NEURAL NETWORKS FOR TEXTUAL EMOTION RECOGNITION AND ANALYSIS

Abstract

Textual Emotion recognition (TER) is an important task in Natural Language Processing (NLP), due to its high impact in real-world applications from health and wellbeing to author profiling, consumer analysis and security. The task of TER is often formalised as the process of detecting, interpreting, and understanding users’ emotions (i.e., the experience of feeling). This process can be performed on different units of analyses like words, phrases, sentences, documents and tweets/posts. Since the majority of existing emotion corpora are collected from social media data, the focus of this thesis is specifically on tweets and posts. This thesis investigates three research questions, as discussed below.

Firstly, we recommend that emotion correlations and associations should be taken into consideration when dealing with the classification and identification of emotion expressions in texts. This aims to enable TER models to account for the ambiguity and complexity of the task by taking into account that certain emotions can be highly correlated with each other. More specifically, we want to leverage information about emotion-to-emotion correlations, as well as associations between emotions and words in the case of multi-label emotion classification. To address the first research question, we propose “a novel model SpanEmo casting multi-label emotion classification as a span-prediction problem”, which can help TER models learn associations between labels and words in an input instance. Furthermore, we introduce a training objective focused on modelling multiple co-existing emotions in the input instance. Experiments performed on a multi-label emotion corpus across multiple languages demonstrate our method’s effectiveness in terms of improving the model performance and learning meaningful correlations as well as associations.

Secondly, existing emotion corpora labelled for a single-label emotion classification problem are more than those labelled for a multi-label emotion classification problem. Through extensive experiments, we observe that certain emotions are highly associated with each other, causing TER models to select incorrect predictions. Therefore, we want to improve TER models ability to handle highly associated emotions by introducing discriminator features. To address this, we introduce an auxiliary task to emotion classification. Furthermore, we introduce a method for evaluating the impact of intra- and inter-class variations on each emotion class. Experiments performed on three emotion corpora demonstrate our method’s effectiveness in terms of improving the prediction scores and producing discriminative features against highly confused emotions.

Thirdly, emotion features can be beneficial for related tasks that share common patterns with emotion. Based on the observation that in social media, negative sentiments and emotions are frequently expressed towards certain topics, such as politics, but also adverse drug reactions and depression. We examine the benefits of emotion features to the last two topics, while at the same time modelling them without the use of handcrafted features. To avoid the use of hand-crafted features, we decide to use transfer learning by training a neural model on sentiment/emotion corpora and then fine-tuning it on the target tasks. We also adapt our proposed model for the first research question to both Adverse Drug Reactions (ADRs) and depression. Experiments performed on different corpora for the topics of ADRs and depression demonstrate our model’s effectiveness in achieving strong performance compared to previous approaches and being easily adapted to other tasks.