In the simplified Deliberate-then-Generate (DTG) method, where the input translation is randomly sampled and not generated by the LLM, a default error type is assigned. For instance, the error can be explicitly labeled as "Incorrect Translation" to direct the model's focus toward generating a new output rather than on error identification.

Google

To focus an LLM's task on deliberation and reduce the burden of error identification, a specific technique can be used. This involves substituting the actual translation with a random one and assigning a default error type, which simplifies the prompt and directs the model's processing away from error analysis.

Technique of Using Random Translation and Default Error Type

Reference of Foundations of Large Language Models Course

An example prompt for the simplified Deliberate-then-Generate (DTG) method, which focuses only on the model's deliberation capabilities, is structured to present a source sentence and a placeholder for a translation. The prompt is formatted as follows:

```
Given the Chinese sentence:

The English translation is:
```

Learn Before

Related