Natural Language Processing Research

Data pre-processing is an important step in machine learning text classification as it improves data quality and hence improves performance of trained algorithms. We experimentally compare the following pre-processing techniques: punctuation removal, lowercasing, typos replacement, slang replacement and stop-word removal on a Swahili short message service (SMS) dataset for topic classification. Different machine learning algorithms are applied such as Random Forest, Stochastic Gradient Descent, RNN LSTM Unidirectional, RNN LSTM Bidirectional and Support Vector Machine. We analyze the impact of the pre-processing techniques on classification accuracy and f1-score. Our experiments show that all pre-processing steps, when applied separately, have a positive impact on the performance of all evaluated classification algorithms. Among all experimented pre-processing steps, stop-word removal has the highest impact on performance of both accuracy and f1-score metrics. Also, of all evaluated algorithms, Random Forest and Stochastic Gradient Descent are the most positively affected with pre-processing steps.

• The study aims to evaluate the performance of classification algorithms which are vital in automating error-prone manual work.

• The study illustrates the importance and effects of different pre-processing steps for Swahili textual data.

• This article will enable future researchers to decide which pre-processing steps for Swahili textual data are best for their respective machine learning tasks.