Post on 15-Jul-2015
{SentiWordNet
Andrea Esuli* and Fabrizio Sebastiani, SentiWordNet: A Publicly Available Lexical Resource for Opinion Mining
1. text의주관성, 객관성판단
-> SO polarity (Pang and Lee, 2004; Yu and Hatzivassiloglou,
2003)
2. 주관성을지닌 text의긍정, 부정판단
-> PN polarity (Pang and Lee, 2004; Turney, 2002)
3. 얼마나긍정/부정인지판단
예) 조금긍정, 약간긍정, 아주긍정
-> strength of text PN polarity(Pang and Lee, 2005; Wilson et al., 2004)
Polarity classificationBo Pang and Lillian Lee, A Sentimental Education: Sentiment Analysis Using Subjectivity Summarization Based on Minimum Cuts
WordNet
synset
- 영어의의미어휘목록
- synset (유의어집단)으로분류하여단어집과유의어,반의어사전의배합을만듬.
- 심리학교수인조지A. 밀러가지도하는프린스턴대학의인지과학연구소에의해만들어졌고, 유지되고있음.
synset
Score : 0.0 ~ 1.0,각 synset의총합은 1
SentiWordNet
Lp, Ln, Lo -> vectorial representation -> label Ci
-> 2개의 classifier 생성
1. Positive / not Positive 로분류하는 classifier
2. Negative / not Negative 로분류하는 classifier
Classifier
“Supervised learner”
Positive∩
Lp, Ln, Lo -> vectorial representation -> label Ci
-> 2개의 classifier 생성
1. Positive / not Positive 로분류하는 classifier
2. Negative / not Negative 로분류하는 classifier
Classifier
“Supervised learner”
Negative∩
Lp, Ln, Lo -> vectorial representation -> label Ci
-> 2개의 classifier 생성
1. Positive / not Positive 로분류하는 classifier
2. Negative / not Negative 로분류하는 classifier
Classifier
“Supervised learner”
Objective
∩ ∩
∪
Precision = Tp / (Tp + Fp)
: True라고예측한것중에서실제로 true인것의비율
Recall = Tp / (Tp + Fn)
: 실제로 true인것중에내가얼마나맞췄는지
(Tp : true라고예측했는데실제로 true,
Fp : true라고예측했는데실제로 false,
Fn : false라고예측했는데실제로 true,
Tn : false라고예측했는데실제로 false)
K를정하려면?
K
Precision
Recall
SmallTraining set
noise
(Andrea Esuli1 and Fabrizio Sebastiani2, Determining Term Subjectivity and Term Orientation for Opinion Mining)
Roccino (Andrew McCallum’s Bow package http://www-2.cs.cmu.edu/~mccallum/bow/)
SVM (6.01 of Thorsten Joachims’
SVMlight - http://svmlight.joachims.org/).
K = 0, 2, 4, 6 -> 8가지의 ternary classifier
-> 1로정규화
K를정하려면?