A Translation Aid System with a Stratified Lookup Interface

Takeshi Abekawa, Kyo KageuraACL 2007

Abstract

Volunteer translators often face a unique set of challenges, balancing non-native source language proficiency with a desire for high agency in the translation process. This paper introduces QRedit, a translation aid system designed to mitigate cognitive overload through a stratified lookup interface. By assigning linguistic units an "awareness score" based on composition, difficulty, speciality, and resource type, QRedit provides just-in-time alerts that highlight complex idioms and technical terms while minimizing visual clutter for common vocabulary.

Introduction

Most Computer-Aided Translation (CAT) tools are optimized for professional environments where speed and automation are paramount. However, volunteer translators—frequently native speakers of the target language with non-native proficiency in the source language—have different needs. They often spend significant time manually looking up references and prefer tools that support their decision-making rather than automating it.

A major challenge in designing such tools is the Notification Paradox: highlighting every potential reference creates unreadable clutter, while highlighting too little risks "mistranslation by oversight." QRedit addresses this by prioritizing "at-a-glance" readability through a novel stratified lookup interface.

The Stratified Lookup Mechanism

The core of QRedit is its scoring algorithm, which categorizes linguistic units into three "awareness levels." The system calculates an awareness score ($A$) for each unit by summing values across four dimensions: composition ($C$), difficulty ($D$), speciality ($S$), and resource type ($R$).

Scoring criteria and alert modes for the stratified lookup interface.

As shown in the table above, the total score determines the visual treatment of the text:

  • Level -2 (Always Emphasized): Units with a score of -2 or lower are underlined. These are high-risk units, such as complex idioms with gaps, that a translator might otherwise read literally.
  • Level -1 (Mouse-over): Units with a score of -1 become bold only when the cursor is nearby. This provides a "just-in-time" reminder for compounds or common idioms without cluttering the interface.
  • Level 0 (Passive): Units with a score of 0 or higher receive no visual alert. These are typically single, common words that the user can still click for information if needed.

For example, the phrasal verb "take off" is a compound with a gap ($C=-2$) but is considered an elementary expression ($D=1$). The resulting score of -1 ensures it is only highlighted upon mouse-over, alerting the user to its status as a multi-word unit without constant visual distraction.

System Architecture and Resources

QRedit integrates several high-quality reference sources to power its lookup engine:

  • Dictionaries: Sanseido’s Grand Concise English-Japanese Dictionary and Eijiro.
  • Technical Terminology: Lists covering 23 distinct domains.
  • Encyclopedic Knowledge: Wikipedia is used for proper names and factual context.
  • Educational Benchmarks: A curated list of 1,654 idioms from Japanese secondary school curricula is used to determine the difficulty ($D$) score.

The UI features a synchronized editor where source and target texts scroll together, maintaining the translator's workflow. The system is particularly adept at handling multi-word units (MWUs) even when they are non-contiguous (e.g., "head screwed on wrong").

Conclusion and Future Work

QRedit demonstrates a human-centric approach to translation aid, focusing on cognitive load management rather than pure automation. Qualitative feedback suggests that translators value the distinction between "must-know" alerts and "nice-to-know" references.

Future work will focus on dynamic detection of new technical terms and proper names not yet present in static dictionaries. Additionally, the authors plan to introduce user customization, allowing translators to adjust scoring weights to match their individual proficiency levels.

Summary updated

BibTeX
@inproceedings{abekawa2007translation,
  author = {Takeshi Abekawa and Kyo Kageura},
  title = {{A Translation Aid System with a Stratified Lookup Interface}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2007},
}
Figure