基於多任務深度學習網路訓練之情感辭典萃取方法

Abstract

情感分析是多年來自然語言處理問題中很受歡迎的一個大類別。例如在屬性層級的情感分析的問題中，研究者經常使用外部的情感辭典來改進其模型的表現。而各種聚焦在情感辭典萃取方法的文獻也逐漸增加，近年來以深度學習模型自動化辭典萃取過程的方法尤其受到歡迎。 然而情感辭典會遇到的問題是許多情感字詞並不見得永遠代表正面或是負面的涵意。根據它們所形容的對象不同，有些情感字詞可能會代表完全相反的情感極性。這使我們認為建立出一個能帶有「意見字詞」(Opinion words)以及「意見-目標字詞之間的相依性資訊」的情感辭典將會對於其他應用辭典的情感分析任務有更大的幫助。 雖然許多文獻著重於萃取用於屬性層級的情感分析之情感辭典，但是只有少數的文獻針對「意見-目標字詞之間的相依性資訊」進行分析與萃取。因此，我們在這篇論文中將提出一種基於多任務深度學習網路訓練之模型架構作為情感辭典萃取方法，不僅針對文字的情感極性做萃取，更能同時聚焦在意見字詞以及其修飾的目標字詞之間的相依性上。實驗證明我們提出的模型架構，確實能有效地捕捉到部分意見-目標字詞相依性的資訊，並獲得更佳的萃取結果。

Sentiment analysis is a popular category of natural language processing tasks over years. In many categories of tasks, such as aspect-based sentiment analysis task, prior studies often use external sentiment lexicons to improve the effectiveness of their models. More and more studies focusing on developing sentiment lexicon extraction methods have been proposed. In recent years, the corpus-based lexicon extraction approach using the deep learning models is particularly popular. The problem with existing sentiment lexicons, however, is that many opinion words do not always have the same positive or negative polarity. Depending on targets, some opinion words may represent completely opposite sentiment polarities. This leads us to consider that building a sentiment lexicon with information of the dependencies between opinion words and their corresponding target words extracted from a given corpus may be helpful for other sentiment analysis tasks using external sentiment lexicons. Although many prior studies focus on extracting lexicons for aspect-based sentiment analysis, only few of them analyze and identify the dependencies between opinions and their target words. Therefore, in this research, we will propose a deep-learning-based sentiment lexicon extraction method using the multi-task learning, which not only focuses on the dependencies between opinions and their target words, but also extracts the sentiment polarity of input documents (product reviews) at the same time. Experiments show that our proposed method can effectively capture the dependencies of opinion words and their target words.