Transferring Labels based on Text Similarity: an Application for e-Portfolio Sentiment Analysis

This paper proposes a method for transferring labels during sentiment analysis based on text similarity. The data is sourced from the e-Portfolio system, which contains employees' feedback. Challenging, the data lacks labels, making sentiment analysis initially impossible. Thus, various similarity algorithms for labeling were examined, including Cosine Similarity, Jaccard Similarity, Jaro-Winkler Similarity, and Pearson Correlation Coefficient. Also, an upscaling technique was employed to ensure balanced labels. Natural Language Processing was used for word tokenization and feature engineering. Sentiment analysis was conducted using several classification algorithms such as Logistic Regression, Decision Tree Classifier, and Support Vector Machines to categorize sentiment (positive or negative). The results showed that the Cosine Similarity algorithm, along with the Bag of Words (BoW) feature through Logistic Regression, achieved a promising accuracy of 0.892 and an F1-Score of 0.883.