| Author: | Xiang, Rong |
| Title: | Knowledge acquisition for affective analysis |
| Advisors: | Lu, Qin (COMP) Li, Wenjie (COMP) |
| Degree: | Ph.D. |
| Year: | 2022 |
| Subject: | Emotion recognition Discourse analysis -- Data processing Machine learning Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Computing |
| Pages: | xxii, 185 pages : color illustrations |
| Language: | English |
| Abstract: | Affective analysis, a broader term for sentiment analysis and emotion recognition, is highly demanded in social media text analysis. Current affective analysis methods mostly use machine learning methods which are heavily dependent on annotated training data. Affective knowledge implicitly presented in text is yet to be fully explored. Moreover, previous work was limited in incorporating knowledge-based resources into machine learning, especially deep learning methods. The main objective of this study is to enhance affective analysis methods by utilizing affective knowledge in several directions. More specifically, this study investigates the use of knowledge in five aspects, including (1) lexical knowledge (intra-word) to be incorporated effectively in machine learning processes, (2) cognition grounded affective lexicon for fine-grained sentiment analysis, (3) ontological knowledge (inter-word) for augmenting semantically sound training data, (4) lexical semantic information across languages through context change related features conveyed by writing system change, and (5) information of pragmatics for genre-specific text. Affective lexicons are valuable resources as affective knowledge. However, lexicons are often not readily available, especially for low-resource languages. This work explores leveraging internet resources to acquire sentiment lexicon to improve machine learning classifiers. A restaurant review website is adopted as the text source, obtaining sentiment lexicon with a hybrid method of manual and automatic collection. Evaluation using multimodel decisions in traditional machine learning and sentiment lexicon-based deep learning models shows that the internet can be an effective source for acquiring affective knowledge for resource-poor languages. Cognitive Studies have long proven that affect measurements with psycho-linguistic backing reveal universal agreement for a shared understanding among all people. Dimensional affective vectors measured by real numbers can provide more fine-grained lexical knowledge grounded by cognition science, yet they are less explored in affective analysis. This work explores using multi-dimensional affective knowledge to fine-tune deep learning classifiers. Affect-driven approaches are devised using affective vectors to highlight the importance of affectively significant words. Performance valuation validates the usefulness of multi-dimensional lexical knowledge even for deep learning-based models as it can better use implicit affective information in text. Deep learning methods demand even more annotated training data to fine-tune learning models. This work explores ontological knowledge to obtain more training data through lexical augmentation. The proposed method explores part-of-speech focused data augmentation using ontology-based semantically sound lexical substitutions to augment training instances. The syntactically bound and semantically sound method can significantly improve the performance of a number of machine learning methods. Text written on social media platforms is mostly informal, short, and filled with mixed language text. The mixed-use of text sometimes is intentional to reflect unfiltered emotion immediately. However, mixed scripts can also break semantic coherence and further complicate affective classifications. This work proposes a novel method for writing system change as an affective feature. A hybrid deep neural network is proposed to better use this feature type through an attention mechanism in the learning process. Evaluations on two Chinese mixed-script written corpora show that semantic knowledge conveyed by writing system changes are practical features for affective analysis in social network analysis. When fine-tuning deep learning models in a specific task or handling a particular text genre, information specific to the task or genre is very helpful. In this work, features in tweets are used as an example genre of text to demonstrate how pragmatics knowledge about a genre can be extracted for fine-tuning deep learning models. Patterns specific to this social media text are extracted and aggregated for a multi-label emotion recognition task. These genre-specific patterns complement general knowledge learning in the deep neural network. Performance gain on a Twitter benchmark dataset validates that pragmatics information can be applied to further improve state-of-the-art methods with extracted pragmatics knowledge. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/11685