This is a model for evaluation of naturalness of short Russian texts. It has been trained to distinguish human-written texts from their corrupted versions.
Corruption sources: random replacement, deletion, addition, shuffling, and re-inflection of words and characters, random changes of capitalization, round-trip translation, filling random gaps with T5 and RoBERTA models. For each original text, we sampled three corrupted texts, so the model is uniformly biased towards the
On our private test dataset, the model has achieved 40% rank correlation with human judgements of naturalness, which is higher than GPT perplexity, another popular fluency metric.
- Downloads last month