Semantic Textual Similarity Analysis of Clinical Text in the Era of LLM

Semantic textual similarity analysis finds its practical usage in many real-world natural language processing applications, including clinical and biomedical settings. Many occurrence frequency based techniques have been proposed in the past, from n-grams to term frequency-inverse document frequency. Semantic relation based methodologies, such as the WordNet and Wu and Palmer measure, have flourished too. Recently, vector embedding based approaches that leverage large language models lifted analysis performance significantly. However, many existing semantic textual similarity analyses in the clinical and biomedical domains only exploited the early generation of large language models, whereas the latter field is advancing rapidly. This paper aims to fill the gap by investigating the potential of the latest foundational language models. This paper proposed MLTE, a feature fusion approach that leverages multi-layer Transformer embeddings for semantic similarity analysis. The effectiveness of the proposed method was evaluated with public clinical datasets in English, Chinese, and Japanese. Experimental outcomes showed that the proposed method is competitive and outperformed related works over two out of four datasets.