description.md

What you need to use this demo?

1. English Sentences that you want to translate at document-level
2. Sentence-level German translations that Llama2 will refine at document-level

You can also simply play around with the examples :D

Going through the examples!

We provide three different type of examples covering contextual phenomena such as terminology, formality level and pronoun disambiguation exploiting manual feedback.

• Terminology (Example 1 & 2): You can see that the term "Large Language Model" is translated into "große Sprachmodelle". Although this is correct, lets say you decide that it should be kept as "Large Language Models" without change in the German translations. Then, you correct the first sentence by replacing it. The model then follows this and uses same convention for the next sentence!
• Formality (Example 3 & 4): The word "you" in English can be translated to German as Sie (Formal) and Du (Informal). The sentence-level model outputs are inconsistent and use both levels. However, you see that the LLM outputs are only formal. If you want it to be informal, you can change the formality level of the first translation and then let the model generate from that point.
• Pronouns (Example 5 & 6): Here, the translation of it is unknown from the sentence alone and can be "Er", "Es" and "Sie" depending on the gender. Since, it is refering to garland, the right choice is Er. However, the LLM mistranslates the garland into "Krone" and uses "Sie" as the pronoun. However, once you fix the mistake and replace "Krone" with "Kränz", LLM refines the translation and use the right pronoun!