Google

The instructional part of the Deliberate-then-Generate (DTG) prompt template for translation tasks directs the model to perform a two-step analysis. The model is first asked to identify the type of error in an existing translation and then to proceed with refining it. This directive is explicitly stated as: `Please first detect the type of error, and then refine the translation.`

Instructional Component of the DTG Prompt Template for Translation Refinement

A complete prompt template for the Deliberate-then-Generate (DTG) method in a Chinese-to-English translation task involves a multi-step structure. First, it presents the source Chinese sentence and its initial English translation. Then, it directs the model to analyze the translation by identifying any errors and subsequently generating a refined version. Finally, it cues the model's output by explicitly starting the response line with 'Error Type:'. The full template is structured as follows:

```
Given the Chinese sentence: {*source*}
The English translation is: {*target*}
Please first detect the type of error, and then refine the translation. 
Error Type:
```

Complete DTG Prompt Template for Chinese-to-English Translation

In the Deliberate-then-Generate (DTG) framework, a model might identify 'Incorrect Translation' as the specific error type when analyzing a given translation. This serves as a concrete example of the output from the 'deliberation' phase, where the model is prompted to first detect errors before proceeding to refine the text.

Example Error Identification in DTG: Incorrect Translation

An engineer is using a large language model to improve automated translations. They provide the model with an original sentence and its initial translation, followed by the simple instruction: `Improve this translation.` The engineer observes that the model's outputs are unpredictable; sometimes it makes only minor stylistic edits, while other times it rewrites the sentence entirely, without a clear rationale for the changes. The goal is to make the model's refinement process more systematic and reasoned. Propose a new, two-part instruction that would guide the model to first analyze the initial translation and then correct it. Explain why structuring the instruction in this sequential way is likely to lead to more consistent and justifiable improvements.

Improving Translation Refinement Prompts

A prompt engineer wants a language model to improve an existing translation by first performing a specific analysis and then generating a corrected version, all within a single turn. Which of the following instructions best encapsulates this two-step 'analyze then correct' directive?

Learn Before

Related