Abstract
This paper presents a new approach called EmbedHDP, which aims to enhance the evaluation models utilized for assessing sentence suggestions in nursing care record applications. The primary objective is to determine the alignment of the proposed evaluation metric with human evaluators who are caregivers. It is crucial due to the direct relevance of the provided provided to the health or condition of the elderly. The motivation for this proposal arises from challenges observed in previous models. Our analysis examines the mechanisms of current evaluation metrics such as BERTScore, cosine similarity, ROUGE, and BLEU to achieve reliable metrics evaluation. Several limitations were identified. In some cases, BERTScore encountered difficulties in effectively evaluating the nursing care record domain and consistently providing quality assessments of generated sentence suggestions above 60%. Cosine similarity is a widely used method, but it has limitations regarding word order. This can lead to potential misjudgments of semantic differences within similar word sets. Another technique, ROUGE, relies on lexical overlap but tends to ignore semantic accuracy. Additionally, while BLEU is helpful, it may not fully capture semantic coherence in its evaluations. After calculating the correlation coefficient, it was found that EmbedHDP is effective in evaluating nurse care records due to its ability to handle a variety of sentence structures and medical terminology, providing differentiated and contextually relevant assessments. Additionally, this research used a dataset comprising 320 pairs of sentences with correspondingly equivalent lengths. The results revealed that EmbedHDP outperformed other evaluation models, achieving a coefficient score of 61%, followed by cosine similarity, with a score of 59%, and BERTScore, with 58%. This shows the effectiveness of our proposed approach in improving the evaluation of sentence suggestions in nursing care record applications.