![? Tokenization
? Feature Extraction: n-grams, semantics, syntactic, etc.
? Classification using different classifiers
? Na?ve Bayes
? MaxEnt
? SVM
? Drawback
? Feature Sparsity
S1: I really like this movie
[...0 0 1 1 1 1 1 0 0 ... ]
8
S1: This phone has a good keypad
S2: He will move and leave her for good](https://image.slidesharecdn.com/random-160118160118/85/Continuous-Sentiment-Intensity-Prediction-based-on-Deep-Learning-8-320.jpg)
![? Using clustering algorithm to aggregate short text to form big clusters, in which
each cluster has the same topic and the same sentiment polarity, to reduce the
sparsity of short text representation and keep interpretation.
S1: it works perfectly! Love this product
S2: very pleased! Super easy to, I love it
S3: I recommend it
it works perfectly love this product very pleased super easy to I recommend
S1: [1 1 1 1 1 1 0 0 0 0 0 0 0]
S2: [0 0 0 1 0 0 1 1 1 1 1 1 0]
S3: [1 0 0 0 0 0 0 0 0 0 0 1 1]
S1+S2+S3: [...0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0...]
9](https://image.slidesharecdn.com/random-160118160118/85/Continuous-Sentiment-Intensity-Prediction-based-on-Deep-Learning-9-320.jpg)
i i i mc f w S b? ?? ? ?
? Dimension Reduced
? Reduce the parameters of the model
? Parameter sharing
f: Activation function, Relu, tanh, sigmoid, ¡
? ? = max(0, ?)
19](https://image.slidesharecdn.com/random-160118160118/85/Continuous-Sentiment-Intensity-Prediction-based-on-Deep-Learning-19-320.jpg)
![[Revised] Intro to CNN](https://cdn.slidesharecdn.com/ss_thumbnails/googletechsprinttalkcnnintro-200730141604-thumbnail.jpg?width=560&fit=bounds)
More Related Content
What's hot (14)
Similar to Continuous Sentiment Intensity Prediction based on Deep Learning (20)
Recently uploaded (20)
Continuous Sentiment Intensity Prediction based on Deep Learning
- 1. Yunchao He (ºÎÔƳ¬) 2015.9.15 @ Yuan Ze University
- 2. ? ¡°unbelievably disappointing¡± ? ¡°Full of zany characters and richly applied satire, and some great plot twists¡± ? ¡°this is the greatest screwball comedy ever filmed¡± ? ¡°It was pathetic. The worst part about it was the boxing scenes.¡± ? Sentiment Analysis ? Using NLP, statistics, or machine learning methods to extract, identify, or otherwise characterize the sentiment content of a text unit ? Sometimes called opinion mining, although the emphasis in this case is on extraction ? Other names: Opinion extraction¡¢Sentiment mining¡¢Subjectivity analysis 2
- 3. 3
- 4. ? Movie: is this review positive or negative? ? Products: what do people think about the new iPhone? ? Public sentiment: how is consumer confidence? Is despair increasing? ? Politics: what do people think about this candidate or issue? ? Prediction: predict election outcomes or market trends from sentiment 4
- 5. ? Short text classification based on Semantic clustering ? Sentiment intensity prediction using CNN ? Transfer Learning* * Future works 5
- 6. ? People express opinions in complex ways ? In opinion texts, lexical content alone can be misleading ? Intra-textual and sub-sentential reversals, negation, topic change common ? Rhetorical devices such as sarcasm, irony, implication, etc. 6
- 8. ? Tokenization ? Feature Extraction: n-grams, semantics, syntactic, etc. ? Classification using different classifiers ? Na?ve Bayes ? MaxEnt ? SVM ? Drawback ? Feature Sparsity S1: I really like this movie [...0 0 1 1 1 1 1 0 0 ... ] 8 S1: This phone has a good keypad S2: He will move and leave her for good
- 9. ? Using clustering algorithm to aggregate short text to form big clusters, in which each cluster has the same topic and the same sentiment polarity, to reduce the sparsity of short text representation and keep interpretation. S1: it works perfectly! Love this product S2: very pleased! Super easy to, I love it S3: I recommend it it works perfectly love this product very pleased super easy to I recommend S1: [1 1 1 1 1 1 0 0 0 0 0 0 0] S2: [0 0 0 1 0 0 1 1 1 1 1 1 0] S3: [1 0 0 0 0 0 0 0 0 0 0 1 1] S1+S2+S3: [...0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0...] 9
- 10. ? Training data labeled with positive and negative polarity ? K-means clustering algorithm is used to cluster positive and negative text separately. ? K-means, KNN, LDA¡ works perfectly! Love this product completely useless, return policy very pleased! Super easy to, I am pleased was very poor, it has failed highly recommend it, high recommended! it totally unacceptable, is so bad works perfectly! Love this product very pleased! Super easy to, I am pleased highly recommend it, high recommended! completely useless, return policy was very poor, it has failed it totally unacceptable, is so bad Topical clusters 10
- 11. Classifier: Multinomial Naive Bayes Probabilistic classifier: get the probability of label given a clustered text , 1 arg max ( | ) arg max ( ) ( | ) Ci i s S i j s S j N s P s C P s P C s ? ? ? ? ? ? ? $ ( ) sN P s N ? , , ( , ) 1 ( | ) ( | ) | | i j i j x V N C s P C s N x s V ? ? ? ?? Bayes¡¯ theory Independent assumption 11
- 12. ? Given an unlabeled text , we use Euclidean distance to find the most similar positive cluster , and the most similar negative cluster ? The sentiment of , is estimated according to the probabilistic change of the two clusters when merging with . (vs. KNN) ? This merging operation is called two-stage-merging method, as each unlabeled text will be merged two times. 0, | ( ) ( ) | | ( ) ( ) | ( ) 1, . m m n n j P NC P C P NC P C f x otherwise ? ? ? ? ? ? ? ? ? ? ? mC ? jx nC ? jx jx 12
- 13. ? Dataset: Stanford Twitter Sentiment Corpus (STS) ? Baseline: bag-of-unigrams and bigrams without clustering ? Evaluation Metrics: accuracy, precision, recall ? The average precision and accuracy is 1.7% and 1.3% higher than the baseline method. Methods Accuracy Precision Recall Our Method 0.816 0.82 0.813 Bigrams 0.805 0.807 0.802 13
- 15. ? Continuous sentiment intensity provides fine-grained representation of sentiment. ? Representing sentiment as Valence-arousal can easily convert to discrete categories. ¡°unbelievably disappointing¡± Model V: -0.5 A: 0.3 15
- 16. ? Lexicon based Method. To find the relationship between word-level and sentence-level sentiment values. Word-level information comes from sentiment lexicon, e.g. ANEW. ? Paltoglou 2013: Weighted Arithmetic Mean¡¢Weighted Geometric Mean ? Malandrakis 2013: linear regression Paltoglou, G., Theunis, M., Kappas, A., & Thelwall, M. (2013). Predicting emotional responses to long informal text. Affective Computing, IEEE Transactions on, 4(1), 106-115. Malandrakis, N., Potamianos, A., Iosif, E., & Narayanan, S. (2013). Distributional semantic models for affective text analysis. Audio, Speech, and Language Processing, IEEE Transactions on, 21(11), 2379-2392. 16
- 17. ? To find the relationship between words and sentence-level sentiment. CNN Method Lexicon-based Methods Word Dense vector VA value Relationship Auto learned Manually specified Training data Many Few or None Word Order Considered* Not Considered Interpretation Black Box Easy 17
- 18. ? To find the relationship between words and sentence-level sentiment. ? Sentence Matrix -> Convolution Operator -> Max Pooling -> Regression ? Word Representation: dense vector, distributed representation ÎÒÃÇµÄ ÐÄ ²»Ïñ Ã÷¾µ ²»¿ÉÒÔ ÃÀ³ó Éƶñ È«²¿ °üÈÝ boat ship vessel good happy Beijing Shanghai glad Semantic information of word is encoded in the dense vector. 18
- 19. ? Sentence Matrix -> Convolution Operator -> Max Pooling -> Regression ÎÒÃÇµÄ ÐÄ ²»Ïñ Ã÷¾µ ²»¿ÉÒÔ ÃÀ³ó Éƶñ È«²¿ °üÈÝSentence Matrix [ : 1,:]( )i i i mc f w S b? ?? ? ? ? Dimension Reduced ? Reduce the parameters of the model ? Parameter sharing f: Activation function, Relu, tanh, sigmoid, ¡ ? ? = max(0, ?) 19
- 20. ? Sentence Matrix -> Convolution Operator -> Max Pooling -> Regression ? Aggregate the information and capture the most important features ÎÒÃÇµÄ ÐÄ ²»Ïñ Ã÷¾µ ²»¿ÉÒÔ ÃÀ³ó Éƶñ È«²¿ °üÈÝ 3 6 79 7 54 79 9 Max pool with 5¡Á1 filters and stride 1 20
- 21. ? Sentence Matrix -> Convolution Operator -> Max Pooling -> Regression ÎÒÃÇµÄ ÐÄ ²»Ïñ Ã÷¾µ ²»¿ÉÒÔ ÃÀ³ó Éƶñ È«²¿ °üÈÝ x1 x2 xn linear ? ??, ? = ???? = ?? Objective function: mean squared error (MSE) 21
- 22. ? Learning Algorithm: stochastic gradient descent (SGD) ? Learning the parameters of the model with labeled data ? Word vectors ? Convolution filters weights ? Linear regression weights ? Labeled data ? Chinese: CVAT dataset ? English: VADER dataset Dataset size #word L Dims CVAT 720 21094 192.1 V+A Tweets 4000 15284 13.62 V Movie 10605 29864 18.86 V Amazon 3708 8555 17.3 V NYT 5190 20941 17.48 V 22
- 23. ? All the dataset is separated into training set, validation set and test set for model training, hyper-parameters selection and model evaluation. ? Evaluation Metrics ? MSE, Mean Square Error ? MAE, Mean Absolute Error ? Pearson¡¯s correlation coefficient r %2 1 1 ( ) n i i i MSE y y n ? ? ?? % 1 1 | | n i i i MAE y y n ? ? ?? % % % % 1 1 1 2 2 1 1 1 1 1 1 ( )( ) 1 1 ( ) ( ) n n n i ii j i j j n n n n i ii j i j i j y y y y n n r y y y y n n ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? 23
- 24. Methods CNN wGW RMAR LCEL RMV Metrics MSE MAE r MSE MAE r MSE MAE r MSE MAE r MSE MAE r valence ratings prediction CVAT 1.17 0.88 0.73 2.30 1.23 0.62 1.89 1.14 0.63 1.81 0.95 0.66 1.49 0.98 0.72 Tweets 1.00 0.76 0.79 2.54 1.25 0.65 1.30 0.89 0.69 1.25 0.85 0.75 1.18 0.86 0.74 Movie 2.14 1.18 0.67 6.46 2.02 0.17 3.54 1.73 0.16 2.54 1.36 0.42 2.25 1.26 0.62 Amazon 1.50 0.95 0.67 3.75 1.51 0.35 2.66 1.38 0.27 1.45 1.14 0.45 2.20 1.19 0.56 NYT 0.84 0.72 0.36 3.47 1.54 0.28 0.79 0.71 0.26 0.83 0.75 0.37 0.61 0.63 0.60 arousal ratings prediction CVAT 0.98 0.81 0.64 1.34 0.94 0.31 1.20 0.89 0.35 1.07 0.91 0.62 0.98 0.79 0.53 CNN method improved the VA prediction experiment performance compared with the lexicon-based and RMV method. Baseline method: ? wGW, Weighted geometric mean method ? RMAR, Regression on mean affective ratings ? LCEL, linear combination using expanded lexicon ? RMV, regression on mean vectors method 24
- 25. ? Using Transfer Learning Techniques to improve VA prediction performance. ? Motivation: There are numerous dataset for sentiment classification but only a few dataset for VA prediction. The sentiment polarity maybe useful for VA prediction. ? Method: Pre-training the classification-CNN model, and then use the parameters of the pre- trained networks as the initial value of VA prediction-CNN model, keep training on VA corpus. 25