360-MAM-Affect: Sentiment Analysis with the Google Prediction API and EmoSenticNet Eleanor Mulholland, Paul Mc Kevitt & Tom Lunney School of Creative Arts & Technologies Ulster University Derry/Londonderry, Northern Ireland John Farren & Judy Wilson 360 Production Ltd. Suite 2.2, The Innovation Centre
Introduction Aims & Objectives Background & Literature Review Design of 360-MAM-Select Software Analysis for 360-MAM-Select SentimentClassifer EmoSenticNetClassifer Relation to Other Work Conclusion & Future Work Acknowledgements
Aims & Objectives Implement online recommender system (360-MAM-Select) Employ sentiment analysis & gamification to achieve higher quality video recommendations Sentiment Analysis facilitates video recommendations Gamification motivates engagement with video content
Recommender Systems Rank products against others for comparison Faridani (2011) trained a recommender model with ratings from OpinionSpace dataset Hanser et al. (2010) implemented NewsViz giving numerical emotion ratings to words Tkalčič et al. (2011) proposed a unifying framework for emotion detection
Unifiying Framework (Tkalčič et al. 2011)
Sentiment Analysis Recognising negative, positive & neutral opinions (Wilson et al. 2005) Methods of opinion collecting Variety of data on the Internet in many different forms Online information is not static (Khan et al. 2009)
Gamification Game mechanics & game design techniques used to enhance non-game scenarios (Deterding et al. 2011) Common gamification methods (Lee et al. 2013) Gamification is popular for monitoring & analysing online communities (Bista et al. 2012) Gamification has improved learning & information retention in education & staff training (Landers & Callan, 2011)
Design of 360-MAM-Select Modules for sentiment analysis & emotion modelling (360-MAM-Affect) & gamification (360-MAM-Gamify) How user responds emotionally to media content 360-MAM-Affect harvests YouTube & Head Squeeze (Head Squeeze, 2014) comments on video content Identify the overall reception
Architecture of 360-MAM-Select
Software Analysis for 360-MAM-Select Google Prediction API (Google, 2015) used to create 2 sentiment analysis prediction models: (1) SentimentClassifer & (2) EmoSenticNetClassifer SentimentClassifer model trained with Google Prediction API with classification accuracy of 61% EmoSenticNetClassifer model trained with Google Prediction API, using EmoSenticNet (Gelbukh, 2015) with classification accuracy of 59%
Testing SentimentClassifier
EmoSenticNet Example Values
EmoSenticNet Concepts Associated With One Emotion
Frequencies (%) of 2 Emotions Per Concept
Combination of 3 Emotions Per Concept
EmoSenticNet Model Test EmoSenticNetClassifer model returned 7 number labels (Anger, Disgust, Joy, Sad, Surprise, Fear & Neutral) with a classification accuracy of 59% 7 concepts failed to return highest rating in expected emotion label Joy had highest frequency of entries in training data
EmoSenticNetClassifer Prediction Results
Relation to Other Work Recommender systems (Ricci et al. 2011) ability to personalise experiences (Linden et al. 2003) to provide high quality recommendations (Bartłomiej 2005) Emotion has been identified as important factor in improving recommender systems (Tkalčič et al. 2011) 360-MAM-Select will advance recommender systems by providing improved user experience
Conclusion & Future Work Architecture of 360-MAM-Select Testing SentimentClassifer (Google Prediction API) & EmoSenticNetClassifer (Google Prediction API + EmoSenticNet) Investigate how sentiment analysis & gamification can be utilised Improve user participation & video retrieval Implementation & testing of 360-MAM-Select
Acknowledgements Dr. Brian Bridges Dr. Kevin Curran Dr. Lisa Fitzpatrick 360 Production Ltd. & Alleycats TV