facebookresearch/fairseq

Biased beam search for re-translation based simultaneous translation

Open

#4,286 opened on Mar 21, 2022

View on GitHub
 (0 comments) (1 reaction) (0 assignees)Python (6,224 forks)batch import
enhancementhelp wantedneeds triage

Repository metrics

Stars
 (29,107 stars)
PR merge metrics
 (No merged PRs in 30d)

Description

🚀 Feature Request

Biased Beam Search

Motivation

Due to incremental nature of source input in streaming translation, the vanilla beam search based model produces unstable translations causing flickers in the translation. The biased beam search reduces flickers and allows stable translation so that users can follow the translation output.

Pitch

Stable Translation for Re-Translation based Simultaneous, Spoken Language Translation,

Additional context

Implements of biased beam search (Section 4.2) as described in Re-Translation Strategies For Long Form, Simultaneous, Spoken Language Translation, ICASSP 2020

Contributor guide

Research direction
Read the referenced paper Section 4.2 to understand the biased beam search algorithm. Then locate the existing beam search implementation in fairseq (likely in fairseq/search.py) and modify it to incorporate a bias term that penalizes large deviations from the previous translation hypothesis. Test the implementation on a simultaneous translation dataset to ensure stability.
Tech stack
python
Domain
machine learning
Issue type
Feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
4
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
3-5 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Fresh
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Mostly clear
Prerequisites
PythonBeam search basics
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
30