Predicting similarity between sentence pairs is essential for applications such as recommender systems and plagiarism detection. There have been several categories of approaches for predicting sentence similarity. This paper combines approaches from two categories, semantic and structural similarity, to find a hybrid approach that aligns more closely with how humans determine similarity.
Extensive background research is conducted to understand the scope of the problem, to be able to understand human psychology, as well as the advantages and disadvantages of already existing approaches. Based on the insights from the background research, we propose a novel hybrid approach, combining semantic and structural similarity metrics. The proposed approach is evaluated on the STSS-131 and MSRP datasets and compared with other common approaches and SentenceBERT, a deep learning algorithm. The proposed approach does not perform as well as SentenceBERT but makes up for this by being more explainable and outperforming all other traditional machine learning approaches in its accuracy in predicting sentence similarity. This paper also provides a critical conclusion with recommendations for further improvements.