Lessons from an Artificial Intelligence Research Project on Metaphor

1
Lessons froman Artificial Intelligence Research
Project on Metaphor

• John Barnden
• School of Computer Science
• University of Birmingham, UK
• Collaborators
• Sheila Glasbey, Mark Lee, Alan Wallington

2
Plan of Talk

• Some general observations
• The ATT-meta approach and system
• Lessons regarding language differences/learning

3
What I Take Metaphor to Be

• Talking/thinking/ about something (the target)
  as if it were something else (the source).
• Taking a Metaphorical view.
• Some examples
• My work will spill over into the weekend.
• Thatcher was the Reagan of Britain.
• Managerialism is creeping into academia.
• Metaphorical views involve mappings.

4
Cross-Linguistic/Cross-Cultural Issues

• Many metaphorical views are common across
  languages.
• e.g. Problem/Issue As Physical Object
• special case Major Issue As Iceberg King 2004
• But some arent
• e.g. Papafragou 1998
• Idea as Edible Object not allowed in Greek.
• Littlemore Low, forthcoming
• English hat metaphors dont translate well into
  French
• And even when they are, word-for-word translation
  often does not give good results King 2004.
• drop of water on a hot stone from German phrase
• behind the tip of the iceberg visible sides of
  the iceberg from Greek phrases

5
Types of Metaphoricity in Language

• When the utterance uses familiar metaphorical
  views
• AND the source aspects used are mapped by them
• AND the wording is standard
• One part of me thought that I should go.
• When the utterance uses familiar metaphorical
  views
• AND the source aspects used are mapped by them
• BUT the wording is not standard
• One component of me thought that I should go.

6

• MAP TRANSCENDING
• When the utterance uses familiar metaphorical
  views
• BUT
• some source aspect is NOT mapped by them
• One part of me was insisting that I should go.
• SnakeByte technologies gobbled up RabbitWare and
  spat its managers out.

7

• NOVEL
• When the utterance does not use familiar views
• John unpeeled the temperature.
• The road was a dolphin.

8
Map-Transcending Metaphor, contd.

• He dredged up his mud-encrusted memories.
• In the far reaches of her mind, Anne knew that
• Men arent islands, but some are peninsulas.
• The middle managers have cricks in their necks
  from talking down to the workers and up to the
  bosses.

9
Language-User Relativity(some links to Cameron
1999, Geeraerts 2002, Ruiz de Mendoza Ibáñez
1999, Radden 2002, Radman 1997, Riemer 2002)

• Relativity of familiarity of views / mappings
• Relativity of lexical senses and standardness of
  wording
• Hence, relativity of
• whether an utterance is actively metaphorical
  (i.e. requires source/target mapping actions)
• what metaphorical views an utterance rests on
• whether an utterance is map-transcending, novel,
  etc.

10
Domains in Metaphor Problems(cf. Barcelona
2002, Cameron 1999, Kittay 1989, Lemmens 2001,
Riemer 2002)

• Domain divisions are context-sensitive and
  arbitrary
• Peter is a fox
• The idea lurked in his mind
• Thatcher was the British Reagan
• Christmas is on the horizon
• The idea had sunk slowly into his subconscious.

11
Domains in Metaphor contd

• Source and target domains can massively overlap
• e.g. -- Mind Parts As Persons
• One part of me says I should go to the party,
  another part insists I should do my tax form.
• Other types of overlap/arbitrariness where
  should the following facts go??
• Minds are not containers.
• Cars can be the setting for passionate love.

12
View-Neutral Mapping Adjunctscf. Carbonell,
(Winston)

• Emotions, value judgments and mental states are
  often implicitly transferred from source to
  target in metaphor in general. (May even be a
  primary function of metaphor.)
• Managerialism is sneaking into academia.
• Poverty is a disease.
• Were conducting a war on terrorism.
• The emotions, etc. can be of agents in the
  source, rather than of the understander.

13
Other VNMAs

• uncertainty
• degrees (intensities)
• causation, enablement, ability, ease, etc.
• event shape, temporal relationships
• sets, qualitative set sizes.

14
VNMAs, contd

• John and Mary are in a race with each other at
  work.
• Rests on Abstract Process as Physical Journey.
• John Mary are viewed as being in a race.
• So each intends to win that race, i.e. to finish
  the race journey first.
• The finishing and the first-ness map by VNMAs.
• So each intends to finish their work first, by
  another VNMA.

15
ATT-Meta Approach

• Reasoning approach for metaphor processing.
• Suited to linguistic non-linguistic metaphor.
• Aimed mainly at map-transcending metaphor. Does
  not discover mappings.
• Exploits view-neutral mapping adjuncts.

16

• Emphasizes degrees (gradations).
• Emphasizes qualitative uncertainty.
• Allows source information to override target.
• Allows combinations of metaphorical views.
• (Parallel serial.)

17
ATT-Meta System

• (Partially) implements the approach.
• Rule-based.
• Much attention to uncertainty handling.
• Metaphor-orientated reasoning thoroughly
  integrated into overall reasoning.
• Metaphor override phenomena absorbed into general
  conflict-resolution approach.

18
Basic Method in ATT-Metacf. Carbonell, Hobbs,
Narayanan, (Lakoff, Martin)

• EXAMPLE
• In the far reaches of her mind, Anne knew that
  Kyle was having an affair. Cosmopolitan, 1994
• Mind as Physical Space,
• Ideas as Physical Objects (implicitly)
• System knows VIEW-SPECIFIC MAPPING
• physical manipulation of ideas ? conscious
  mental usage

19

• PRETEND that the utterances source-domain
  meaning is true.
• REASON (VIA SEVERAL STEPS) within a special
  computational pretence environment that,
• presumably, Anne can physically operate upon the
  Kyle-affair idea only to a very low degree.
• Apply VIEW-SPECIFIC MAPPING VNMAs
• presumably, Anne can consciously mentally use the
  Kyle-affair idea only to a very low degree.

20
Another Example

• A part of Mary was insisting that she was
  right.
• Mind Parts as Persons
• a part believes X ? whole agent has motivation to
  believe X
• REASONING in the PRETENCE
• The mentioned part believes that Mary is right.
• There is another part that has stated that Mary
  is not right.
• That other part believes that Mary is not right.
• Result of KNOWN MAPPING
• Mary has a motivation to believe she is right.
• Mary has a motivation to believe she is not right.

21
Map-Extension Minimization

• Metaphor understanding should try to avoid
  creating new mapping relationships for
• map-transcending aspects of the utterance
• far reaches, mud-encrusted, neck cricks,
  peninsulas, insisting,
• general source-domain knowledge exploited
• Much within-pretence knowledge and reasoning
  serves merely to facilitate and warrant the
  application of already-known mapping
  relationships (view-specific VNMAs).

22
Where Mappings Go

• They go from pretence environments to surrounding
  environments (usually reality)
• NOT from source domains to target domains.
• If were pretending that an idea is a physical
  object, and that X physically manipulates it,
• then (in reality) X is consciously using it.
• This stance sidesteps the problems with domains.

23
Context-Sensitivity(Leezenberg 1995, Stern 2000,
)

• Peters a tank.
• What might this convey??
• KEY
• The example is only likely to arise in an already
  established context.

24
Context-Drivenness

• Peters colleagues are badly affected by
  criticism, but hes a tank.
• First clause and the but raise issue of how
  badly affected PETER is by criticism.
• Suppose we know a metaphorical mapping from
  physical attack in a battle to criticism.
• Then, in a metaphorical pretence, raise issue of
  how badly affected Peter-as-tank is by physical
  attack.
• Address the issue by knowledge about tanks.
• NB Backwards use of metaphorical mapping.

25
Context-Drivenness in ATT-Meta

• In ATT-Meta, reasoning is directed backwards
  from goals (a standard technique in AI).
• Enables metaphor understanding to be
  context-driven.
• In the far reaches of her mind, Anne knew Kyle
  was having an affair, but to acknowledge the
  betrayal to herself would have meant she would
  have had to take a stand.
• acknowledge to herself and the but raise the
  issue of Annes conscious awareness. Mapped
  backwards to within-pretence issue of physical
  manipulability.

26
Lessons (Suggestions) concerning Language
Differences, Learning, etc.

• Emphasis on
• within-pretence reasoning
• view-neutral mapping adjuncts, and hence
• small number of view-specific mappings per view,
• allows
• Relatively little learning of view-specific
  mappings in L2

27

• Some issues raised
• To what extent are view-neutral mapping adjuncts
  (VNMAs) universal?
• (Many discussions assume that, e.g., value
  judgments are carried over in any language.)
• We need to attend to different types of value,
  emotion, mental state, event conceptions, etc. in
  different languages/cultures.

28

• Littlemore Low (forthcoming) stress connections
  between metaphoric thinking and learning
  L2-culture-specific within-source connotations of
  words/concepts.
• Silence is golden.
• Our emphasis on VNMAs (if largely universal) and
  on avoiding new mappings allows
• Focussing on specific sorts of connotation those
  that connect to known mappings

29

• Language-user relativity even within a given
  language/culture means
• There is less pressure on teachers/learners to
  attend to culture-specific metaphorical language,
  mappings and domain divisions as opposed to
• culture-specific NON-metaphorical connotations
  (golden ? valuable, good)
• and
• general principles of metaphoric processing.

30

• Our focus on uncertainty reveals that
  metaphor-derived information often overrides
  target defaults.
• The company nursed its competitor back to
  health.
• (Conjecture This is a major function of
  metaphor.)
• Metaphor is therefore even more important in
  language learning than it would be otherwise.

31

• An emphasis on context means that
• Teachers/learners can rest assured that usually
  context will help a lot in understanding metaphor
  (especially when unconventional).
• Best not to base metaphor-related learning on
  isolated sentences.
• Caution students may not use the right part of
  the context Littlemore

32
Conclusions

• Importance of map-transcendence in mundane uses
  of metaphor.
• Treatment by a reasoning-heavy approach without
  (in general) creation of new mappings.
• Nice fit with a pretence view of metaphor, which
  also sidesteps domain problems.
• Importance of relativity, uncertainty, degrees,
  context, and view-neutral mapping principles.
• Some suggestions regarding cross-culture/cross-lan
  guage differences/learning/teaching.
• Heightening of the problem of the
  metaphor/metonymy distinction.

33
Variation of Metaphorical Idiomcf. Moon 1998

• in the recesses of Xs mind could be in a
  lexicon, a WordNet, etc.
• But productive variation is possible
• in the dim recesses of Xs mind
• in the deep recesses of Xs mind
• in the distant recesses of Xs mind
• in the unlit recesses of Xs mind
• Such variation is often map-transcending.

34
Other Practical Features of Metaphor

• Metaphor as exception-handler. My conjecture.
• She knew it in the dark recesses of her mind.
• Copular metaphor (A is B) as summarizer
  Kupferberg Green 1998, Drew Holt 1995.
• Marys a real bulldozer. Do you know what she did
  
• Metaphorical idioms/proverbs as topic pivots
  Drew Holt 1995.
• Yep, too many cooks spoil the broth. Now what I
  wanted to say was
• Meta-discourse metaphor Cameron 1998.
• Lets circle back now to the first principle I
  mentioned.

35
Lessons from Being Computational

• Context-drivenness.
• Grappling with the problems with domains.
• Settling on pretence/reality distinction as
  opposed to source/target domain distinction.
• Importance of uncertainty and reasoning-conflict
  resolution.
• Metaphorical inference can defeat target
  defaults.
• Importance of degrees.
• Need for view-neutral mapping adjuncts.

36
Cross-Linguistic/Cultural Issues, contd.

• Productivity of metaphor.
• Example could learn a translation for the phrase
  race condition in distributed computing but
  then words such as win, defeat, tie, drop
  out etc. etc. need to be treated systematically.

37

• No need to have view-specific mapping that deals
  with degrees of , ability etc. of physical
  manipulation of ideas.
• Contrast with a Lakoff conceptual metaphor
• ((FILL IN))

38
The Method Not!

• No necessary reliance on LITERAL meaning. A
  source-domain meaning is not necessarily
  literal.
• But even when source-domain meaning can be said
  to be literal, the method is NOT literal-first.