MiSS: An Assistant for Multi-Style Simultaneous Translation

Zuchao Li, Kevin Parnow, Masao Utiyama, Eiichiro Sumita, Hai ZhaoEMNLP 2021 · DOI

Introduction

Neural Machine Translation (NMT) has achieved significant milestones in recent years, yet it often remains a "black box" for users who do not speak the target language. This lack of transparency creates a trust gap: users cannot easily verify the accuracy of a translation or adjust its tone to suit specific contexts. Furthermore, traditional NMT systems typically wait for a full sentence to be completed before translating, which introduces significant latency in live conversational or drafting scenarios.

To address these challenges, we present MiSS (Multi-Style Simultaneous Translation), a comprehensive web-based assistant designed to transform translation from a passive output process into a collaborative, interactive experience. MiSS integrates simultaneous translation, style control, and real-time grammatical error correction (GEC) to provide a transparent loop of automated quality feedback.

The MiSS interface provides a comprehensive dashboard for simultaneous translation, style selection, and real-time feedback.

System Architecture and Methods

MiSS is built upon a Transformer-based architecture, enhanced by several specialized modules that create a holistic translation environment:

1. Simultaneous Translation via wait-k

To minimize latency, MiSS employs a wait-k strategy. Unlike standard models that require the entire source sentence, the wait-k decoder begins generating translation after reading only $k$ words of input. This allows the system to provide real-time updates as the user types, making it ideal for live interactions.

2. Multi-Style Control (CTRL-NMT)

Rather than deploying separate models for different domains, MiSS utilizes "control codes" inspired by the CTRL language model. By prepending natural language tags such as "Formal" (for business) or "Oral" (for casual conversation) to the input, a single model learns to steer its output style while sharing underlying linguistic knowledge across domains.

3. Grammatical Error Correction (GEC)

NMT models are notoriously sensitive to source-text errors. MiSS incorporates a tag-based GEC model (similar to GECToR) that predicts "g-transformations" (e.g., $KEEP`, `$REPLACE, or $DELETE). This pre-emptive fix acts as a shield, correcting user typos or grammatical slips before they reach the translation engine, thus preventing "garbage in, garbage out" scenarios.

Interactive Feedback and Quality Assurance

A core innovation of MiSS is the "Back-translation Feedback" loop. To solve the trust problem, the system automatically translates the generated target text back into the source language. It then calculates quantitative confidence scores—BERTScore and BT-BLEU—to show the user how well the original meaning was preserved.

The system provides target translations alongside back-translation scores and specific revision suggestions to improve source input.

Key Results and Insights

Evaluation across English (EN), Chinese (ZH), and Japanese (JA) corpora demonstrated the system's efficacy:

  • Style Fidelity: The CTRL-NMT model achieved performance comparable to specialized baselines. For EN→ZH, the formal style reached 37.5 BLEU, nearly matching the 37.6 BLEU of a model trained exclusively on formal data.
  • Latency vs. Accuracy: While simultaneous translation ($k=3$) involves a slight dip in accuracy (e.g., ~31.1 vs ~35.4 BLEU for EN→ZH), the reduction in latency significantly improves the user experience in interactive settings.
  • GEC Performance: The integration of pre-trained language models like XLNet and BERT allowed the English GEC module to reach an $F_{0.5}$ of 65.47, a substantial improvement over the 54.22 baseline.

Conclusion and Future Work

MiSS moves beyond simple text-in, text-out translation by providing a malleable, transparent interface. By treating style as a parameter and incorporating GEC as a pre-processing step, the system offers a more robust tool for non-native speakers. Future developments will focus on document-level contextual awareness, expanding the range of available styles (e.g., "Academic"), and utilizing crowdsourced interaction data from the MiSS platform to further fine-tune model performance.

Summary updated

BibTeX
@inproceedings{li2021miss,
  author = {Zuchao Li and Kevin Parnow and Masao Utiyama and Eiichiro Sumita and Hai Zhao},
  title = {{MiSS: An Assistant for Multi-Style Simultaneous Translation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2021},
  doi = {10.18653/v1/2021.emnlp-demo.1},
}
Figure