A Sentiment Classifier of Allegorical Sayings (Xie-Hou-Yu) Based on Deep Neural Network

Autor: CHEN, CHIH-CHENG, 陳志正
Rok vydání: 2018
Druh dokumentu: 學位論文 ; thesis
Popis: 106
Allegorical Sayings (Xie-Hou-Yu) are treasures in Chinese literature. A speaker uses Xie-Hou-Yu to wrap his/her original meaning, of which listeners might not be able to understand the true meaning and the sentiment of it. This study is therefore to provide a classifier that can help learners easily judge whether a Xie-Hou-Yu is of positive or negative sentiment. First, this study uses 1,202 Taiwanese Xie-Hou-Yu’s adopted from Huang Zhong-Jie's research to test the feasibility of deep learning in classifying the sentimental polarities of the Taiwanese Xie-Hou-Yu. The polarities of the Xie-Hou-Yu include commendatory (labeled 1), neutral (label 2), and derogatory (labeled 3). We also explored the mapping/translation between "word face" (puzzles) and "word meaning" (answers) by using sequence-to-sequence alignment. It is our idea that the deep learning can help to get higher level of representation of Xie-Hou-Yu, and thus will help to build a more robust sentiment classifier. This study also collected data of Chinese Xie-Hou-Yu and further explored the differences between Chinese and Taiwanese Xie-Hou-Yu in the following aspects: 1) The effects of attention mechanism on the mapping between "word face" and "word meaning", and further on the classification accuracy. 2) The effects of three different parts ("word face" only, "word meaning" only, and the whole Xie-Hou-Yu) on the classification accuracy. 3) the comparison between the proposed model with Multi-nominal Bayes and Support Vector Machine. Because the distribution of three categories are highly unbalanced, most classifier cannot perform well. We believe that deep learning can extract the deeper abstraction of the text and will help establish a more robust sentimental polarity classifier.
Databáze: Networked Digital Library of Theses & Dissertations
Externí odkaz: