A Grading Methodology for E-commerce Reviews based on a Mixed Lexical-semantic Approach

We introduce a methodology for grading sentiment in e-commerce reviews based on a mixed lexical and semantic approach, with a strong focus on interpretability. The methodology employs a simple, replicable, and operational procedure to identify sentiment-bearing substantives in textual reviews and uses them to assign granular sentiment scores on a 1–10 scale. Experimental results show that our method achieves performance comparable to zero-shot large language models (LLMs) when benchmarked against human-assigned grades. Unlike black-box LLM approaches, we offer enhanced transparency by explicitly highlighting the linguistic elements that drive its grading decisions.