Affective texts play a key role in sentiment classification/prediction and decision making. They are being increasingly used to form and/or share sentiments in financial, economic and/or political applications. However, the processing time is exponentially increased for large affective textual datasets. Moreover, casual expressions such as emoji, slang, abbreviation and misspelling words usually make data analysis (i.e., text classification) complicated. This paper proposes a pipeline model consisting of data pre-processing, feature extraction and classification model training to classify affective text datasets. It offers three contributions including Emoji recovery, misspelling word correction and abbreviation translation that results in maximised classification accuracy. A rigorous experimental plan is designed to evaluate the performance of the proposed approach according to three factors including dataset size (i.e., small, medium and large), NLP feature extraction technique (i.e., TF-IDF, word2vec and BERT) and classification model (i.e., MLP, Logistic Regression, Naive Bayes and SVM). In addition, the proposed approach is compared with a well-known Deep Learning sentiment analysis approach, named sentimentDLmodel, which addresses a pre-trained sentiment analysis. According to the results, the proposed approach significantly outperforms benchmarks in terms of classification model accuracy for most cases.