الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 123 |  |  |
| | | from | 123 | | |
Abstract
People are sharing their thoughts, their experiences, their recommendations, their feedback, their opinions and their emotions throughout social media and recommendation systems. Analysing people’s emotions is an evolving Natural Language Processing (NLP) challenge. NLP role is to extract information from text, detecting emotion of text or extracting emotion causes. Understanding the content of text helps a lot in NLP tasks in general.
The thesis will tackle two important fields in emotion analysis which are emotion detection and emotion cause extraction. Both emotion detection and emotion cause extraction are main tasks in Emotion Analysis. Emotion detection is a field where emotions are detected from text. Emotion cause extraction is a field where causes behind emotions are extracted from text.
The main objective of the thesis is to find the best model for emotion detection then extract the causes behind emotions. We have applied different techniques to detect emotions from text and analysed the results. We have applied different features representation methods using different models such as Convolutional Neural Networks (CNN), Convolutional Neural Networks- Long Short-Term Memory (CNN-LSTM), Convolutional Neural Networks- Bidirectional Long Short-Term Memory (CNN-BiLSTM). Also, Convolutional Neural Networks- Bidirectional Gated Recurrent Network (CNN-BiGRU), Bidirectional Long Short-Term Memory (BiLSTM), Bidirectional Encoder Representation Transformers (BERT) and Embeddings Language Model (ELMo) models have been used. The comparative analysis study objective is to compare different techniques, analyze results, and reach the best solution for emotion detection. This comparative study has been applied to the International Survey on Emotion Antecedents and Reactions (ISEAR) benchmark dataset. We have experimented the models using different number of emotions of the ISEAR dataset seven emotions, five emotions and four emotions. The seven emotions are joy, fear, anger, sadness, disgust, shame, and guilt. The five emotions are joy, sadness, fear, shame, and guilt. The four emotions are joy, sadness, fear, and shame. Using four emotions of the dataset has achieved better results that five and seven emotions. We have observed that Bidirectional Long Short-Term Memory BiLSTM and BERT models are the best models used in our study in emotion detection due to their ability in capturing context. BERT outperforms BiLSTM in detecting emotions using F1-score evaluation metrics. Using four emotions of ISEAR dataset, BERT has achieved 0.84 F1-score while BiLSTM has achieved 0.75 F1-score.
We have introduced emotion cause extraction in Arabic Language. Arabic language resources are very limited in this field. There are some corpora built for western languages like English and far east languages like Chinese. We have applied emotion detection and extracted the causes behind emotions from text. An Annotated Dialectal Arabic corpus has been constructed for the purpose of both emotion detection and emotion cause extraction tasks. We have used two annotation approaches which are keyword spotting technique and BERT as a learning-based approach to annotate emotions in our Arabic constructed dataset. We have compared the two annotation approaches. BERT has achieved 0.49 F1-score while keyword spotting technique has achieved 0.69 F1-score. But keyword spotting has failed to detect emotions of some reviews completely. On the other hand, BERT has detected emotions of all reviews. We have calculated the inter-annotator agreement between the two annotators. It is a fair agreement. Arabic emotion detection has been tackled using BERT on our annotated constructed dataset as it has been observed the best model to achieve results in our comparative analysis study.
Causes are then annotated manually. Arabic emotion cause extraction has been tackled using two models which are Bidirectional Long-Short Term Memory-Conditional Random Fields (BiLSTM-CRF) and Bidirectional Encoder Representation Transformers – Conditional Random Fields (BERT-CRF) models. They have achieved the best results in emotion detection comparative analysis study. We have applied both models but with a modification which is adding CRF layer at the end of each model. This CRF layer has been added so that the models can be tuned with the emotion cause extraction task which is tackled as sequence labelling task. The two models have been evaluated using the Dialectal Arabic constructed corpus. BERT-CRF model has achieved better F1-score than BiLSTM-CRF model. BERT-CRF has achieved 0.84 F1-score while BiLSTM has achieved 0.56 F1-score using token-level evaluation measure.