Excited to share that we have released RuBLiMP (Russian Benchmark of Linguistic Minimal Pairs), a novel benchmark for evaluating Russian language models (LMs).

❓RuBLiMP consists of 45,000 minimal pairs and includes 12 grammatical phenomena well-represented in Russian linguistics, covering morphology, syntax, and semantics. A minimal pair consists of a grammatical and an ungrammatical sentence (e.g., The cat is on the mat / *The cat are on the mat), and an LM is expected to prefer the grammatical one based on the scoring function.

Our approach allows to:
🔸generate minimal pairs at scale from any text domain
🔸estimate if a grammatical sentence appears in the LM's pretraining corpus

💡RuBLiMP can be used for evaluating the sensitivity of LMs to grammatical phenomena in Russian and for developing ranking and grammatical error detection methods.

🔸 Read more in our pre-print: https://arxiv.org/abs/2406.19232
🔸 HuggingFace: https://huggingface.co/datasets/RussianNLP/rublimp
🔸 GitHub: https://github.com/RussianNLP/RuBLiMP

исследовано

Excited to share that we have released RuBLiMP (Russian Benchmark of Linguistic Minimal Pairs), a novel benchmark for evaluating Russian language models (LMs).

❓RuBLiMP consists of 45,000 minimal pairs and includes 12 grammatical phenomena well-represented in Russian linguistics, covering morphology, syntax, and semantics. A minimal pair consists of a grammatical and an ungrammatical sentence (e.g., The cat is on the mat / *The cat are on the mat), and an LM is expected to prefer the grammatical one based on the scoring function.

Our approach allows to:
🔸generate minimal pairs at scale from any text domain
🔸estimate if a grammatical sentence appears in the LM's pretraining corpus

💡RuBLiMP can be used for evaluating the sensitivity of LMs to grammatical phenomena in Russian and for developing ranking and grammatical error detection methods.

🔸 Read more in our pre-print: https://arxiv.org/abs/2406.19232
🔸 HuggingFace: https://huggingface.co/datasets/RussianNLP/rublimp
🔸 GitHub: https://github.com/RussianNLP/RuBLiMP

arXiv.org

RuBLiMP: Russian Benchmark of Linguistic Minimal Pairs

Minimal pairs are a well-established approach to evaluating the grammatical knowledge of language models. However, existing resources for minimal pairs address a limited number of languages and...

🔥10👍5

hottg.com/nlp_seminar/130

1.64K viewsKatya Artemova, Jul 10, 2024 at 12:29