Please use this identifier to cite or link to this item:
Title: A study on using personal profiles for a biased reader emotion prediction model
Authors: Long, Yunfei.
Degree: Ph.D.
Issue Date: 2019
Abstract: In the age of social media, users express their personal feelings and emotions through the Web. In addition to understanding the emotion of the public, it is also important to learn how individual subjectivity and their bias affect emotion analysis especially in social media and review texts. The main objective of this study is to investigate the effect of personal profiles in emotion analysis. This thesis focuses on emotion analysis from social media and review text, and studies four areas in subjectivity linked emotion analysis, including (1) improving emotion analysis from cognitive perspective by identifying linguistic features more appropriate for social media text, (2) using cognition grounded data to improve emotion prediction models, (3) learning the representation of user profiles by addressing the data sparseness through two methods, and (4) incorporating user profiles into emotion analysis model to take subjectivity as a bias into consideration. Based on the premise that emotion is a personalized cognitive process encoded in different types of linguistic features, we first explore additional linguistic features relevant to social media and review text. In addition to the traditional lexical features, we propose a linguistic-driven model to explore the use of morpho-syntactic features such as passive construction, verb order as well as some less explored orthographic text features such as unusual use of punctuations and code-switches which are often seen in this genre of text. Evaluation on both a personally generated micro-blog dataset and a formal news collection shows that incorporating our proposed linguistic features can improve emotion analysis by introducing genre and stylist information encoded in social media text. Cognitive studies support the linguistic fact that not all words contribute equally to the semantic and affective meaning of sentences. Some words are more important than others in conveying semantic meanings. Computational attention models are proposed to give different weights to different words in text. However, many attention models are built using local text features through distributional similarity which lack the theoretical foundation to reflect the cognitive basis. This motivates us to explore the use of eye tracking data as cognition based information to train attention models to further improve the performance of linguistic-driven models. Our proposed method can capture attention of words more comprehensively using a two level approach. Evaluations show that our method outperforms the state-of-the-art methods significantly. We prove that cognition grounded data can be used to improve attention mechanisms and thus indirectly improves the performance of sentiment analysis.
Presenting user profile using dense vector representation through user activities is the key to build user profiling models. However, like many social media data, user activities follow the long-tail distribution. Thus, the key to obtain a better representation of user profiles is to address the data sparseness issue. Inspired by the stimulus generalization theory and the halo effect in cognitive science, we first propose a novel approach to predict user preferences by learning from both observed comments and missing comments based on the missing-not-at-random hypothesis. Then we explore methods to extend context for user profiles through network links in social media data. We propose a novel approach to learn node embedding through a joint learning framework of both network links and text associated with nodes. The method can handle both homogeneous networks and heterogeneous networks with multiple types of links. A novel attention mechanism is also proposed to make good use of text extended through links to obtain a larger network context. Finally, emotion as a cognitive process is largely subjective and user bias plays a significant role in emotion analysis. Lenient users tend to give higher ratings than finicky ones even if they review the same products with similar wording, On the other hand, popular products do receive higher ratings than those unpopular ones because the aggregation of user reviews still shows the difference in opinions for different products. In this work, we propose a deep learning method to incorporate biased user profile into emotion analysis for review text. Individual user bias as user profiles, is learned through a neural network model. Then, user profiles as a collection are aggregated by a memory network to encode the user bias. A separate memory network is also used to learn product information. In this way, user profiles and product information can be captured more efficiently as they are different by nature. Lastly, the dual memory networks is merged into a unified classifcation model for joint optimization. Evaluations on three commonly used benchmark datasets show that our dual memory network model is more effective than the state-of-art methods.
Subjects: Hong Kong Polytechnic University -- Dissertations
Machine learning
Emotions -- Data processing
Social media
Pages: xxii, 193 pages : color illustrations
Appears in Collections:Thesis

Show full item record

Page views

Citations as of May 15, 2022

Google ScholarTM


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.