100 likes | 211 Views
Tim Rumbell , John Barnden (speaking), Mark Lee & Alan Wallington School of Computer Science University of Birmingham. Affect in Metaphor: Developments with WordNet. Initial Motivating Context.
E N D
Tim Rumbell, John Barnden (speaking), Mark Lee & Alan Wallington School of Computer Science University of Birmingham Affect in Metaphor: Developments with WordNet
Initial Motivating Context • Online affect detection from text by an automated conversational agent, in contexts where considerable inaccuracy is tolerable. • But the intentions are broader/deeper/higher/sharper. • Relatively shallow techniques used (though involving parsing and some semantic analysis) … • … but intended to be consistent with our deeper theory of metaphor understanding (ATT-Meta).
Metaphor and Affect • Affect and metaphor are important to each other: • Affect is often conveyed/described metaphorically • Metaphor is often affective • Emotion states are often described metaphorically • “He was boiling inside” [not discussed here] • Affect of metaphorical source terms typically carries over • “My son's room is a bomb site” [this phenomenon underlies aspects of the present talk] • Both phenomena are important aspects of ATT-Meta approach.
Metaphorical Phenomena for this talk • Someone as an animal • “You piglet” • Someone as a supernatural being • “You’re an angel” • Someone as a special type of human • “Lisa is such a baby” • [This case not yet addressed] • Metaphorical use of size adjectives • “You big/little bully”, “Mike is a little rat”
Examples of Results of Current Implemented System • “You cow” • negativeanimal metaphor • “She's an absolute angel” • positivesupernatural being metaphor • “You are a little rat” • negativeanimal metaphor with added contempt • “You piglet” • negativeanimal metaphor meant affectionately • “He is an elephant” • positive-or-negativeanimal metaphor • “He's a rock” • positivenatural object metaphor (NEW) • “She’s a bit of a bag” • negativeartefact metaphor (NEW)
The Recognition Component • Heuristic metaphoricity signals looked for: • X is/are Y • You Y • '[looks] like', 'a bit of a ', 'such a' • Signals detected using the RASP robust parser (with some post-processing) • Information extracted: • X (pro)noun • Y noun • Y noun’s modifiers
WordNet-based Analysis Component • Animal • Chordate • Vertebrate • Mammal • Ungulate • Swine • Person • Unwelcome person • Unpleasant person • Selfish person • (a person who is unusually selfish) • Person • Unwelcome person • Unpleasant person • (a person who is not pleasant or agreeable) • Vulgarian • (a vulgar person) • Pig (domestic swine) • Pig (a coarse obnoxious person) • Pig (a person regarded as greedy and pig-like) • Piglet (a young pig) “You piglet”
Little: If negative metaphor: Contempt added to evaluation If positive metaphor ORaffection already added (= through baby animal metaphor): (Extra) Affection added to evaluation Size Adjectives • Big: • Emphasis added to existing metaphorical evaluation
Problems and Ongoing/Future Work • Only searching for individual words in WN glosses: no parsing etc. of them yet. • Positive/negative feature counting is simplistic! • For non-WN-metaphorical animals etc.: affective carry-over needs more sophisticated affective-feature selection than our current one. • Go beyond metaphorical animals and supernatural beings (and newly: natural objects and artefacts). In particular, add special types of human (baby, freak, lunatic, etc.). • Improve/generalize the size-adjective processing. • Integrate processing with ATT-Meta system.