Guided decoding with "enhanced" keyterms
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 39
- Forks
- 7
- Avg merge
- 1d 9h
- Merged PRs (30d)
- 5
Description
By using some combination of alignments, LLM's and Levenshtein distance, we should be able to determine the "proper form" of keywords in the target sentence, and tell NLLB-200 to include that specific surface form of the word. This could be done by doing the following:
- Determining the matching of proper names for each verse from source to target
- Giving this information to an LLM to say "when this word is used here in this context, the surface form looks like this" - give it a lot of examples. Put them in the context window
- Ask the LLM: "For this new name in this context, what should the surface form be?"
- Take those surface forms and feed them to the LLM to "guide" the decoding.
Implementing this relies upon the successful implementation of #178.
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by reading issue #178, which this proposal depends on, then locate the NLLB-200 decoding entry point. Define how alignment, LLM-produced surface forms, and Levenshtein matching connect; done means proper-name forms can guide decoding as described, with the behavior and coverage verified.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100