As a fundamental natural language processing (NLP) task, text emotion classification aims to correctly identify the emotion expressed by short texts from a given set of emotion labels.

With the advent of social media and e-commerce, the volume of online generated content, such as tweets, news feeds, and customer reviews, has far outpaced an individual’s reading and comprehending ability. Thus, automatic emotion detection and classification play an important role in opinion mining and has extensive application in various fields. However, applying supervised models in such a domain-dependent task as emotion classification is limited by two major challenges.

The first challenge is the domain-specific feature insufficiency. The advance in pretrained language models has greatly promoted the development of NLP. However, such representation learning models, following the distributional hypothesis that contextual texts shall present similar meanings, only encode semantic continuity and lack critical features in a domain-dependent task.

The second challenge is the emotional label ambiguity issue. Due to the subjectiveness and fuzziness of emotions, researchers have been aware that it is ubiquitous to observe multiple emotions in a sentence. The one-hot label approach is not informative enough to reflect the relationships between the text and emotion labels in emotion-related classification tasks and compromises the performance of classifiers. To address the challenges above, we will focus on the following objectives in this thesis:

1. To propose an emotion lexicon technique incorporating domain knowledge that is suitable for employment in deep learning models.

2. To exploit the proposed emotion lexicon as an emotion representation to address the domain-specific feature insufficiency.

3. To augment label information and produce an emotion distribution label using the proposed emotion lexicon technique for emotion distribution learning to address the emotional label ambiguity issue.

4. To exploit global statistical information as an additional feature to address the feature insufficiency.

Existing works adopt the categorical lexicons tagged by predefined emotion taxonomies to link affective words with emotions. However, in these lexicons, emotion tags are restricted to a specific set of basic emotions. Moreover, the emotional intensity is ignored, making such a method less flexible to be employed in a deep learning model. Our first work proposes the word-level emotion distribution (WED) vector by incorporating domain knowledge and a dimensional lexicon to achieve Objective 1. The proposed method can link a word with more generic and fine-grained emotion taxonomies with quantitatively computed intensities. We propose two schemas to utilize the WED vector implicitly and explicitly to facilitate classification. The explicit approach exploits WED as emotional word embedding to enhance sentiment features in conformity with Objective 2. The implicit approach implements a rule-based conversion strategy to augment information in the label space to address Objective 3.

Limited studies attempt to investigate the emotional perspectives, which are to understand how the emotion of a sentence is constructed. In the second work, we propose EmoChannel capture the intensity variation of a particular emotion in time series based on Objective 2. The resulting information of 151 available fine-grained emotions is utilized to comprise the sentence-level emotion construction. Furthermore, we employ the dependency relationship within all emotions to enhance emotion classification performance.

An emotion distribution label may cause inter-class confusion between similar emotion categories and introduce undesirable noise during training, which limits the improvements of the implicit approach in our first work. Meanwhile, current studies neglect the fine-grained emotions associated with these coarse-grained labels. To address the issue caused by utilizing a coarse-grained distribution and pave ways to achieve Objective 3, we propose a general and novel emotion label extension method based on fine-grained emotions in the third work. Specifically, we first identify a mapping function between coarse-grained emotions and fine-grained emotion concepts and extend the original label space with specific fine-grained emotions. Then, we generate a fine-grained emotion distribution and utilize it as a model constraint to incorporate the dependencies among fine-grained emotions.

Besides enhancing semantic features using lexicons, we also incorporate some intrinsic statistical features of the corpus, like word frequency and distribution, to fulfil Objective 4. Compared with external knowledge, the statistical features are deterministic, easy-to-retrieve, and naturally compatible with corresponding tasks. In the fourth work, we propose an Adaptive Gate Network (AGN) to consolidate semantic representation with statistical features selectively. AGN encodes statistical features through a variational component and merges information via a well-designed valve mechanism. The experiments indicate the robustness of AGN against adversarial attacks of manipulating statistical information.