A Simplification-Translation-Restoration Framework for Cross-Domain SMT Applications

Han-Bin Chen, Hen-Hsen Huang, Hsin-Hsi Chen, Ching-Ting TanCOLING 2012

TL;DR

This paper introduces a novel "Simplification-Translation-Restoration" (STR) framework to improve cross-domain Statistical Machine Translation (SMT). By simplifying domain-specific text segments before translation and then restoring them, the framework significantly boosts translation quality, especially in domains with limited parallel data.

What Problem Does This Paper Address?

Statistical Machine Translation (SMT) systems rely heavily on large amounts of parallel (bilingual) text data to learn translation patterns. However, when applying an SMT system to a specialized domain (e.g., medical texts) for which limited bilingual data exists, performance degrades significantly. This is primarily due to the "data sparseness" problem, where the system encounters out-of-vocabulary (OOV) words and phrases that are rare or absent in its general training data. Even if domain-specific terms are known (e.g., through a bilingual dictionary), integrating them naively can still lead to poor translations because the SMT system lacks the contextual understanding of these terms within the general model. This paper aims to bridge this domain gap by making domain-specific text more amenable to a general-purpose SMT system.

Methods

The proposed Simplification-Translation-Restoration (STR) framework consists of four key steps:

  1. Identification: The system first identifies "in-domain segments" within the source text. These are segments that are rare in the general SMT model's training data but crucial for domain-specific translation. This identification relies on bilingual in-domain resources like dictionaries or phrase pairs.
  2. Simplification: Identified in-domain segments are then simplified into more general expressions that are more likely to be present and understood by the background SMT system. This simplification can be:
  3. Terminological: Replacing rare domain-specific terms with their more common synonyms or hypernyms.
  4. Syntactic: Simplifying complex grammatical structures (like noun phrases or verb phrases) by retaining the head word and removing modifiers, making them easier for the SMT system to parse and translate.
  5. Translation: The simplified source text (with in-domain segments replaced by general ones) is then translated using a standard out-domain SMT system. Because the text is now more "general," the SMT system can better estimate contextual distributions and produce a more coherent translation of the simplified segments.
  6. Restoration: Finally, the translated simplified segments are replaced with their original, correct in-domain translations. This step uses alignment information (phrase, word, or probability-based) to ensure accurate mapping back to the original domain-specific terms and phrases.

Datasets

The experiments were conducted on an English-to-Chinese translation task in the medical domain.

  • Training Data: A general-domain SMT system (Moses) was trained on a large corpus of 6.8 million sentence pairs from the Hong Kong corpus and the UN corpus. A trigram language model was trained on the Chinese portion of this corpus and the Central News Agency corpus.
  • Test Data: The test set consisted of 1,077 parallel sentences from 18 medical summaries from the National Taiwan University Hospital (NTUH). This dataset represents a significant domain gap compared to the general training corpus, characterized by shorter sentences and a high proportion of domain-specific terminological and syntactic units, as well as OOV words.

Key Insights and Ideas

  • Bridging the Domain Gap: The core idea is that instead of trying to force a general SMT system to understand rare domain-specific terms, we can temporarily "translate" them into something the general system does understand, then translate back to the correct domain-specific term. This is a clever way to leverage existing general SMT models for specialized tasks.
  • Simplification as a Pre-processing Step: The simplification step acts as a crucial pre-processing module that makes the input text more palatable to an out-domain SMT system, effectively reducing the impact of data sparseness.
  • Restoration is Key: The final restoration step is vital for ensuring the translation is not only fluent but also accurate with respect to the original domain-specific terminology. The paper explores different granularities of alignment for this restoration.
  • Context Matters: The example of "bone lesion" translation highlights how out-domain SMT systems can mess up word order for domain-specific phrases. The STR framework, by simplifying and then restoring, aims to maintain correct contextual translation.

Key Results

The STR framework achieved significant improvements over baseline systems:

  • Baseline Moses: BLEU-4 score of 14.35, TER score of 64.911.
  • Moses+Terminology (back-off model): BLEU-4 score of 20.56, TER score of 55.712.
  • Naïve Integration (using XML markup): BLEU-4 score of 26.29, TER score of 47.683.
  • STR Framework: Achieved the highest performance with a BLEU-4 score of 35.87 and a TER score of 40.650.

The paper also reports that the STR framework outperforms baselines in terms of better translations and reorderings on general segments. The restoration methods were successful on over 93% of the test sentences.

Limitations and Future Work

  • Rule-Based Simplification: The current simplification approach is rule-based and covers only four common syntactic units (NP, VP, PP, S). The authors acknowledge that this might not be robust enough for all cases and suggest exploring statistical evidence for simplification.
  • Feedback Mechanism: The paper suggests investigating a feedback mechanism to improve the STR system. This could involve analyzing post-editing logs from human translators to refine identification, simplification, translation, and restoration steps.
  • More Robust Simplification: Developing more robust methods for simplification, potentially beyond rule-based approaches, is a key area for future work.
  • Handling Translation Errors: While successful for most cases, failures in the restoration step can occur if the SMT system makes significant translation errors on the simplified phrases, impacting word alignments.

Summary updated

BibTeX
@inproceedings{chen2012simplification,
  author = {Han-Bin Chen and Hen-Hsen Huang and Hsin-Hsi Chen and Ching-Ting Tan},
  title = {{A Simplification-Translation-Restoration Framework for Cross-Domain SMT Applications}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2012},
}
Figure