Transcript Automatic Metaphor Interpretation as a Paraphrasing Task

Automatic Metaphor Interpretation
as a Paraphrasing Task
Ekaterina Shutova
Computer Lab, University of Cambridge
NAACL 2010
Outline
•
•
•
•
•
•
What is metaphor
The idea and overview of the system
Experimental data
Method
Evaluation
Conclusion
Metaphor in this paper
• Metaphor arise when one concept is viewed
in terms of the properties of the other.
– News travels fast.
– How can I kill the process?
– And then my heart with pleasure fills, And dances
with the daffodils.
Automatic processing of metaphor
• Metaphor recognition
– Distinguishing between literal and metaphorical
language in text.
• Metaphor interpretation
– Indentifying the intended literal meaning of a
metaphorical expression.
Interpretation as paraphrasing
• Phrases
– All of this stirred and uncontrollable excitement in
her.
– A carelessly leaked report.
• Their paraphrases
– All of this provoked an uncontrollable excitement
in her.
– A carelessly disclosed report.
System overview
• Produces a list of all possible paraphrases for a
metaphorical expression.
• Ranks the paraphrases according to their
likelihood derived from the corpus.
• Discriminates between literal and figurative
paraphrases by detecting selectional preference
violation and output the literal ones.
• Disambiguates the sense of the paraphrases
using WordNet inventory of senses.
Experimental data
• Focused on single-word metaphors expressed
by a verb.
– Classification according to whether the verbs are
used metaphorically or literally.
– Some verbs have weak or no potential of being a
metaphor are excluded.
•
•
•
•
Auxiliary verbs
Model verbs
Aspectual verbs
Light verbs
Dataset
• The corpus is a subset of the British National
Corpus (BNC) that contains 761 sentences and
13642 words.
• Annotated based on the principles of the
metaphor identification procedure (MIP).
– A form of word sense disambiguation with an
emphasis on metaphoricity.
Annotation
• Discriminate between verbs used metaphorically
and literally.
– For each verb establish its meaning in context and try
to imagine a more basic meaning of this verb on other
contexts.
•
•
•
•
More concrete
Related to bodily action
More precise (as opposed to vague)
Historically older
– If you can establish the basic meaning that is distinct
from the meaning of the verb in this context, the verb
is likely to be used metaphorically.
Example
• If he asked her to post a letter or buy some
razor blades from the chemist, she was
transported while pleasure.
– The verb transport in its basic sense is used in the
context of “goods being transported/carried by a
vehicle”.
Phrase selection
• Only the phrases that were tagged as metaphorical by all
the annotators were included in the test set.
– memories were slipping away.
– hold the truth back.
– factors shape results
• Some phrases are excluded from the test set
–
–
–
–
Subject or object referent is unknown.
Metaphorical meaning is realized solely in passive constructions.
Subject or object of interest are represented by a named entity.
Multiword metaphors.
• The resulting test set contains 62 metaphorical expression.
Paraphrase ranking
• The likelihood of a particular paraphrase as a
joint probability
– Carelessly [w1] leaked [i] report [w2]
Paraphrase ranking
• Rank the possible replacements of the term
used metaphorically in the fixed context
according to the data by the likelihood.
Filter the paraphrases
• Filter out the terms whose meaning does not
share any common features with that of the
metaphorical term.
– Kill a process <-> Terminate a process
• Kill implies an end or termination of life
• Identifying shared meanings using hyponymy
relations in the WordNet taxonomy.
Filtering based on selectional
preferences
• Remove the irrelevant paraphrases and
paraphrases where the substitute is used
metaphorically again.
– Suppress the truth.
• The verbs used metaphorically are likely to
demonstrate strong semantic preference for
the source domain than for the target domain.
– Suppress would select for movements (political)
rather than ideas or truth (target domain).
Association measure
• Resnik (1997) models selectional preference
of a verb in probabilistic terms as the
difference between the posterior distribution
of noun classes in a particular relation with
the verb and their prior distribution using the
relative entropy (Kullback-Leibler distance):
Evaluation
• 7 annotators were asked to mark the ones that
have the same meaning as the term used
metaphorically and are used literally in the
context of the paraphrase expression as correct.
• Agreement is 0.62 (k)
• Accuracy measure the proportion of correct
literal interpretations among the paraphrase in
rank 1.
Conclusion
• A system that produces literal paraphrases for
metaphorical expressions.
• Directly transferable to other applications that
can benefic from a metaphor processing
component.
• Does not rely on any hand-crafted knowledge
other than WordNet.