Chain-of-Thought Reasoning Improves Context-Aware Translation with Large Language Models

This dataset contains the results of the Generative Translation Task from several Large Language Models. The Generative Translation Task evaluates Large Language Models (LLMs) on two context-dependent translation challenges: anaphora resolution and lexical cohesion. The models are prompted with Chain-of-Thought Reasoning prompts. They translate pairs of sentences, where the second sentence depends on the first one. The second sentence is contextually dependent on the first one. The evaluation scores measure whether the translation from the LLMs is contextually correct. A full description of the task has been published separately. All experiments use the 'discourse-mt-test-sets' dataset and compare two prompting strategies: no reasoning and structured reasoning.