Reasoning in Metaphor Understanding:
The ATT-Meta Approach and System
John Barnden, Sheila Glasbey, Mark Lee, Alan Wallington
School of Computer Science
University of Birmingham
Birmingham B15 2TT
{J.A.Barnden, S.R.Glasbey, M.G.Lee, A.M.Wallington}@cs.bham.ac.uk
Abstract
A detailed approach has been developed for
core aspects of the task of understanding a
broad class of metaphorical utterances. The
utterances in question are those that depend
on known metaphorical mappings but that
nevertheless contain elements not mapped
by those mappings. A reasoning system has
been implemented that partially instantiates
the theoretical approach. The system, called
ATT-Meta, will be demonstrated. The paper
briefly indicates how the system works, and
outlines some specific aspects of the system,
approach and the overall project.
Introduction
The sentence In the far reaches of her mind,
Anne believed that Kyle was having an affair1
can be analyzed as depending on metaphorical
views of MIND AS PHYSICAL SPACE and
IDEAS AS PHYSICAL OBJECTS (see Barnden
2001a). These views are, plausibly, familiar to
typical users of English. However, it is
reasonable to assume that typical users do not
already have a mapping into the mental domain
of the physical notion of "far reaches". Our
approach to metaphor is predicated on the notion
that one should, when possible, avoid
constructing source-to-target mappings for such
elements of a metaphorical utterance that
transcend the already known mappings in the
metaphorical views underlying the utterance.
Instead, we advocate the use of metaphor-
pretence "cocoons" (reasoning spaces) where the
utterance is taken as literally true. Within-
                                                     
1 Slightly adapted from real discourse.
cocoon reasoning will attempt to connect these
"map-transcending" elements to aspects of the
source using a set of specified conversion rules.
The far-reaches qualification in the sentence
implies by source-domain reasoning that Anne
could only to a very low degree physically
manipulate the idea that Kyle was having an
affair. Then, if we can appeal to a conversion
rule, i.e. a known mapping of ability to
physically manipulate an idea to ability to
consciously process it, we can draw the
reasonable conclusion that Anne only had a very
low degree of ability to consciously process the
idea. In our presentation, we shall demonstrate
how the ATT-Meta system deals with this
example.
Note that the rules of reasoning are given a
qualitative certainty level, and that predicates
can be graded, using a scale of qualitative
degrees. For instance, someone can be
represented as understanding a situation to a
"medium" degree.
Our approach also makes heavy use of "view-
neutral mapping adjuncts" (VNMAs). These are
general mapping principles (inspired by the
work of Carbonell 1982) that apply, though only
by default, no matter what metaphorical views
are in play. For instance, the ability to do things
and the degrees with which states of affairs hold
are automatically mapped by VNMAs. In many
examples of metaphor, most of the real mapping
work is done by VNMAs.
Much of the approach has been implemented in
the ATT-Meta system, which is an uncertain
rule-based system operating by backchaining
(see also Barnden 1998, Barnden 2001, Lee &
Barnden, 2001a). ATT-Meta performs
reasoning, but does not yet interface directly to
natural language. Instead, hand-constructed
logical forms couching the source-domain
meaning of metaphorical sentences are passed to
it. In the above example, the source-domain
meaning is that Anne’s believing was literally
physically located in the physical far reaches of
her mind.
The following sections summarize various
abilities of the system, principles of the
approach, and aspects of ongoing theoretical
work aimed at further extensions to the system.
A major item of current implementational work
is a fuller realization of VNMAs.
1 Uncertainty
Although reasoning conflict and uncertainty are
intricately involved in metaphor, very few
approaches attempt to grapple with the issues.
Propositions and reasoning within both the
target and source domains, being largely of
common-sense varieties, are typically uncertain.
It can be uncertain what metaphorical views are
involved; information transferred from the
source domain can conflict with target-domain
information; and transfers can even conflict with
each other. The ATT-Meta system handles all
these types of uncertainty and conflict. Its
uncertainty handling is based on fairly crude
qualitative uncertainty annotations on rules and
propositions, but there is a sophisticated
conflict-resolution mechanism.
The uncertainty-handling and conflict-resolution
are almost entirely orthogonal to the provisions
for metaphor. This leads to clean design and
helps to address long-standing issues about
metaphor. One such issue is the conflict between
information transferred from the source domain
and the target information. ATT-Meta allows
either side to win, depending on standard
specificity principles. This goes against a naive
assumption in most of the literature that target
information should automatically override
transfers. But, this is only convincing when the
target information is certain. Indeed, we claim
that metaphor is often used precisely to describe
an exception to a target-domain default.
2 Mixed Metaphors
Issues such as reasoning about uncertainty are
particularly important in the processing of mixed
metaphors. Mixed metaphors need not feature
obvious cases of conflict but can include
graceful combinations of metaphors, such as the
following sentence to be examined below: One
part of John hotly resented the verdict. This
combines a view of John as made up of sub-
agents and a view of agents’ emotional states as
things that can have temperature. It is possible to
distinguish two types of mixed metaphor:
parallel mixes and serial mixes. In a parallel
mixed metaphor, the target (A) is seen partly
through an A-as-B metaphor and partly through
another metaphor, A-as-B’. B and B’ are in
general different domains, but may overlap.
Also, different aspects of A may be involved in
the two metaphors. In a serial mixed metaphor
(commonly called a chained metaphor), the
target (A) is seen as a source (B), which is in
turn then seen as a different source (C).
Previous work on the understanding of metaphor
has assumed that mixing is a relatively rare
phenomenon that can be handled once a more
theory of simple metaphor is developed. We
argue that this assumption is detrimental to
progress since mixed metaphors rely on the
same conceptual knowledge as simple
metaphors and can, therefore, provide valuable
insight into the processes and representations
underlying metaphorical reasoning. Moreover,
we claim that the reasoning processes and data
structures involved in understanding mixed
metaphors are identical to those used in
understanding simple metaphors. Therefore, any
current theory of metaphor should (at least in
principle) be extensible to deal with mixing. To
this end, ATT-Meta handles mixed metaphor in
a manner consistent with the way it handles
simple metaphors. The two types of metaphor
are processed in subtly different ways. Parallel
mixed metaphors create separate pretence-
cocoons that are mapped in parallel to the target
domain where their respective contributions are
understood. Serial mixed metaphors create
nested pretence cocoons where the metaphorical
view of B as C is nested within a pretence
cocoon with the view of A as B.
3 Reverse Transfers in Metaphor
The use of metaphor involves a flow of effects
of some kind from the source domain to the
target domain, where effects can include insights
into the target, hypotheses about the target, or
the highlighting of parts of the target. However,
although the overall effect flow is always from
source to target, in many cases, this does not
preclude a reverse flow where a literal
proposition, command, or question is mapped
onto an equivalent within the current
metaphorical domain.
The ATT-Meta system allows conversion rules
to map from propositions in the source domain
to propositions in the target domain and also in
the opposite direction. So a source domain
proposition such as "Socrates was the midwife
for an idea" might be mapped onto the target
domain proposition "Socrates helped in the
production of the idea". However, the rules
would equally allow the proposition "Socrates
helped in the production of the idea" to be
mapped to the source domain proposition
"Socrates was the midwife for an idea". We
argue that there are at least three reasons why
ATT-Meta should have this ability:
(1) Given that metaphors are ultimately used to
have an effect on the target domain, the use of a
metaphorical utterance can be seen as
answering, in some sense, a target domain
query. This sets up a choice between taking the
metaphorical utterance and applying all
conversion rules to it in the hope that one of the
resulting propositions might provide a suitable
answer, or taking the question and converting it
into a question in terms of the current metaphor.
We argue that the latter is often more efficient.
(2) Certain source domain propositions would
allow ATT-Meta to draw a tentative conclusion,
which would, were it more strongly supported,
provide an argument via a chain of reasoning for
some other, target level, proposition or query. A
target-level statement might give the added
support, but for this to be the case it would first
need to be converted into its source-level
equivalent.
(3) The combination of source and target domain
information within a discourse that only
intermittently maintains a metaphorical view of
the target domain may best be done in the source
domain after the target domain information has
been "metaphorized". This would be especially
so if the source domain was information-rich
compared to the target domain, so allowing
much more reasoning to be carried out than
would be possible in the target domain.
4 Non-Declarative Metaphor
Almost all examples of metaphorical language
discussed in the literature are of declarative
utterances rather than questions, commands,
ejaculations, etc. However, these other forms of
utterance can obviously occur. For instance, just
as one can state "John is a steamroller’’ one can
ask "Is John a steamroller?’’ Just as one can state
"The champion knocked the cream-puff out’’ one
can issue the command "Knock that cream-puff
out!’’ The observation that questions, in
particular, can be metaphorical, plays a
significant role in our theoretical approach. This
is because their processing is contiguous with
that of implicit queries generated within the
metaphorical pretence cocoon (see Introduction)
during ATT-Meta’s goal-directed reasoning.
However, the theoretical significance of non-
declarative metaphorical utterances is even
greater, because such utterances call into
question accounts of metaphor that assume the
task of understanding is to work out what claim
about the target domain the metaphorical
utterance is making.
Compiling such examples is an additional goal
of our corpus work (see section 6).
5 Time and change
Work is ongoing which addresses the temporal,
aspectual and causal facets of metaphor. A
survey of metaphors in the ATT-Meta Databank
reveals, unsurprisingly, that the metaphorical
expressions there involve a wide range of tense
and aspectual constructions in English, including
past, present and future tenses, simple and
progressive aspects, and the full set of aspectual
classes. A wide variety of temporal adverbials is
also present. A key topic under investigation is
the mapping of temporal and aspectual
information between source and target domains.
For example, if an event is telic in the source
domain, to what extent does that telicity carry
over to the target domain? Preliminary
investigations confirm the expectation that such
aspectual information is preserved in the
majority of cases. Exceptions exist, however,
and these merit further study.
The mapping of temporal duration between
domains is also being investigated. In some
cases, a mapping appears to exist whereby an
event of long duration relative to the source
domain maps to an event with long duration
relative to the target domain. This can be
captured by an appropriate VNMA, which maps
relative durations between domains. The logic of
ATT-Meta is episode-based, which means that it
is relatively straightforward to express this kind
of constraint and employ it in reasoning.
Currently underway is a detailed examination of
metaphorical expressions involving both explicit
and implicit temporal durations. This will result
in a set of VNMAs covering a wide range of
tense/aspect/temporal-adverbial constructions.
A second strand of the work on time involves a
detailed study of the metaphors used to describe
times, states and events, including spatial
metaphor for time (Lakoff 1994).
6 Corpus Studies of Metaphor
As an adjunct to the development of the ATT-
Meta approach and system, we have been
conducting corpus studies of metaphor, mainly
using the British National Corpus but also using
the Bank of English and, to a limited extent, web
search engines. We have used both hand-
annotation of small numbers of documents from
the BNC and automated search for particular
types of metaphorical phraseology (mainly
relatively fixed metaphorical phrases concerning
mental states) over the whole of corpora.
Current objectives are (a) to develop large
databanks of examples of various types of
metaphorical utterance, for the benefit of
metaphor researchers in general, (b) to
demonstrate more extensively and objectively
the importance in discourse of "map-
transcending" metaphorical utterances (see
Introduction), (c) relatedly, to reveal the degree
to which relatively conventional metaphor
phraseology can be varied in real discourse (cf.
Moon 1998), and (d) to uncover (in small
numbers of documents) the degree to which
metaphorical utterances relate to context: how
much their understanding depends on context
and how much the understanding of the context
depends in turn on them. We are interested in (d)
because in the ATT-Meta approach the process
of metaphorical understanding is partially
guided by discourse goals set up by context.
This feature goes a long way to side-stepping
problems of apparent indeterminacy of meaning
of metaphorical utterances when taken in
isolation.
We also have the methodological objectives of
developing a good annotation regime for
metaphor and better-automated search
techniques for metaphor. As part of the latter,
we plan to investigate the usefulness of a large
set of morphological, lexical, syntactic and
phraseological clues to the presence of
metaphor, inspired by the clues discussed by
Goatly (1997). These clues are only present in a
minority of metaphorical utterances but could
nevertheless form a useful weapon in the
automated search armoury.
Metaphor detection techniques developed for
corpus study should also help with developing a
means for an understanding system to notice the
presence of metaphor. Such noticing is not
currently performed by ATT-Meta but is an
important topic for future research.
Conclusion
The ATT-Meta project is making headway in
showing how metaphorical utterances can be
computationally processed. It is based on a
distinctive set of principles as to how to
understand metaphor, some of which are original
and some related to those of previous
researchers. In particular, it seeks to avoid
expensive computation of new analogical
mappings between domains as a regular part of
metaphorical understanding. This is inspired
partly by the observation that genuinely novel
pairings of domains are relatively rare in real
discourse. What are more common are novel
extensions of familiar metaphorical views, and
novel mixes of views. This is true even in poetry
(Lakoff and Turner 1989). The project is also
seeking to take full account of the important role
that uncertainty, gradedness and dynamism of
situations plays in metaphor.
The approach and system have been evaluated in
a number of ways. We have applied the
implemented system or the theoretical approach
to (simplified versions of) selected real-
discourse examples from an existing databank
(http://www.cs.bham.ac.uk/~jab/
ATT-Meta/Databank): see Barnden (2001a),
Barnden & Lee (2001a) and Barnden & Lee
(2001b). We have applied the implemented
system to examples of all the metaphors of
mental states listed in the Master Metaphor List
(Lakoff 1994, Lee & Barnden 2001a). The
examples here were found by search over the
Bank of English. Finally, we have applied the
theoretical approach to various real-discourse
examples included in Goatly (1997): see
Barnden (2001b).
Acknowledgements
This research is supported by grant GR/M64208
from the Engineering and Physical Sciences
Research Council of the UK.

References

Barnden, J. A. (1998a)  Combining uncertain belief
reasoning and uncertain metaphor-based
reasoning. In "Procs. Twentieth Annual Meeting of
the Cognitive Science Society", Lawrence Erlbaum
Associates, Hillsdale, N.J, pp.114-119.

Barnden, J.A. (2001a)  Application of the ATT-Meta
Metaphor-Understanding Approach to Various
Examples in the ATT-Meta Project Databank.
Technical Report CSRP-01-02, School of
Computer Science, The University of Birmingham,
U.K.

Barnden, J.A. (2001b)  Application of the ATT-Meta
metaphor-understanding approach to selected
examples from Goatly. Technical Report CSRP-01-
01, School of Computer Science, The University of
Birmingham, U.K.

Barnden, J.A (2001c)  Uncertainty and conflict
handling in the ATT-Meta context-based system for
metaphorical reasoning. In “Lecture Notes in
Artificial Intelligence”, J.G.Carbonell and
J.Siekmann, eds, Springer Verlag, Berlin.

Barnden, J.A. and Lee, M.G. (1999)  An implemented
context system that combines belief reasoning,
metaphor-based reasoning and uncertainty
handling. In " Modelling and Using Context", P.
Bouquet, P. Brezillon, L. Serafini ed., Lecture
Notes in Artificial Intelligence, No. 1688,
Springer-Verlag, pp. 28-41.

Barnden, J.A. and Lee, M.G. (2001a)  Understanding
open-ended usages of familiar conceptual
metaphors: An approach and artificial intelligence
system. Technical Report CSRP-01-05, School of
Computer Science, The University of Birmingham,
U.K.

Barnden, J.A. and Lee, M.G. (2001b)  Application of
the ATT-Meta metaphor-understanding system to
an example of the metaphorical view of MIND
PARTS AS PERSONS. Technical Report CSRP-01-
09, School of Computer Science, The University of
Birmingham, U.K.

Carbonell, J.G. (1982)  Metaphor: an inescapable
phenomenon in natural-language comprehension.
In "Strategies for Natural Language Processing",
W. Lehnert & M. Ringle ed., Lawrence Erlbaum,
Hillsdale, N.J, pp.415-434.

Goatly, A. (1997)  The language of metaphors.
Routledge, London and New York.

Lakoff, G. (1994)  The Master Metaphor List.
http://cogsci.berkeley.edu/, University of
California, Berkeley.

Lakoff, G. and Turner, M. (1989)  More than cool
reason: a field guide to poetic metaphor.
University of Chicago Press, Chicago.

Lee, M. G. and Barnden, J. A. (2001a)  Reasoning
about mixed metaphors within an implemented AI
system. Metaphor and Symbol 16, 1& 2, pp. 29-42.

Lee, M. G. and Barnden, J. A. (2001b)  Mental
Metaphors from the Master Metaphor List:
Empirical Examples and the Application of the
ATT-Meta System Technical Report CSRP-01-03,
School of Computer Science, The University of
Birmingham, U.K.

Moon, R. (1998)  Fixed idioms and expressions in
English. Clarendon Press, Oxford, U.K.
