Premium
Three predictions are better than one: Sentence multi‐emotion analysis from different perspectives
Author(s) -
Wu Yug,
Kita Kenji,
Matsumoto Kazuyuki
Publication year - 2014
Publication title -
ieej transactions on electrical and electronic engineering
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.254
H-Index - 30
eISSN - 1931-4981
pISSN - 1931-4973
DOI - 10.1002/tee.22020
Subject(s) - sentence , probabilistic logic , context (archaeology) , sentiment analysis , perspective (graphical) , word (group theory) , focus (optics) , computer science , natural language processing , task (project management) , baseline (sea) , macro , artificial intelligence , cognitive psychology , psychology , mathematics , paleontology , oceanography , physics , geometry , management , optics , economics , biology , programming language , geology
Emotion prediction has been a core task in affective computing, which aims at finding the thorough human mental states by analyzing people's activities. In this paper, we focus on predicting emotions in the public online blogs from different people, by extracting as many reasonable emotions for each blog sentence as possible. Concretely, we consider three different perspectives for analyzing the multiple emotions in a sentence: (i) predict sentence emotions by examining the emotion related topics in a global sense; (ii) predict the sentence emotions from the context‐sensitive word emotions; and (iii) predict sentence emotions by considering the emotional significance in the local bag of words. We build different probabilistic models from each perspective, to separately generate the sentence emotion probabilities. We then integrate these probabilistic models to jointly predict the emotion probabilities. Because the component models are based on different emotional assumptions with distinct features, the integrated predictions should predict emotions from more general perspectives and therefore yield better results. In the experiment, we employ different evaluation criteria to compare the multi‐emotion predictions from the single and integrated models. Compared to the results in the baseline model, our bi‐integrated models achieve 8.69% higher Micro F1 and 7.78% higher Macro F1 scores, on average. Moreover, our tri‐integrated model acquires 10.00% higher Micro F1 and 9.19% higher Macro F1 scores than the baseline results, which proves our assumption, and suggests interesting features in the different emotion perspectives. © 2014 Institute of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc.