Text-Translation

Basic template for translating comments between Greek ↔ English, English ↔ Italian, and English ↔ German. Translations are generated using prompts sent to LLMs hosted on AWS Bedrock.

LLMs used

1. Claude 3.5
2. Llama 3.3 70b
3. Deepseek R1

Pipeline

1. We fetch data from the Helsinki-NLP/europarl dataset on HuggingFace Dataset.
2. Each comment (a paragraph of conversation) is sent to the LLM for translation.
3. For each comment, we create a dictionary with the following keys:

• from_language: The source language sent to the model
• to_language: The target language to be translated into
• ai_response: The translated output generated by the model

4. This dictionary is then sent for evaluation.

Evaluation

We evaluate the translations using:

1. BertScorer
2. Meteor_Score
3. Rougue