MATCHA: Matching Text via Contrastive Semantic Alignment

Researchers introduced MATCHA, a new automatic metric for evaluating text similarity and correctness.

It uses a dual-view contrastive approach: aligning a generated text with a reference while penalizing an adversarial contradiction.

Across eight benchmarks, MATCHA outperformed common overlap- and embedding-based metrics such as ROUGE and BERTScore.

It showed especially strong gains on TruthfulQA and was also rated effective by human evaluators.

The metric addresses a key weakness in LLM evaluation: contradictory texts can otherwise score as deceptively similar.