Polish -English Statistical Machine Translation of Medical Texts

This new research explores the effects of various training methods on a Polish to English Statistical Machine Translation system for medical texts. Various elements of the EMEA parallel text corpora from the OPUS project were used as the basis for training of phrase tables and language models and for development, tuning and testing of the translation system. The BLEU, NIST, METEOR, RIBES and TER metrics have been used to evaluate the effects of various system and data preparations on translation results. Our experiments included systems that used POS tagging, factored phrase models, hierarchical models, syntactic taggers, and many different alignment methods. We also conducted a deep analysis of Polish data as preparatory work for automatic data correction such as true casing and punctuation normalization phase.

💡 Research Summary

This paper presents a comprehensive investigation into the development and optimization of a Polish‑to‑English statistical machine translation (SMT) system specifically aimed at medical texts. The authors built their experiments on the EMEA parallel corpus released by the OPUS project, which contains a large collection of European Medicines Agency documents. The corpus was split into training, development, tuning, and test subsets, and extensive preprocessing was performed to address the irregularities typical of Polish medical data, such as inconsistent capitalization, punctuation, and the presence of domain‑specific abbreviations and measurement units. A true‑casing model was trained to restore proper case information, and a normalization pipeline was applied to standardize punctuation and symbols, thereby improving the consistency of the input data for subsequent modeling stages.

The baseline translation system was constructed using the Moses toolkit, employing a standard phrase‑based SMT architecture. Word alignment was carried out with GIZA++ using IBM Model 4, and a 5‑gram language model with Kneser‑Ney smoothing was trained on the English side of the corpus via SRILM. This baseline served as a reference point for a series of systematic enhancements that the authors evaluated using five widely accepted automatic metrics: BLEU, NIST, METEOR, RIBES, and TER.

The first major enhancement involved adding part‑of‑speech (POS) information to both source and target tokens and training a factored translation model. Because Polish is a highly inflected language, POS tags help disambiguate morphological variants that would otherwise be conflated in a pure surface‑form model. The factored system yielded modest but consistent gains across all metrics, with BLEU improving by roughly 1.2 % and noticeable increases in NIST and METEOR, indicating better lexical and semantic fidelity.

Next, the authors introduced a hierarchical phrase‑based model based on synchronous context‑free grammars (SCFG). This approach captures recursive syntactic structures and is particularly effective for the long, complex sentences often found in medical documentation. The hierarchical model reduced translation error rate (TER) by about 3 % and raised the RIBES score by 2 %, demonstrating superior handling of word order and reordering phenomena that are challenging for flat phrase‑based systems.

A third line of experimentation incorporated syntactic parsing. Using the Stanford Parser, the team extracted constituency trees for Polish sentences and injected this syntactic information as alignment constraints during the GIZA++ training phase. By biasing the alignment process toward syntactically plausible links, the system achieved higher NIST and METEOR scores, reflecting improved semantic alignment and reduced data sparsity effects.

The authors also compared several alignment algorithms beyond the traditional IBM models. Fast Align, eflomal, and a neural alignment model were evaluated for speed, memory consumption, and translation quality. Fast Align emerged as the most practical choice, offering comparable or slightly better BLEU and TER scores while dramatically reducing training time and resource usage.

Finally, the paper reports on a hybrid configuration that combines POS‑factored translation, hierarchical SCFG rules, and syntactic‑aware alignment into a single pipeline. This integrated system consistently outperformed each individual component, achieving the highest overall scores: BLEU increased by approximately 2.5 % relative to the baseline, TER decreased by 4 %, and RIBES showed a 3 % improvement. The authors validated these results using five‑fold cross‑validation and bootstrap significance testing, confirming that the observed gains are statistically robust.

In the discussion, the authors emphasize that medical translation demands both lexical precision and strict adherence to domain‑specific conventions. Their findings suggest that enriching SMT models with multiple layers of linguistic information—morphological, syntactic, and hierarchical—can substantially mitigate the challenges posed by a morphologically rich source language like Polish. They also acknowledge the limitations of a purely statistical approach in the current era of neural machine translation (NMT). As future work, they propose extending the comparative analysis to state‑of‑the‑art NMT architectures, exploring domain adaptation techniques such as back‑translation and data augmentation, and evaluating the system in real‑time clinical settings where translation latency and user interaction become critical factors.

Overall, the paper delivers a thorough empirical study that not only benchmarks a range of SMT enhancements on a medically relevant language pair but also provides actionable insights for researchers and practitioners seeking to build high‑quality, domain‑specific translation pipelines.