Touch Editing: A Flexible One-Time Interaction Approach for Translation

Qian Wang, Jiajun Zhang, Lemao Liu, Guoping Huang, Chengqing Zong

Abstract

Post-editing machine translation (MT) outputs is a critical but labor-intensive task, particularly on mobile devices where text entry is cumbersome. We present Touch Editing, a novel "one-time" interaction paradigm that allows users to refine translations using non-lexical gestures. By specifying error types—insertion, deletion, substitution, and reordering—without providing lexical corrections, users can guide a neural model to significantly improve translation quality. Our approach achieves gains of up to 25 BLEU points over raw MT output and outperforms existing interaction methods by substantial margins.

Introduction

While Neural Machine Translation (NMT) has reached impressive levels of fluency, it frequently suffers from errors such as under-translation (missing words), over-translation (unnecessary words), and incorrect word ordering. Traditional post-editing requires full keyboard interaction to correct these errors, creating a significant bottleneck for mobile users. Previous "lightweight" approaches like QuickEdit allow users to mark words as incorrect, but this feedback is often too sparse for the model to determine the necessary corrective action (e.g., whether to delete a word or replace it).

The Touch Editing Framework

Touch Editing addresses the sparsity of previous methods by introducing four specific non-lexical actions:

  1. Insertion (I): Indicates a missing word at a specific position.
  2. Deletion (D): Identifies a word that should be removed.
  3. Substitution (S): Marks a word that is incorrect and needs replacement.
  4. Reordering (R): Indicates that a segment of text should be moved to a different position.
Comparison of QuickEdit and Touch Editing interaction models.

To incorporate these actions, we utilize a dual-encoder Transformer architecture. A Source Encoder processes the original sentence, while a Hypothesis Encoder processes the draft translation alongside the user's touch actions. Crucially, we replace standard positional encodings in the hypothesis encoder with Action Positional Embeddings. These specialized embeddings inform the model of the intended operation at each token position, allowing the decoder to generate a refined result that respects the user's constraints without requiring the user to type the correct words manually.

Experimental Results

We evaluated Touch Editing on several standard benchmarks, including IWSLT’14 (EN-DE), WMT’14 (EN-DE), and WMT’17 (ZH-EN). To train the model, we simulated user interactions by using the Translation Edit Rate (TER) algorithm to determine the optimal set of edits required to transform a machine-generated hypothesis into a reference translation.

Quantitative performance comparison across multiple translation benchmarks.

As shown in the results table, Touch Editing consistently and significantly outperforms both the raw Transformer baseline and the QuickEdit approach. On the WMT’14 EN-DE task, Touch Editing improved the baseline from 26.69 BLEU to 50.49 BLEU—a jump of nearly 24 points. On the WMT’17 EN-ZH task, the improvement reached over 25 BLEU points. Furthermore, it outperformed QuickEdit by up to 16.64 BLEU points, demonstrating the immense value of richer, action-oriented feedback.

Discussion and Conclusion

The success of Touch Editing stems from its "One-Time" philosophy: minimizing the interaction tax by allowing users to provide all necessary feedback in a single pass. By focusing on user intent rather than requiring lexical input, the system leverages the underlying NMT model's ability to fill in the blanks based on the source context.

While current results are based on simulated interactions, user studies with a prototype mobile application suggest that gestures like swiping and tapping are highly intuitive for proofreading. Future work will focus on measuring actual productivity gains in professional translation workflows and addressing potential ambiguities in user feedback.

Summary updated

BibTeX
@inproceedings{wang2020touch,
  author = {Qian Wang and Jiajun Zhang and Lemao Liu and Guoping Huang and Chengqing Zong},
  title = {{Touch Editing: A Flexible One-Time Interaction Approach for Translation}},
  booktitle = {AACL/IJCNLP},
  year = {2020},
  doi = {10.18653/v1/2020.aacl-main.1},
}
Figure