landis

Words with Attitude: 

Words with Attitude Jaap Kamps Maarten Marx

Paper’s Goal: 

Paper’s Goal Judge the emotive or affective meaning of a text Use WordNet to determine values of words with Osgood’s semantic differential technique

Osgood’s Semantic Differential Technique: 

Osgood’s Semantic Differential Technique Judge words, phrases, texts by asking subjects to rate them on scales of bipolar adjectives A subject might be asked to rate “proper” on scales like optimistic-pessimistic, serious-humorous, and active-passive. It turns out that good-bad, strong-weak, and active-passive values account for most variance in judgment

Using WordNet with Osgood’s theory: 

Using WordNet with Osgood’s theory Authors want to get values for words from WordNet They define MPL(w1,w2) as the minimal path length between w1 and w2, using only same-synset relations Allowing more than just same-synset damages metric

MPL Examples: 

MPL Examples MPL(good, proper) = 2 (good,right,proper) MPL(good, neat) = 3 MPL(good, noble) = 4 Can we use this to rate “proper”, “neat”, and “noble” on a good-bad scale?

MPL: 

MPL MPL(good, bad) = 4 If we just look at MPLs, “noble” is as good as “bad” We need to do something a bit more complicated

TRI: 

TRI To determine the good-bad (“evaluative”) value of wi, examine TRI(wi;good,bad) Define EVA(w) = TRI(w;good,bad)

EVA results: 

EVA results There are 5410 adjectives linked to “good” or “bad”. Average value of EVA for these 5410 words is –0.0089

Other scales: 

Other scales Define POT as TRI(w;strong,weak) Define ACT as TRI(w;active,passive) EVA, POT, ACT are well-defined for exactly the same set of 5410 adjectives.

EVA*, POT*, ACT*: 

EVA*, POT*, ACT* Define EVA*(w) to be EVA(w) if a path exists between w and “good”, and 0 if it doesn’t This gives us a well-defined function for all w Do the same thing to get POT* and ACT*

Application: 

Application We can now take the sum of EVA*, POT*, ACT* for all words in a text to get an idea of the good-bad, strong-weak, active-passive values for the text as a whole

Accuracy: 

Accuracy No corpus existed that had already been rated for these values, so accuracy could not be tested on a large scale Tests on small numbers of Internet discussions show correspondence between results of this method and actual value of texts, but questionable accuracy for short texts Works better for long texts

Accuracy problems: 

Accuracy problems With longer texts, false positives and false negatives cancel each other out; doesn’t help for shorter texts Longer texts yield scores of higher magnitude, in general – need to normalize scores Apparent bias to positive words (positive opinions more extensively elaborated, affecting a text’s score more than negative opinions)

Author’s closing notes: 

Author’s closing notes Authors of texts on Internet discussion sites must be less subtle about good/bad Little NLP research addresses subjective aspects; this paperhelps fill the gap