Words with AttitudeJaap KampsMaarten MarxPaper’s Goaln Judge the emotive or affectivemeaning of a textn Use WordNet to determine values ofwords with Osgood’s semanticdifferential techniqueOsgood’s SemanticDifferential Techniquen Judge words, phrases, texts by askingsubjects to rate them on scales of bipolaradjectivesn A subject might be asked to rate “proper”on scales like optimistic-pessimistic,serious-humorous, and active-passive.n It turns out that good-bad, strong-weak,and active-passive values account formost variance in judgmentUsing WordNet withOsgood’s theoryn Authors want to get values forwords from WordNetn They define MPL(w1,w2) as theminimal path length between w1 andw2, using only same-synset relationsn Allowing more than just same-synset damages metricMPL Examplesn MPL(good, proper) = 2n (good,right,proper)n MPL(good, neat) = 3n MPL(good, noble) = 4n Can we use this to rate “proper”,“neat”, and “noble” on a good-badscale?MPLn MPL(good, bad) = 4n If we just look at MPLs, “noble” is asgood as “bad”n We need to do something a bit morecomplicatedTRIn To determine the good-bad (“evaluative”)value of wi, examine TRI(wi;good,bad)n Define EVA(w) = TRI(w;good,bad)),(),(),();( ,jkjikikjiwwMPLwwMPLwwMPLwwwTRI−=EVA resultsn There are 5410 adjectives linked to “good”or “bad”.n Average value of EVA for these 5410words is –0.00891440)(1404)(25.0445)(0433)(1426),(),(),(),;()(−=−==−==−==−==−=−==badEVAgoodEVAnobleEVAneatEVAbadgoodMPLgoodproperMPLbadproperMPLbadgoodproperTRIproperEVAOther scalesn Define POT as TRI(w;strong,weak)n Define ACT as TRI(w;active,passive)n EVA, POT, ACT are well-defined forexactly the same set of 5410adjectives.EVA*, POT*, ACT*n Define EVA*(w) to be EVA(w) if apath exists between w and “good”,and 0 if it doesn’tn This gives us a well-defined functionfor all wn Do the same thing to get POT* andACT*Applicationn We can now take the sum of EVA*,POT*, ACT* for all words in a text toget an idea of the good-bad, strong-weak, active-passive values for thetext as a wholeAccuracyn No corpus existed that had already beenrated for these values, so accuracy couldnot be tested on a large scalen Tests on small numbers of Internetdiscussions show correspondencebetween results of this method and actualvalue of texts, but questionable accuracyfor short textsn Works better for long textsAccuracy problemsn With longer texts, false positives andfalse negatives cancel each other out;doesn’t help for shorter textsn Longer texts yield scores of highermagnitude, in general – need tonormalize scoresn Apparent bias to positive words (positiveopinions more extensively elaborated,affecting a text’s score more thannegative opinions)Author’s closing notesn Authors of texts on Internetdiscussion sites must be less subtleabout good/badn Little NLP research addressessubjective aspects; this paperhelpsfill the
View Full Document