BERTScore
BERTScore is a text generation evaluation metric based on pre-trained BERT models. It measures semantic fidelity by computing the contextual word-embedding similarity between candidate and reference texts.
Timeline (last 90 days)
BERTScore对事实性错误可能给出高分，因此必须与事实性评审器搭配使用。
BERTScore使用预训练语言模型的上下文嵌入，计算参考句与预测句词语间的余弦相似度，汇总为precision、recall、F1。
Metrics based on pretrained embeddings, such as BERTScore, attempt to introduce semantic similarity but still operate at the sentence level and struggle to track whether specific factual claims in an answer are consistent with the context.
BERTScore measures representational similarity rather than factual accuracy, which is the reason the research team additionally introduced expert review.
NepKANUN's evaluation uses BERTScore as the measurement metric and tests across question difficulty levels.
BERTScore measures quality by computing the cosine similarity between candidate and reference texts in a semantic embedding space, distinguishing it from lexical-overlap-based metrics such as BLEU and ROUGE.
Output fidelity evaluation combines multiple metrics including BLEU, ROUGE-1/2/L, BERTScore-F1, and SentenceBERT cosine similarity
对于追求智能而非逐字照搬的转写系统，可能需要引入语义相似度（如 BERTScore）、可读性评分、信息完整度等更综合的评估维度
All Facts (8)
BERTScore对事实性错误可能给出高分，因此必须与事实性评审器搭配使用。
50%UnverifiedBERTScore使用预训练语言模型的上下文嵌入，计算参考句与预测句词语间的余弦相似度，汇总为precision、recall、F1。
50%UnverifiedMetrics based on pretrained embeddings, such as BERTScore, attempt to introduce semantic similarity but still operate at the sentence level and struggle to track whether specific factual claims in an answer are consistent with the context.
50%UnverifiedBERTScore measures representational similarity rather than factual accuracy, which is the reason the research team additionally introduced expert review.
50%UnverifiedNepKANUN's evaluation uses BERTScore as the measurement metric and tests across question difficulty levels.
50%UnverifiedBERTScore measures quality by computing the cosine similarity between candidate and reference texts in a semantic embedding space, distinguishing it from lexical-overlap-based metrics such as BLEU and ROUGE.
50%UnverifiedOutput fidelity evaluation combines multiple metrics including BLEU, ROUGE-1/2/L, BERTScore-F1, and SentenceBERT cosine similarity
50%Unverified对于追求智能而非逐字照搬的转写系统，可能需要引入语义相似度（如 BERTScore）、可读性评分、信息完整度等更综合的评估维度
50%