160 likes | 573 Vues
Ensemble Learning for Sentiment Analysis. Robert Christensen, Haibo Ding, Mengyang Wang, Fei Luo Dec 10 2013. Sentiment Analysis. Sentiment Analysis Research area of NLP Analyze people’s opinion, sentiments, emotions etc. One basic task is to classify the polarity of a given text.
E N D
Ensemble Learning for Sentiment Analysis Robert Christensen, Haibo Ding, Mengyang Wang, FeiLuo Dec 10 2013
Sentiment Analysis • Sentiment Analysis • Research area of NLP • Analyze people’s opinion, sentiments, emotions etc. • One basic task is to classify the polarity of a given text. • Why? • business • People’s opinions influence out behaviors ( Choice we make, and what we will buy)
The Problem • Sentiment polarity classification • Classify the given text as positive or negative • In our experiments, we try to classify the movie reviews • For example: • Positive: “A deep and meaningful film” • Negative: “It’s like watching a nightmare made flesh.”
Why Ensemble Learning? • INTUITION: Combining predictions of multiple classifiers(an ensemble)is more accurate than a single classifier. • Justification: • easy to find quite good “rules of thumb” however hard to find single highly accurate prediction rule. • If the training set is small and the hypothesis space is large then there may be many equally accurate classifiers. • Exhaustive global search in the hypothesis space is expensive so we can combine the predictions of several locally accurate classifiers.
Ensemble with different algorithms Train Set Classify algorithms CA 1 CA 2 CA n ……………. Model 1 Model2 Model n ……………. Test Set Result 1 Result2 Resultn ……………. Ensemble Final results
Ensemble on sampled data Train Set Random sampling sampletrain1 sampletrain1 sampletrain1 ……………. Same Classify algorithms Model 1 Model2 Model n ……………. Test Set Result 1 Result2 Resultn ……………. Final results Ensemble
Experiments and Results • Data • Stanford sentiment analysis data (download on web) • Train set size: 6920 • Test set size: 1821 • Development set size: 872 • Classification algorithms • Naïve Bayes, SVM, MaxEnt, Logistic Regresion • Measurement • Classification Accuracy
Baseline Results • Features: • Bag of words Table 1. Baseline Results
Results of Ensemble Method 1 • Ensemble with different classification algorithms • Features: • Probabilities from stage-1 classifiers (NB, SVM, MaxEnt) Table 2. Performance of our ensemble system using different classifiers.
Results of Ensemble method 2 • Ensemble by sampling train data • Features: • Probabilities from classifiers trained on samples (using MaxEnt classifier) • Stage-1 classifiers’ accuracy 76.83% (max), 70.47%(avg) Table 3. Performance of our ensemble system using sampled train data.
Results of Ensemble method 2 • Ensemble by sampling train data • Features: • Probabilities from classifiers trained on samples (using NB classifier) • Stage-1 classifiers’ accuracy 78.03% (max), 71.56%(avg) Table 4. Performance of our ensemble system using sampled train data.
Conclusion • Ensemble method slightly improves accuracy of multiple stage-1 classifiers. • Ensemble method provides a flexible method to effectively combine multiple trained classifiers. • Future work includes studying the effectiveness of various stage-2 classifier features.
References • http://classes.engr.oregonstate.edu/eecs/fall2011/cs434/notes/ensemble.pdf • Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank Richard Socher et al. EMNLP-2013 • Opinion mining and sentiment analysis. B. Pang and L. Lee. 2008. Foundations and Trends in Information Retrieval