Abstract
Domain-specific neural machine translation (NMT) systems are socially significant with the potential to help make information accessible to a diverse set of users in multilingual societies. Such NMT systems should be lexically constrained and draw from domain-specific dictionaries. Dictionaries could present multiple candidate translations for a source word/phrase due to the polysemous nature of words. The onus is then on the NMT model to choose the contextually most appropriate candidate. Prior work has largely ignored this problem and focused on the single candidate constraint setting wherein the target word or phrase is replaced by a single constraint. In this work, we present DictDis, a lexically constrained NMT system that disambiguates between multiple candidate translations derived from dictionaries. We achieve this by augmenting training data with multiple dictionary candidates to actively encourage disambiguation during training by implicitly aligning multiple candidate constraints. We demonstrate the utility of DictDis via extensive experiments on English-Hindi, English-German, and English-French datasets across a variety of domains including regulatory, finance, engineering, health and standard benchmark test datasets. In comparison with existing approaches for lexically constrained and unconstrained NMT, we demonstrate superior performance for the copy constraint and disambiguation-related measures on all domains, while also obtaining improved fluency of up to 2-3 BLEU points on some domains.
Introduction
In domain-specific translation—such as engineering manuals or financial reports—terminological accuracy is critical. Standard Neural Machine Translation (NMT) models often struggle with rare technical terms. To mitigate this, researchers use "lexical constraints" to force the model to use specific words from a domain dictionary.
However, language is inherently polysemous. A single English word like "speed" might have several different translations in Hindi depending on whether the context is physics, mechanics, or general driving. Most existing constrained NMT systems assume a "one-to-one" mapping, requiring a human or a pre-processing script to select the correct constraint beforehand. If the wrong constraint is chosen, the translation fails. DictDis addresses this "one-to-many" problem by allowing the NMT model to choose the most contextually appropriate translation from a list of multiple dictionary candidates.
Methodology
DictDis (Dictionary Disambiguation) is an end-to-end "copy-and-disambiguate" framework.
Input Augmentation
Rather than feeding only the source sentence, DictDis appends a list of constraints to the input. Special tokens create a hierarchy: <sep> separates different source phrases, while <isep> (intra-separator) separates multiple candidate translations for the same phrase.
Architecture
The system extends a standard Transformer with two critical components:
- Pointer Network: A "copy" mechanism that allows the model to point at words in the input constraints and pull them directly into the output.
- Disambiguation Network: This component calculates a score for each candidate translation by comparing the candidate's embedding with the source sentence context.
- Gating Mechanism: A learnable gate ($g_t$) decides at each time step whether the model should generate a word from its general vocabulary or utilize the disambiguation network to select a constraint.
Soft Ingestion
Unlike "hard" constraint methods that can disrupt grammatical flow, DictDis employs "soft" ingestion. The model is trained to prefer constraints but retains the flexibility to ignore them if they compromise sentence fluency.
Experimental Setup
The framework was evaluated on English-Hindi, English-German, and English-French language pairs.
- Training: Models were trained on large-scale general corpora, such as Samanantar (8.4M sentences for En-Hi).
- Benchmarks: The authors released a new 4K sentence English-Hindi test set manually curated across specialized domains: Regulatory, Aerospace, and Banking. Standard benchmarks like Flores-101, WAT2021, and WMT were also used.
Results and Discussion
DictDis demonstrates significant improvements in both fluency and terminological accuracy.
Performance Metrics
DictDis improved BLEU scores by up to 2-3 points over unconstrained models in certain domains and significantly outperformed previous constrained methods like VDBA and Leca. Using the Copy Success Rate (CSR) metric, DictDis proved superior at picking the correct constraint from a list, especially in high-polysemy domains like Aerospace.
Key Insights
- Integrated Disambiguation: By showing the model multiple options during training, it learns the relationship between context and specific dictionary senses, removing the need for separate pre-processing.
- Controllable Fluency: A user-controllable parameter ($\alpha$) allows for balancing technical accuracy with natural flow.
- Robustness to Polysemy: While baseline models' performance degrades as the number of dictionary candidates increases, DictDis remains robust even with 5+ possible translations.
Conclusion
DictDis represents a significant step forward in making NMT systems more practical for domain-specific applications. While it introduces some computational overhead and depends on dictionary quality, its ability to handle polysemy end-to-end provides a more flexible and accurate solution for technical translation. Future work will focus on improving efficiency and exploring real-time dictionary retrieval.
Summary updated