Abstract
Despite the fact that multilingual agreement (MA) has shown its importance for multilingual neural machine translation (MNMT), current methodologies in the field have two shortages: (i) require parallel data between multiple language pairs, which is not always realistic and (ii) optimize the agreement in an ambiguous direction, which hampers the translation performance. We present Bidirectional Multilingual Agreement via Switched Back-translation (BMA-SBT), a novel and universal multilingual agreement framework for fine-tuning pre-trained MNMT models, which (i) exempts the need for aforementioned parallel data by using a novel method called switched BT that creates synthetic text written in another source language using the translation target and (ii) optimizes the agreement bidirectionally with the Kullback-Leibler Divergence loss. Experiments indicate that BMA-SBT clearly improves the strong baselines on the task of MNMT with three benchmarks: TED Talks, News, and Europarl. In-depth analyzes indicate that BMA-SBT brings additive improvements to the conventional BT method.
Introduction
Multilingual Neural Machine Translation (MNMT) aims to translate between multiple language pairs using a single model. A key technique to enhance these models is Multilingual Agreement (MA), which encourages the model to generate consistent target translations regardless of the source language, provided the semantic meaning is preserved. For instance, a Chinese source sentence and its German equivalent should ideally result in the same English output.
However, traditional MA faces two significant hurdles. First, it typically requires multi-way parallel data—aligned sentences across multiple source languages—which is scarce in practice. Second, existing methods often rely on word-level code-switching, leading to ambiguous optimization signals that fail to fully exploit the linguistic nuances of different languages. This paper introduces BMA-SBT to overcome these limitations by synthesizing auxiliary data and employing a bidirectional alignment strategy.
Methodology: BMA-SBT
The proposed Bidirectional Multilingual Agreement via Switched Back-translation (BMA-SBT) framework introduces two core innovations to refine pre-trained MNMT models during fine-tuning.
Switched Back-Translation (SBT)
To bypass the need for authentic multi-way parallel data, BMA-SBT uses the target text to generate synthetic auxiliary source sentences. By back-translating the target (e.g., English) into an auxiliary language (e.g., Japanese), the model creates a synthetic pair of source sentences that share the same semantic meaning. This allows the model to learn multilingual agreement using only standard bilingual data.
Bidirectional KL Divergence Loss
Rather than simple code-switching, BMA-SBT aligns the output probability distributions of the original source and the synthetic auxiliary source. By applying a bidirectional Kullback-Leibler (KL) divergence loss, the model is forced to reconcile its internal representations from both languages simultaneously. This ensures that the model's understanding of one language directly informs and improves its understanding of the other.
Experimental Setup
The framework was evaluated on three major MNMT benchmarks for X → English translation:
- TED Talks (IWSLT15): Evaluated at both sentence and document levels.
- News Benchmark: Trained on News Commentary v11 and tested on
newstestsets. - Europarl-v7: Tested across 9 source languages to demonstrate broad applicability.
Results and Discussion
BMA-SBT consistently outperformed strong baseline models, including M2M-100 and mBART, across all tested datasets.
Performance on TED Talks
On the TED Talks benchmark, BMA-SBT achieved a state-of-the-art average document-level BLEU score of 40.85, a significant improvement over the Baseline + standard Back-Translation (BT) score of 39.65. The results demonstrate that synthetic data generated via SBT is highly effective for teaching the model cross-lingual consistency.
Key Insights
- Complementary Strengths: Different languages often have different grammatical constraints. For example, while Chinese might be ambiguous regarding verb tense, an auxiliary language like German can provide explicit tense markers. BMA-SBT allows the model to use these explicit cues to resolve ambiguities in the primary source language.
- Additive Gains: BMA-SBT provides improvements that are additive to standard BT. This suggests that the multilingual agreement signal captures different linguistic information than simple data augmentation.
- Document-Level Coherence: Beyond BLEU scores, the model showed improved document-level coherence, as evidenced by higher BlonDe scores. It successfully resolved complex plural/singular and tense ambiguities that often challenge commercial translation systems like Google Translate.
Conclusion
BMA-SBT presents a robust and universal framework for improving MNMT through bidirectional multilingual agreement. By leveraging switched back-translation, it removes the dependency on rare multi-way parallel data. Future work may explore the use of multiple auxiliary languages in an ensemble approach to further enhance translation accuracy and robustness across diverse linguistic contexts.
Summary updated