A Sentiment Analysis Method to Better Utilize User Profile and Product Information Capstone Project Presentation Mingyu MA (Derek) derek.ma@connect.polyu.hk BSc (Hons) Computing, 14110562D Su Supervisor: Prof. Qin n LU Co Co-ex exam aminer er: Dr. Ajay ay Kumar ar PATHAK nd As 2 nd Assessor: Dr. Richard LUI
Contents Introduction Related Work Model Design Evaluation and Analysis Conclusion and Future Work 1 / 34
Businesses would like to know users’ opinions Introduction Users can be benefited from others’ opinions users’ opinions to improve services post product reviews reviews data post video reviews ratings and opinions of other customers 2 / 34
Introduction methods of detecting, analyzing, and Sentiment evaluating people’s state of mind towards events, issues, or any other Analysis interest. (Yadollahi et al., 2017) 3 / 34
Introduction Background Info Is Available user profile provide user’s history domain user’s preferences knowledge ... user product information more facts product property and reviews: possibilities other user’s opinions main document … product 4 / 34
Introduction Background Information Is Not Unified • User’s perspective • Mean/lenient user • Product’s perspective + • Type, category • Different background information influences the results in different perspectives 5 / 34
Introduction Objectives A new sentiment analysis model - utilize user and product information - reflect impacts from user profile and product information separately 6 / 34
Related Work Machine-Learning-based Sentiment Analysis (Tang et al., 2014), (Kim, 2014) (Yang et al., 2016) (Wang and Manning, 2012) (Long et al., 2017) NN as classifier for text Linear model or Focus more on important text classification kernel methods on and add more associate data RNN , LSTM lexical features like eye-tracking data Neural-network- Traditional Attention based Approaches Way 7 / 34
Related Work User and Product Info in Sentiment Analysis • Memory network xt Utilizing User Profile and (Tang, Qin and Liu, 2015; Dou, 2017) a • RNN + external memory Product Information in • Use external info as attention Sentiment Analysis (Chen et al., 2016) • State-of-the-art • All consider user profile and product information as single > representation 8 / 34
Model Design JUPMN Joint User and Product Memory Network 9 / 34
Model Design Model Overview Sentiment Prediction Input & Output Joint Mechanism • Input • Document d PMN UMN • A writer u ^ ^ P(d) U(d) • A target p • Output (numeric vector) Document d • Discrete sentiment Hierarchical LSTM with Attention prediction Document d (text) D 10 / 34
Model Design Model Overview Sentiment Prediction Structure Joint Mechanism PMN UMN Part 2: Memory Networks ^ ^ P(d) U(d) (numeric vector) Document d Part 1: Document Hierarchical LSTM with Attention Embedding Document d (text) D 11 / 34
Model Design > Part 1: Document Embedding Hierarchical Long Short-Term Memory Network Hierarchical LSTM with Attention 12 / 34
Model Design > Part 1: Document Embedding Hierarchical Long Short-Term Memory Network • Word-sentence- document level convention (Chen et al., 2016) • Add attention in LSTM Hierarchical LSTM with Attention layers • With user and product attention • With eye-tracking cognition attention 13 / 34
Model Design Part 2: Memory Networks Sentiment Prediction Joint Mechanism PMN UMN Part 2: Memory Networks ^ ^ P(d) U(d) (numeric vector) Document d Part 1: Document Hierarchical LSTM with Attention embedding Document d (text) D 14 / 34
Model Design > Part 2: Memory Networks Sentiment Softmax Prediction w U w P W P d 3 d 3 W U U P a 3 a 3 Attention Layer Attention Layer Attention Layer 3 3 3 d 2 d 2 P U a 2 a 2 Attention Layer Attention Layer M M ... ... 2 2 N d 1 N d 1 ^ ^ U(d) P(d) a 1 a 1 Attention Layer Attention Layer 1 1 (embedded by (embedded by hierarchical LSTM) hierarchical LSTM) d 0 d 0 Document d (embedded by hierarchical LSTM) 15 / 34
Model Design > Part 2: Memory Networks Structure of Attention Layers • Attention weight Output a k External ATT W Memory Attention Layer • Output of attention layer 3 Input d k-1 Attention Layer k 16 / 34
Evaluation and Analysis Benchmark Datasets and Performance Metrics Three Benchmark Datasets • IMDB • Diao et al., 2014 • Yelp 13, Yelp 14 • Tang et al., 2015a 17 / 34
Evaluation and Analysis Benchmark Datasets and Performance Metrics Three Benchmark Datasets 18 / 34
Evaluation and Analysis Benchmark Datasets and Performance Metrics Performance Metrics 19 / 34
Evaluation and Analysis JUPMN and Comparison Models Experimental Results Group 1: simple methods based on language features Group 2: models using machine learning Group 3: models with user profile and product information in machine learning 20 / 34
Evaluation and Analysis JUPMN and Comparison Models Experimental Results Findings • JUPMN outperforms the state-of-the-art model • Generally Group 2 performs better than Group 1, Group 3 performs better than Group 2 • Exceptions exist • TextFeature • LSTM+CBA 21 / 34
Evaluation and Analysis JUPMN with Different Configurations Sentiment Four aspects of Prediction configurations Joint Weights Joint Mechanism PMN UMN Memory Size ^ ^ P(d) U(d) (numeric vector) Document d Number of Hops Importance of User vs Product Memory Network Hierarchical LSTM with Attention Document d (text) D 22 / 34
Evaluation and Analysis > JUPMN with Different Configurations Importance of User vs Product Memory Network Experimental Results Observations • User profile influences sentiments of movie reviews more • Product information influences sentiments of restaurants reviews more • JUPMN-U • With only User Memory Network • JUPMN-P • With only Product Memory Network 23 / 34
Evaluation and Analysis > JUPMN with Different Configurations Importance of User vs Product Memory Network Investigating by Checking Joint Weights Average joint weight for three datasets • Verified the hypothesis Joint weights for three datasets 24 / 34
Evaluation and Analysis > JUPMN with Different Configurations Importance of User vs Product Memory Network For IMDB dataset Investigating by Word Frequency Plotting 10 users give average 10 movies have average highest/lowest rating score highest/lowest rating score 25 / 34
Evaluation and Analysis > JUPMN with Different Configurations Importance of User vs Product Memory Network For IMDB dataset Investigating by Word Frequency Plotting For movies reviews • Users’ words are very different • Products’ words are very objective 26 / 34
Evaluation and Analysis > JUPMN with Different Configurations Importance of User vs Product Memory Network For Yelp dataset Investigating by Word Frequency Plotting For restaurants reviews • Users’ words are not distinguishable • Products’ words shows the sentiments 27 / 34
Evaluation and Analysis > JUPMN with Different Configurations Number of Hops (Computational Layers) Experimental Results Observations • Smaller hop works better • Possible explanations • Data distortion • Over-fitting 28 / 34
Evaluation and Analysis > JUPMN with Different Configurations Memory Size • Larger memory helps • When memory size reaches 75, no longer improve • There is not enough documents 29 / 34
Evaluation and Analysis > JUPMN with Different Configurations Joint Weights JUPMN (not weighted) JUPMN • Weighted version works better • Weight help to balance the influences of UMN and PMN 30 / 34
Evaluation and Analysis Case Study • What is this user’s opinion? • Cite negative reviews to praise • JUPMN can learn the features of this user • This user is a science fiction movie • JUPMN can learn the features of this movie (product) • This movie is relative great according to other reviews 31 / 34
Conclusion and Future Work Conclusion • Proposed JUPMN • JUPMN outperforms the state-of-the-art sentiment analysis model • Analysis on different configuration is employed • Research paper Yunfei Long*, Mingyu Ma*, Rong Xiang, Qin Lu, Chu-Ren Huang. Fusing User Memory and Product Memory for Sentiment Classification. (*: Equal contribution) Future Work • More knowledge in memory network • Application of JUPMN in more languages datasets 32 / 34
References 33 / 34
References 34 / 34
Thanks! A Sentiment Analysis Method To Better Utilize User Profile and Product Information Mi Mingyu MA MA (Dere rek) supervised by Pr Prof. f. Qin LU
Recommend
More recommend