A New Input Method for Human Translators: Integrating Machine Translation Effectively and Imperceptibly

Guoping Huang, Jiajun Zhang, Yu Zhou, Chengqing ZongIJCAI 2015 · OpenReview

Abstract

Computer-aided translation (CAT) system is the most popular tool which helps human translators perform language translation efficiently. To further improve the efficiency, there is an increasing interest in applying the machine translation (MT) technology to upgrade CAT. Post-editing is a standard approach: human translators generate the translation by correcting MT outputs. In this paper, we propose a novel approach deeply integrating MT into CAT systems: a well-designed input method which makes full use of the knowledge adopted by MT systems, such as translation rules, decoding hypotheses and n-best translation lists. Our proposed approach allows human translators to focus on choosing better translation results with less time rather than just complete translation themselves. The extensive experiments demonstrate that our method saves more than 14% time and over 33% keystrokes, and it improves the translation quality as well by more than 3 absolute BLEU scores compared with the strong baseline, i.e., post-editing using Google Pinyin.

Introduction

In Computer-Aided Translation (CAT), the standard workflow is Post-Editing (PE), where a human translator corrects a machine-generated draft. However, this method faces two significant challenges. First, low-quality machine translation (MT) output can be more mentally taxing to fix than translating from scratch. Second, for languages like Chinese that require an Input Method Editor (IME) to convert phonetic Pinyin into characters, standard IMEs are context-blind. They do not leverage the source sentence context, forcing translators to manually select from long lists of irrelevant candidates.

This paper, presented at IJCAI 2015, introduces CoCat, a novel input method that "imperceptibly" integrates MT knowledge directly into the typing process. By embedding MT translation rules and hypotheses into the keyboard itself, CoCat allows translators to benefit from machine intelligence without the friction of traditional post-editing.

Methodology

CoCat enhances the standard IME by incorporating data from a Statistical Machine Translation (SMT) decoder. The system consists of two primary components:

1. Phrase Generation Model

Standard Pinyin IMEs rank candidates based on general word frequency. CoCat employs a log-linear model that re-ranks these candidates using three "MT-aware" features:

  • Translation Rules: Whether a candidate phrase exists in the MT system's rules for the specific source sentence.
  • Decoding Hypotheses: Whether the phrase was considered by the MT engine during its internal search.
  • N-best List: Whether the phrase appears in the top $N$ full-sentence translations generated by the machine.
Overview of the CoCat system architecture, showing how SMT decoder knowledge (rules, hypotheses, and n-best lists) is integrated into the input method interface.

2. N-gram Prediction Model

CoCat provides advanced predictive text by matching the translator's current input against the MT's n-best translation list. It suggests the most likely suffixes as single completions, allowing the translator to accept entire multi-word phrases with a single keystroke.

Experimental Setup

The underlying SMT system was trained on 10 million parallel English-Chinese news sentence pairs. Evaluation was conducted using 160 sentences from China Daily news with 12 professional translators. The study compared four scenarios: standard Google Pinyin, CoCat, Post-editing with Google Pinyin, and Post-editing with CoCat.

Results and Discussion

CoCat demonstrated significant improvements over the strong baseline of Google Pinyin:

  • Efficiency: CoCat saved 14% more time and 33% more keystrokes than the standard post-editing workflow.
  • Keystroke Reduction: One of the most effective features is CoCat's ability to resolve extreme abbreviations (acronyms of Pinyin initials) based on the MT hypothesis space. For example, typing "zgklgg" can immediately surface "China mulls reform" (Zhōngguó kǎolǜ gǎigé).
  • Quality: Translations produced with CoCat scored 3 absolute BLEU points higher than traditional post-editing, suggesting that reduced mechanical effort allows translators to focus more on linguistic nuance.
Example of keystroke efficiency in CoCat: using abbreviations and n-gram predictions to complete a phrase in just 5 keystrokes.

Conclusion

By moving machine translation from a separate suggestion box into the translator's keyboard, CoCat provides "ambient" assistance that is both effective and unobtrusive. While the current study focused on SMT, future work could adapt this framework to Neural Machine Translation (NMT) architectures, leveraging attention maps and neural hidden states to further refine candidate ranking and prediction.

Summary updated

Abstract

Computer-aided translation (CAT) system is the most popular tool which helps human translators perform language translation efficiently. To further improve the efficiency, there is an increasing interest in applying the machine translation (MT) technology to upgrade CAT. Post-editing is a standard approach: human translators generate the translation by correcting MT outputs. In this paper, we propose a novel approach deeply integrating MT into CAT systems: a well-designed input method which makes full use of the knowledge adopted by MT systems, such as translation rules, decoding hypotheses and n-best translation lists. Our proposed approach allows human translators to focus on choosing better translation results with less time rather than just complete translation themselves. The extensive experiments demonstrate that our method saves more than 14% time and over 33% keystrokes, and it improves the translation quality as well by more than 3 absolute BLEU scores compared with the strong baseline, i.e., post-editing using Google Pinyin.
BibTeX
@inproceedings{huang2015new,
  author = {Guoping Huang and Jiajun Zhang and Yu Zhou and Chengqing Zong},
  title = {{A New Input Method for Human Translators: Integrating Machine Translation Effectively and Imperceptibly}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
  url = {https://openreview.net/forum?id=HJWmgSMu-r},
}
Figure