Using LLMs for Evaluating QA Systems: Exploration and Assessment

The evaluation process of detecting the similarity between reference text and Large Language Model (LLM)-generated text is a challenging problem. It is difficult to measure it automatically. Traditional metrics, such as ROUGE and BERTScore, have been shown to have some limitations. They showed a relatively low correlation with humans. Also, they penalize LLM-generated text, and have difficulty recognizing noise in text and qualitatively evaluate similarity in texts.In this paper, we introduce an approach to studying the use of LLM to evaluate the similarity between LLM-generated answers and reference answers. The approach framework includes classes, prompt, and classification schemes. The framework defines classes to measure similarity. The prompt will include the classes and their definitions. We use two types of prompts: instructed and uninstructed. The instructed prompt contains specific evaluation rules for conducting the LLM to evaluate the similarity between LLM-generated answers and reference answers. The uninstructed prompt, though it will have an evaluation request, will not contain evaluation rules. The classification schemes range from binary classification to multi-class classification; the latter is finer-grained and more informative. Also, the classification schemes group into two categories, where in one category the classes are viewed as mutually exclusive and the classification is unguided, while in the other category the classes are not all mutually exclusive but can be viewed as partly hierarchical, and the classification is guided, i.e., totally ordered or partially ordered (a hybrid). The approach will use a classification scheme to allow LLMs to invoke prompts and classes to perform the evaluation. We experimented with the similarity evaluation with GPT-4 and Gemini.Our results indicate that the binary classification scheme shows significant accuracy results. In the multi-class classification schemes, the unguided classification showed very poor evaluation performance, while the guided classification showed excellent performance in the 90-percentiles of accuracy. Furthermore, both Gemini and GPT-4 perform very well as evaluators, especially under totally ordered classification, while in the case of partially ordered classification, the winning combination turned out to be GPT-4 as evaluator and Gemini as answerer.