MRPC

MRPC, or Microsoft Research Paraphrase Corpus, is a benchmark dataset used for paraphrase identification in natural language processing. It consists of pairs of sentences labeled to indicate whether the two sentences convey the same meaning. The corpus was created by researchers at Microsoft Research in the mid-2000s, and the sentence pairs were drawn from online news sources and other published texts. Each data point contains two sentences and a binary label: 1 if the sentences are paraphrases (semantically equivalent) and 0 otherwise.

The MRPC dataset is widely used to train and evaluate models for sentence similarity and paraphrase detection, including traditional machine learning approaches and modern neural networks. It was originally distributed with train, development (or validation), and test splits to support supervised learning and model comparison. In later years, MRPC became part of the GLUE benchmark, where it is used as one of several tasks to assess general language understanding. The task emphasizes detecting semantic equivalence rather than surface similarity, requiring models to handle nuances such as negation, tense changes, paraphrastic expressions, and syntactic variation.

Compared with some other paraphrase datasets, MRPC is relatively small, which has encouraged researchers to use data augmentation and transfer learning. Despite its size, it remains a standard resource for benchmarking paraphrase detection systems and for evaluating sentence-pair modeling capabilities.