Deep learning with knowledge graphs for fine-grained emotion classification in text

This PhD thesis investigates two key challenges in the area of fine-grained emotion detection in textual data. More specifically, this work focuses on (i) the accurate classification of emotion in tweets and (ii) improving the learning of representations from knowledge graphs using graph convolutional neural networks.The first part of this work outlines the task of emotion keyword detection in tweets and introduces a new resource called the EEK dataset. Tweets have previously been categorised as short sequences or sentence-level sentiment analysis, and it could be argued that this should no longer be the case, especially since Twitter increased its allowed character limit. Recurrent Neural Networks have become a well-established method to classify tweets over recent years, but have struggled with accurately classifying longer sequences due to the vanishing and exploding gradient descent problem. A common technique to overcome this problem has been to prune tweets to a shorter sequence length. However, this also meant that often potentially important emotion carrying information, which is often found towards the end of a tweet, was lost (e.g., emojis and hashtags). As such, tweets mostly face also problems with classifying long sequences, similar to other natural language processing tasks. To overcome these challenges, a multi-scale hierarchical recurrent neural network is proposed and benchmarked against other existing methods. The proposed learning model outperforms existing methods on the same task by up to 10.52%. Another key component for the accurate classification of tweets has been the use of language models, where more recent techniques such as BERT and ELMO have achieved great success in a range of different tasks. However, in Sentiment Analysis, a key challenge has always been to use language models that do not only take advantage of the context a word is used in but also the sentiment it carries. Therefore the second part of this work looks at improving representation learning for emotion classification by introducing both linguistic and emotion knowledge to language models. A new linguistically inspired knowledge graph called RELATE is introduced. Then a new language model is trained on a Graph Convolutional Neural Network and compared against several other existing language models, where it is found that the proposed embedding representations achieve competitive results to other LMs, whilst requiring less pre-training time and data. Finally, it is investigated how the proposed methods can be applied to document-level classification tasks. More specifically, this work focuses on the accurate classification of suicide notes and analyses whether sentiment and linguistic features are important for accurate classification