Bridging Background Knowledge Gaps in Translation with Automatic Explicitation

HyoJung Han, Jordan L. Boyd-Graber, Marine CarpuatEMNLP 2023 · DOI

Introduction

In the field of machine translation, the goal has traditionally been to achieve the most accurate literal mapping between source and target languages. However, literal translations often fail to communicate the full intended meaning when the target audience lacks the cultural or historical context inherent to the source text. For example, a French news article mentioning "Dominique de Villepin" requires no explanation for a French reader, but an English reader may benefit significantly if the text is rendered as "the former French Prime Minister Dominique de Villepin."

Professional human translators bridge these gaps using a technique called explicitation—the practice of adding explanatory information to make implicit context explicit. Despite its importance, explicitation has been largely ignored in Natural Language Processing (NLP) because standard evaluation metrics like BLEU penalize additions as "extraneous" noise. This paper introduces WIKIEXPL, a new dataset of human-annotated explicitations, and proposes an automated pipeline to detect context gaps and generate grounded explanations using structured knowledge bases.

The Problem: The Literalism Trap

Literal translation is often insufficient for cross-cultural communication. Current machine translation (MT) models are trained to stay as close to the source text as possible. This creates a "BLEU paradox": because metrics reward word-for-word overlap with a reference, they actively discourage the very techniques human experts use to ensure readability and comprehension. Furthermore, there is a severe lack of labeled data to train models to recognize when a reader might be confused by a culturally specific reference.

Methodology: Automating Explicitation

The researchers developed a three-stage pipeline to identify and fill these knowledge gaps:

1. Detection

The authors created an algorithmic method to identify naturally occurring explicitations in "noisy" parallel text, such as Wikipedia. By analyzing unaligned segments in the target language that appear near named entities, they were able to extract real-world examples of how humans provide context during translation.

2. Decision: Measuring Cultural Distance

To automate the decision of when to add information, the study defined a metric for "cultural distance." Using Wikidata, they measured the graph topology (how many "hops" an entity is from a specific country) and compared the popularity of Wikipedia pages across different languages. If an entity is highly prominent in Poland but obscure in the English-speaking world, the system flags it as a candidate for explicitation.

3. Generation

The system tests three levels of context grounded in Wikidata and Wikipedia:

  • SHORT: 1-2 word appositives (e.g., "Sambre river").
  • MID: Phrases or parentheticals (e.g., "Sambre, a river in France and Belgium").
  • LONG: 1-3 sentence footnotes providing deeper historical or situational context.

Evaluation via Multilingual Question Answering

To move beyond subjective measures of "usefulness," the authors employed a Multilingual Question Answering (QA) framework using the XQB (Cross-lingual Quizbowl) dataset. This dataset consists of pyramidal trivia questions where clues move from obscure to well-known.

The logic of the evaluation is that if adding a snippet of information helps an English-speaking AI answer a question about a foreign entity correctly—but does not change the answer for a native-speaking AI who already possesses that background knowledge—the explicitation is objectively useful for bridging a specific gap.

Key Results and Insights

  • High Precision: Human evaluators verified that 70% of the system's automated explicitation decisions were valid and helpful.
  • Significant QA Boost: In Polish-to-English tasks, adding "MID" length explicitations improved "Expected Wins" (a metric measuring both speed and accuracy) by 47.4%.
  • The "Mid" Length Sweet Spot: While "LONG" footnotes provided the most information and received high human quality ratings (0.95), "MID" length parentheticals were the most effective for QA systems, offering a balance of context and seamless integration.
  • Validation of the Gap: The system showed a much higher increase in accuracy for the target language (English) than the source languages, proving the additions were specifically addressing a knowledge deficit rather than simply providing general hints.

Limitations and Future Work

While the WIKIEXPL pipeline represents a significant step forward, the authors identify several areas for future research:

  • Beyond Named Entities: The current system focuses on proper nouns (people, places, organizations). Future iterations could address culturally specific concepts or historical events that are not named entities.
  • Nuanced Cultural Mapping: The current model maps languages to single "majority" countries (e.g., French to France), which overlooks the diversity of the broader Francophone or Hispanophone worlds.
  • Recursive Gaps: Occasionally, an explanation requires its own context (e.g., explaining "J'accuse...!" by mentioning the "Dreyfus affair," which the reader may also not know).
  • LLM Integration: The authors suggest that Large Language Models (LLMs) could be used to better blend these generated explicitations into the source text to improve overall narrative flow and fluency.

Conclusion

By quantifying "cultural boundedness" and providing a task-based evaluation for explicitation, this research challenges the industry's reliance on literal translation. The WIKIEXPL dataset and the associated pipeline demonstrate that providing targeted, grounded context can significantly improve the utility of translated text for global audiences.

Summary updated

BibTeX
@inproceedings{han2023bridging,
  author = {HyoJung Han and Jordan L. Boyd-Graber and Marine Carpuat},
  title = {{Bridging Background Knowledge Gaps in Translation with Automatic Explicitation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2023},
  doi = {10.18653/v1/2023.emnlp-main.603},
}
Figure