from a single CM while foils are taken from different CMs.
pendency Parser (De Marneffe, MacCartney, and Manning
2006), from which were automatically extracted expressions Condition Expressions Trials validated
with the syntactic forms of Types 1, 2, and 3. validated by
Four research assistants annotated each of these expres- majority
Implied LMs 0.77 0.76
sions to determine whether the key word in the expression is
Foils 0.66 0.69
used in its most salient embodied/concrete sense or in a sec-
ondary, metaphorical sense. For instance, in the case of “bit-
ter lemon” the first embodied definition of the adjective “bit-
ter” is “Having an acrid taste.” Hence in this case, it is used “governance”. These terms were: “governance,” “govern-
literally. When asked to judge, on the other hand, whether ment,” “authority,” “power,” “administration,” “administra-
the phrase “bitter relationship” is literal or metaphorical, tor,” “politics,” “leader,” and “regime.” The highest ranked
that basic meaning of “bitter” is used to make a decision; conceptual metaphors found by the system are in Table 2.
as “relationship” cannot have an acrid taste, the expression For evaluation, we recruited 21 naive subjects, 11 female,
is judged as metaphorical. 10 male, with native fluency in American English. Subject
Annotators were given the entire sentence in which each age ranged from 19 to 50, averaging 25.5 years old. Educa-
expression appears, with the target word and related lexical tion ranged from subjects currently in college to possessing
units visually marked. Inter-annotator agreements, measured master’s degrees; just over half of all subjects (12) had bach-
by Cronbach’s Alpha, were 0.78, 0.80 and 0.82 for expres- elor’s degrees.
sions of Type 1, 2, and 3 respectively. The majority decision Each subject was first given an explanation of what
was used as the label for the expression in the testing corpus. conceptual metaphors are in lay terms. The subject was
Results (precision and recall) are given in Table 1. Both then given a sequence of conceptual metaphors and asked
precision and recall are significantly above baseline for all whether each conceptual metaphor reflects a valid and mean-
metaphor types. We do note that both precision and recall ingful conceptual metaphor in the subject’s own language.
are notably lower for Type 3 metaphors, partly due to their Some of the conceptual metaphors presented were those
much lower frequency in the data. generated by the system, and others were foils, generated
by randomly pairing source and target concepts from the
Conceptual Metaphors system-generated results.
Our evaluation metrics were the fraction of CMs vali-
To evaluate the validity of nominal analogies and concep- dated by a majority of subjects and the fraction of the to-
tual metaphors found by our system, we ran the system to tal number of trials that were validated by subjects. We also
generated a set of linguistic metaphors, nominal analogies, separately evaluated validation for the highest-scoring CMs,
and conceptual metaphors starting with slightly larger but which were the top 19 out of the 32 total CMs found.
more focused set of target terms related to the concept of
Results are given in Table 3. We note first of all that a clear
majority of CMs produced by the system were judged to be
meaningful by most subjects. Furthermore, CMs produced
Table 2: Top conceptual metaphors generated by the system by the system are more likely to be validated by the major-
for governance-related targets. ity of subjects than foils, which supports the hypothesis that
system-produced CMs are meaningful; higher ranking CMs
IDEA is a GARMENT were more validated yet.
POWER is a BODY PART To evaluate the system’s framing of its conceptual
POWER is a CONTAINER metaphors, subjects were presented with a succession of
POWER is a CONTESTANT word pairs and asked to judge the meaningfulness of each
POWER is an ARMY UNIT expression as a metaphor. Each primary stimulus rep-
POWER is a BUILDING resented a linguistic metaphor implied by a conceptual
POWER is a CHEMICAL metaphor found by the system, comprising a pair of a facet
REGIME is a CONTAINER word and a target word taken from the conceptual metaphor.
REGIME is a ROOM (Note that such pairs are not necessarily linguistic metaphors

332
found originally in the data.) In addition, we included a num- statements would start by finding additional facets related
ber of foils, which were based on randomly matching facets to the nominal analogies for each conceptual metaphor. For
from one conceptual metaphor with target terms from differ- A REGIME IS A ROOM, for example, it would be nice to
ent conceptual metaphors. We constructed foils in this way see facets such as ”encompassing”, ”inescapable”, or ”af-
so as to give a meaningful baseline, however, this method fecting”. In addition we can cluster some of the facets such
also tends to increase the rate at which subjects will accept as ”burning/cold” into larger categories such as ”tempera-
the foils. ture”, which give clearer pictures of the different aspects of
Results for framing are given in Table 4. A large ma- the source concept that are being mapped.
jority of CM framing statements were judged valid by most
subjects, supporting the validity of the systems framing of Conclusions
conceptual metaphors. We have described a system which, starting from cor-
pus data, finds linguistic metaphors and clusters them to
Error analysis find conceptual metaphors. The system uses minimal back-
To better understand what is needed to improve the system, ground knowledge, consisting only of lexical semantics in
we consider selected conceptual metaphors proposed by the the form of Wordnet and Wiktionary, and generic syntac-
system and their component framing statements. tic parsing and part-of-speech tagging. Even with this fairly
Consider first the proposed conceptual metaphor AN knowledge-poor approach, experiments show the system to
IDEA IS A GARMENT. The framing statements given for effectively find novel linguistic and conceptual metaphors.
it are as follows: Our future work will focus on two main goals—
improving overall effectiveness of the system, and reducing
• An idea can be matching and a garment can be matching our already low reliance on background knowledge.
• An idea can be light and a garment can be light One key area for improvement is in the underlying LM
• An idea can be tight and a garment can be tight identification algorithms, by incorporating more linguistic
• An idea can be blue and a garment can be blue cues and improving the underlying syntactic processing. De-
• An idea can be old and a garment can be old veloping a good measure of system confidence in its clas-
sifications will also aid higher-level processing. More fun-
• An idea can be net and a garment can be net
damentally, we will explore implementation of an itera-
• An idea can be pink and a garment can be pink tive EM-like process to refine estimations at all three lev-
While the notion of an idea being like a garment may be els of metaphor representation. Hypothesized conceptual
plausible, with someone being able to put the idea on (con- metaphors can help refine decisions about what expressions
sider it true) or take it off (dismiss it), etc., many of the fram- are linguistic metaphors, and vice versa. We will also apply
ing statements are not clearly relevant, and a number seem clustering to the facets within conceptual metaphors to im-
to be simply mistaken (e.g., “old idea” is likely literal). prove framing by finding the key aspects which are mapped
We see a similar pattern for A REGIME IS A ROOM. between source and target domains.
The framing statements the system gives for this conceptual To reduce dependence on background knowledge, we
metaphor are as follows: plan to use unsupervised clustering methods to find seman-
tically coherent word groups based on corpus statistics, such
• A regime can be cold and a room can be cold
as latent Dirichlet allocation (Blei, Ng, and Jordan 2003)
• A regime can be old and a room can be old or spectral clustering (Brew and Schulte im Walde 2002),
• A regime can be supported and a room can be supported which will reduce or eliminate our reliance on Wordnet. To
• A regime can be shaped and a room can be shaped eliminate our need for Wiktionary (or a similar resource)
• A regime can be burning and a room can be burning to determine polysemy, we will apply statistical techniques
• A regime can be new and a room can be new such as in (Sproat and van Santen 1998) to estimate the level
of polysemy of different words from corpus data.
• A regime can be austere and a room can be austere Finally, to explore application of these methods to lan-
Again, the notion of thinking of a regime as a room is cogent, guages where corpus data may be limited, we will examine
with the members of the regime in the room, and people can methods for triangulating multiple partial data sources in a
enter or leave, or be locked out, and so forth. However, not given language/culture (such as combining information from
all of these framing statements make sense. One issue is that Dari, Classical Persian, and Quranic materials with Modern
some of the facets are not specific to rooms per se, but apply Iranian Persian) to improve performance where quantities of
to many physical objects (old, shape, burning, new). labeled or parsed training data may be limited.
The two main (interrelated) difficulties seem to be (a)
finding the right level of abstraction for the concepts in the Acknowledgements
conceptual metaphor, as related to the framing statements, This work is supported by the Intelligence Advanced Re-
and (b) finding good framing statements. search Projects Activity (IARPA) via Department of De-
One way to improve the system would be to first use the fense US Army Research Laboratory contract number
current system to find proposed CMs and associated fram- W911NF-12-C-0021. The U.S. Government is authorized to
ing statements, and then refine both the concepts in the CMs reproduce and distribute reprints for Governmental purposes
and the associated framing statements. Refining the framing notwithstanding any copyright annotation thereon.

333
Disclaimer: The views and conclusions contained herein are Krishnakumaran, S., and Zhu, X. 2007. Hunting elusive
those of the authors and should not be interpreted as neces- metaphors using lexical resources. In Proceedings of the
sarily representing the official policies or endorsements, ei- Workshop on Computational approaches to Figurative Lan-
ther expressed or implied, of IARPA, DoD/ARL, or the U.S. guage, 13–20. Association for Computational Linguistics.
Government. Lakoff, G., and Johnson, M. 1980. Metaphors we live by.
University of Chicago Press: London.
References Lewis, D.; Yang, Y.; Rose, T.; and Li, F. 2004. RCV1: A
Ahrens, K.; Chung, S.; and Huang, C. 2004. From lexical new benchmark collection for text categorization research.
semantics to conceptual metaphors: Mapping principle ver- The Journal of Machine Learning Research 5:361–397.
ification with Wordnet and SUMO. In Recent Advancement Mason, Z. 2004. CorMet: A computational, corpus-based
in Chinese Lexical Semantics: Proceedings of 5th Chinese conventional metaphor extraction system. Computational
Lexical Semantics Workshop (CLSW-5), Singapore: COL- Linguistics 30(1):23–44.
IPS, 99–106.
Nayak, S., and Mukerjee, A. 2012. A grounded cognitive
Allbritton, D.; McKoon, G.; and Gerrig, R. 1995. Metaphor- model for metaphor acquisition. In Twenty-Sixth AAAI Con-
based schemas and text representations: Making connec- ference on Artificial Intelligence.
tions through conceptual metaphors. Journal of Exper-
imental Psychology: Learning, Memory, and Cognition Pantel, P., and Lin, D. 2002. Discovering word senses from
21(3):612. text. In Proceedings of the eighth ACM SIGKDD interna-
tional conference on Knowledge discovery and data mining,
Assaf, D.; Neuman, Y.; Cohen, Y.; Argamon, S.; Howard, 613–619. ACM.
N.; Last, M.; Frieder, O.; and Koppel, M. 2013. Why “dark
thoughts” aren’t really dark: A novel algorithm for metaphor Rohrer, T. 2007. The cognitive science of metaphor from
identification. In Proc. IEEE Symposium Series on Compu- philosophy to neuropsychology. Neuropsychological Trends
tational Intelligence 2013. 2:31–40.
Birke, J., and Sarkar, A. 2006. A clustering approach for the Shutova, E.; Teufel, S.; and Korhonen, A. 2012. Statistical
nearly unsupervised recognition of nonliteral language. In metaphor processing. Computational Linguistics 39(2):1–
Proceedings of EACL, volume 6, 329–336. 92.
Blei, D.; Ng, A.; and Jordan, M. 2003. Latent dirichlet Sproat, R., and van Santen, J. 1998. Automatic ambiguity
allocation. the Journal of machine Learning research 3:993– detection. In Proceedings ICSLP.
1022. Thibodeau, P., and Boroditsky, L. 2011. Metaphors we
Boers, F. 2003. Applied linguistics perspectives on cross- think with: The role of metaphor in reasoning. PLoS One
cultural variation in conceptual metaphor. Metaphor and 6(2):e16782.
Symbol 18(4):231–238. Turney, P.; Neuman, Y.; Assaf, D.; and Cohen, Y. 2011. Lit-
Brew, C., and Schulte im Walde, S. 2002. Spectral cluster- eral and metaphorical sense identification through concrete
ing for german verbs. In Proceedings of the ACL-02 confer- and abstract context. In Proceedings of the 2011 Conference
ence on Empirical methods in natural language processing- on the Empirical Methods in Natural Language Processing,
Volume 10, 117–124. Association for Computational Lin- 680–690.
guistics. Wickman, S.; Daniels, M.; White, L.; and Fesmire, S. 1999.
De Marneffe, M.; MacCartney, B.; and Manning, C. 2006. A primer in conceptual metaphor for counselors. Journal of
Generating typed dependency parses from phrase structure Counseling & Development 77(4):389–394.
parses. In Proceedings of LREC, volume 6, 449–454.
Fass, D., and Wilks, Y. 1983. Preference semantics, ill-
formedness, and metaphor. Computational Linguistics 9(3-
4):178–187.
Freeman, M. 2003. Poetry and the scope of metaphor:
Toward a cognitive theory of literature. Metaphor and
Metonymy at the Crossroads: A Cognitive Perspective
30:253.
Gibbs Jr, R. 1992. Categorization and metaphor understand-
ing. Psychological review 99(3):572.
Goguen, J., and Harrell, D. 2010. Style: A computational
and conceptual blending-based approach. In Argamon, S.;
Burns, K.; and Dubnov, S., eds., The Structure of Style.
Springer-Verlag: Berlin, Heidelberg. 291–316.
Johnson, M.; Lakoff, G.; et al. 2002. Why cognitive lin-
guistics requires embodied realism. Cognitive linguistics
13(3):245–264.

334