Language enables humans to communicate with others. For instance, we talk, give our opinions and suggestions all using natural language; to be more precise, we use words while communicating with others. However, in today’s world, we wish to communicate with computers, just like humans. It is not an easy task because humans communicate in an unstructured and informal way, whereas computers need structured and clean data. So it is essential for computers to understand and classify text accurately for proper human-computer interactions. For classifying a text, the first question we must address is how to improve the low-quality text. The next immediate challenge is to have the best representation so that text can be classified accurately. The way text is organized reflects polysemy, semantic and syntactical coupling relationships which are embedded in its contents. The effective capturing of such content relationships is thereby crucial for a better understanding of text representations. This is especially challenging in the environments where the text messages are short, informal and noisy, and involves natural language ambiguities. The existing sentiment classification methods are mainly for document and clean textual data which can not capture relationships, different attributes and characteristics within tweet messages.

Social media analysis, especially the analysis of tweet messages on Twitter has become increasingly relevant since the significant portion of data is ubiquitous in nature. The social media-based short text is valuable for many good reasons, explored increasingly in text analysis, social media analysis and recommendation. In the same time, there is a number of challenges that need to be addressed in this space. One of the main issues is that the traditional word embeddings are unable to capture polysemy (assigns the same representation of a word irrespective of its context and meaning) and out of vocabulary words (assigns a random representation). Furthermore, traditional word embeddings fail to capture sentiment information of words which results in similar word vector representations having the opposite polarities. Thus, ignoring polysemy within the context and sentiment polarity of words
in a tweet reduces the performance for tweets classification.

In order to address the above-mentioned research challenges and limitations associated with word-level representations, this thesis focuses on improving the representation of low-quality text by improving the unstructured and informal nature of tweets to utilize the information thoroughly and manages the natural language ambiguities to build a more robust sentiment classification model. As compared to previous studies, the proposed models can deal with the ubiquitous nature of the short text, polysemy, semantic and syntactical relationships within a content, thereby addressing the natural language ambiguity problems.

Chapter 4 presents the effects of pre-processing techniques using two different word representation models with the machine and deep learning classifiers. Then, we present our recommended combination (approach) of different pre-processing techniques which improves the low quality, by performing sentiment-aware tokenization, correction of spelling mistakes, word segmentation and other techniques to utilize most of the information hidden in unstructured text. The experimental result shows that the proposed combination performs well as compared to other combinations.

Chapter 5 presents the hybrid words representation. In this chapter, we proposed our Deep Intelligent Contextual Embedding for Twitter sentiment analysis. The proposed model addresses the natural language ambiguities and is devised to capture polysemy in context, semantics, syntax and sentiment knowledge of words. Bi-directional Long-Short Term Memory wth attention is employed to determine the sentiment. We evaluate the proposed model by performing quantitative and qualitative analysis. The experimental results show that the proposed model outperforms various word embedding models in the sentiment analysis of tweets. Above mentioned methods can be applied to any social media classification task. The performance of proposed models is compared with different models that support the effectiveness of the proposed models and bound the information loss in their generated high-quality representations.