Md Shofiqul Islam, <br/><br/>
A hybrid deep learning method for multilabel sentiment analysis using attentive capsule and dual structured recurrent network / 
Md Shofiqul Islam 
 - xii, 107 pages  illustrations (some color) ;  30 cm. +  1 CD-ROM 
<br/><br/> Faculty of Computing 
<br/><br/>Thesis (Master of Science) -- Universiti Malaysia Pahang – 2022 
<br/><br/>Includes bibliographical references 
<br/><br/> People express opinions freely in social media platforms such as Facebook, Twitter and Instagram. Sentiment analysis is a machine learning based technique that takes the conversation of social media users about certain products or issues and puts them into context to understand their opinion on that particular topic of interest. Recent developments in machine learning show that they are helpful in sentiment classification particularly to classify human-like concentration words. However, the existing models are typically inadequate when dealing with certain characteristics of sentences. They still have limitations in handling sentences that contain multilabel texts and the sentences that are syntactically complex structured. Due to these two limitations, the accuracy of existing models can still be improved and will be done so when the limitations are properly handled. In order to solve these limitations, a hybrid deep learning method (HACDD) for sentiment analysis to recognize multilabel text emotion is proposed. At first, the data will have to undergo a pre-processing. After preprocessing, the data is embedded using GloVe vector that handles semantic relation among similar words. This proposed model is using a hierarchical attention-based capsule network based on dual and bidirectional recurrent neural networks (RNN). The BiGRU-BiLSTM (dual structured) based framework has been implemented to automatically extract and handle long ranged features. Then, the Dilated Convolutional Neural Network (CNN) layer is implemented to handle specific features with less convolutional operation and less usage of weights. This hybrid model is constructed to handle most relevant and informative feature by dynamically routing its hierarchical system to an attentive capsule. The model is also fitted with a hierarchical attention mechanism to allow full use of the details to effectively acquire semantic knowledge to handle complex sentences. The proposed models were trained and tested on ER (Emotion Recognition) dataset. This dataset was then trained and tested with nine other benchmark models. Our method obtained higher accuracy of 95.46% on ER. The experimental findings reveal that HACDD outperformed all benchmark models and also the best one in terms of training time, accuracy, and handling of complex sentence. 
<br/><br/><label>ISBN: </label>THE0009544 (Local) 
<br/><br/><label>Subjects--Corporate Names: </label><br/>Faculty of Computing--Dissertations
<br/><br/><label>Subjects--Topical Terms: </label><br/>Universities and colleges--Dissertations<br/>Theses--Dissertations