Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Language-Informed Beam Search Decoding for Multilingual Machine Translation

Yilin Yang, Stefan Lee, Prasad Tadepalli

TL;DR

This work tackles the persistent problem of off-target translations in multilingual NMT during beam search decoding. It introduces Language-informed Beam Search (LiBS), which augments standard beam search with a fast Language Identification (LiD) signal to bias candidates toward the intended language, without requiring extra training data. Through analysis of beam search dynamics and extensive experiments on WMT and OPUS-100, LiBS demonstrably reduces off-target rates (e.g., from 22.9% to 7.7% on WMT and from 65.8% to 25.3% on OPUS) while yielding BLEU gains of approximately +1.1 on WMT and +0.9 on OPUS. LiBS operates as a post-hoc, model-agnostic decoding enhancement with modest speed costs (roughly 3.5–7.5x slower depending on infrastructure) and shows strong effectiveness on zero-shot translations, providing a practical path to safer, higher-quality multilingual MT deployments.

Abstract

Beam search decoding is the de-facto method for decoding auto-regressive Neural Machine Translation (NMT) models, including multilingual NMT where the target language is specified as an input. However, decoding multilingual NMT models commonly produces ``off-target'' translations -- yielding translation outputs not in the intended language. In this paper, we first conduct an error analysis of off-target translations for a strong multilingual NMT model and identify how these decodings are produced during beam search. We then propose Language-informed Beam Search (LiBS), a general decoding algorithm incorporating an off-the-shelf Language Identification (LiD) model into beam search decoding to reduce off-target translations. LiBS is an inference-time procedure that is NMT-model agnostic and does not require any additional parallel data. Results show that our proposed LiBS algorithm on average improves +1.1 BLEU and +0.9 BLEU on WMT and OPUS datasets, and reduces off-target rates from 22.9\% to 7.7\% and 65.8\% to 25.3\% respectively.

Language-Informed Beam Search Decoding for Multilingual Machine Translation

TL;DR

This work tackles the persistent problem of off-target translations in multilingual NMT during beam search decoding. It introduces Language-informed Beam Search (LiBS), which augments standard beam search with a fast Language Identification (LiD) signal to bias candidates toward the intended language, without requiring extra training data. Through analysis of beam search dynamics and extensive experiments on WMT and OPUS-100, LiBS demonstrably reduces off-target rates (e.g., from 22.9% to 7.7% on WMT and from 65.8% to 25.3% on OPUS) while yielding BLEU gains of approximately +1.1 on WMT and +0.9 on OPUS. LiBS operates as a post-hoc, model-agnostic decoding enhancement with modest speed costs (roughly 3.5–7.5x slower depending on infrastructure) and shows strong effectiveness on zero-shot translations, providing a practical path to safer, higher-quality multilingual MT deployments.

Abstract

Beam search decoding is the de-facto method for decoding auto-regressive Neural Machine Translation (NMT) models, including multilingual NMT where the target language is specified as an input. However, decoding multilingual NMT models commonly produces ``off-target'' translations -- yielding translation outputs not in the intended language. In this paper, we first conduct an error analysis of off-target translations for a strong multilingual NMT model and identify how these decodings are produced during beam search. We then propose Language-informed Beam Search (LiBS), a general decoding algorithm incorporating an off-the-shelf Language Identification (LiD) model into beam search decoding to reduce off-target translations. LiBS is an inference-time procedure that is NMT-model agnostic and does not require any additional parallel data. Results show that our proposed LiBS algorithm on average improves +1.1 BLEU and +0.9 BLEU on WMT and OPUS datasets, and reduces off-target rates from 22.9\% to 7.7\% and 65.8\% to 25.3\% respectively.
Paper Structure (25 sections, 3 figures, 12 tables)

This paper contains 25 sections, 3 figures, 12 tables.

Table of Contents

  1. Motivation
  2. Experiment Setup
  3. Dataset
  4. Model Training and Evaluation
  5. Analyzing Off-Target Occurrence During Beam Search
  6. Multilingual Beam Search Curse
  7. Off-Target Error Analysis
  8. "$\to$Source" errors
  9. "$\to$English" errors
  10. Beam Search Process Analysis
  11. Possible Solutions
  12. Language-informed Beam Search (LiBS)
  13. Design Choice and Speed Concern
  14. Experiment Results
  15. WMT Results
  16. ...and 10 more sections

Figures (3)

  • Figure 1: The sentence BLEU distribution between source and system translation from WMT Fr$\to$De "$\to$Source" errors, with an average BLEU of 85.3.
  • Figure 2: The sentence BLEU distribution between WMT Fr$\to$De "$\to$English" errors and Fr$\to$En translation with the same source. The average BLEU is 55.9.
  • Figure 3: Translation performance (BLEU and off-target rates) with different $\alpha$ values on OPUS-100 Fr$\to$De test set.