Publication: Emocnn: Encoding emotional expression from text to word vector and classifying emotions—a case study in thai social network conversation
Issued Date
2021-01-01
Resource Type
ISSN
01258281
Other identifier(s)
2-s2.0-85112480821
Rights
Mahidol University
Rights Holder(s)
SCOPUS
Bibliographic Citation
Engineering Journal. Vol.25, No.7 (2021), 73-82
Suggested Citation
Konlakorn Wongpatikaseree, Yongyos Kaewpitakkun, Sumeth Yuenyong, Siriwon Matsuo, Panida Yomaboot Emocnn: Encoding emotional expression from text to word vector and classifying emotions—a case study in thai social network conversation. Engineering Journal. Vol.25, No.7 (2021), 73-82. doi:10.4186/ej.2021.25.7.73 Retrieved from: https://repository.li.mahidol.ac.th/handle/20.500.14594/76979
Research Projects
Organizational Units
Authors
Journal Issue
Thesis
Title
Emocnn: Encoding emotional expression from text to word vector and classifying emotions—a case study in thai social network conversation
Abstract
We present EmoCNN, a collection of specially-trained word embedding layer and convolutional neural network model for the classification of conversational texts into 4 types of emotion. This model is part of a chatbot for depression evaluation. The difficulty in classifying emotion from conversational text is that most word embeddings are trained with emotionally-neutral corpus such as Wikipedia or news articles, where emotional words do not appear very often or at all, and the language style is formal writing. We trained a new word embedding based on the word2vec architecture in an unsupervised manner and then fine-tuned it on soft-labelled data. The data was obtained from mining Twitter using emotion keywords. We show that this emotion word embedding can differentiate between words which have the same polarity and words which have opposite polarity, as well as find similar words with the same polarity, while the standard word embedding cannot. We then used this new embedding as the first layer of EmoCNN that classifies conversational text into the 4 emotions. EmoCNN achieved macro-averaged f1-score of 0.76 over the test set. We compared EmoCNN against three different models: a shallow fully-connected neural network, fine-tuning RoBERTa, and ULMFit. These got the best macro-averaged f1-score of 0.5556, 0.6402 and 0.7386 respectively.