Abstract
Large language models (LLMs) can generate synthetic video captions at scale, offering a practical way to enrich datasets that lack multiple human references. However, evaluating the semantic reliability of such captions is challenging, particularly when traditional lexical-overlap metrics (e.g., BLEU, ROUGE, METEOR, CIDEr, SPICE) are sensitive to surface-level wording and often misalign with human judgments. This paper investigates whether sentence embedding models provide more reliable evaluation metrics for LLM-generated video captions. We first compare five embedding models—Sentence-BERT, SimCSE, MPNet, and two GPT-based embeddings—against traditional lexical metrics using 450 caption pairs from VATEX, a benchmark video captioning dataset. Through validation with a linguistic specialist, we find that sentence embedding-based approaches achieve substantially higher agreement with expert judgments using Kendall’s τ correlation. We then used GPT-4o to generate ten synthetic captions per video from VATEX and apply the validated embedding metrics to compare semantic similarities with the human-human caption pairs. We applied the same pipeline to YuWa, a single-reference accessibility video dataset, which demonstrates a real-world use case. Our findings suggest that synthetic captions achieve more consistency in semantic similarity levels compared to independently written human captions. Notably, augmenting YuWa’s single human caption per video with ten LLM-generated captions produces semantic similarity scores matching or surpassing those of VATEX’s ten independently written human captions. This work provides the first expert-validated framework for evaluating LLM-generated video captions and demonstrates a practical pipeline for quality-controlled dataset augmentation in accessibility research.