Skip to content

Conversation

SimBe195
Copy link
Collaborator

Implement a lexicon-free label-synchronous search algorithm (based on SearchAlgorithmV2). This is mostly similar to LexiconfreeTimesyncBeamSearch with the most notable differences being:

  • Instead of a blank token, this search needs a sentence-end token.
  • There is a distinction between terminated hypotheses and active hypotheses, depending on whether they have produced sentence-end yet.
  • Since hypotheses may be terminated at different lengths, length normalization is introduced for score comparison.
  • The search terminates when all hypotheses in the beam are terminated or when an upper limit of decoding steps has been reached. This limit depends on the number of input features that have been fed.

Simon Berger and others added 30 commits February 19, 2025 19:10
@SimBe195 SimBe195 requested review from curufinwe and larissakl May 12, 2025 14:04
@larissakl
Copy link
Contributor

I have just noticed that you have not registered this algorithm as a SearchTypeV2 in Search/Module yet. I would say that we should implement this directly when introducing a new search algorithm.

@SimBe195 SimBe195 requested a review from larissakl May 28, 2025 15:58
Base automatically changed from cleanup_inactive_contexts to master June 6, 2025 12:14
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants