Synslator: An Interactive Machine Translation Tool with Online Learning

Jiayi Wang, Ke Wang, Fengming Zhou, Chengyu Wang, Zhiyong Fu, Zeyu Feng, Yu Zhao, Yuqi Zhang

Abstract

Interactive machine translation (IMT) has emerged as a progression of the computer-aided translation paradigm, where the machine translation system and the human translator collaborate to produce high-quality translations. This paper introduces Synslator, a user-friendly computer-aided translation (CAT) tool that not only supports IMT, but is adept at online learning with real-time translation memories. To accommodate various deployment environments for CAT services, Synslator integrates two different neural translation models to handle translation memories for online learning. Additionally, the system employs a language model to enhance the fluency of translations in an interactive mode. In evaluation, we have confirmed the effectiveness of online learning through the translation models, and have observed a 13% increase in post-editing efficiency with the interactive functionalities of Synslator.

Introduction

Traditional Machine Translation (MT) workflows often follow a static "translate-then-edit" pattern. A significant drawback of this approach is the system's inability to learn from human corrections in real-time. If a translator corrects a domain-specific term in one sentence, the MT system typically repeats the same error in subsequent segments.

Synslator addresses this by bridging the gap between human expertise and machine efficiency. It provides a user-friendly interface that supports both private environments (where model fine-tuning is permitted) and public cloud environments (where privacy constraints necessitate retrieval-based learning).

The Synslator user interface, showing (a) project settings for MT engine selection and (b) the interactive translation workbench.

System Architecture and Online Learning

Synslator integrates two primary engines to facilitate online learning across different deployment scenarios:

1. Adaptive-TM-MT

Designed for environments where retraining is feasible, this engine utilizes a Transformer-based architecture. It incorporates retrieved Translation Memories (TMs) directly into the model layers, effectively "fine-tuning" its output based on similar historical translations.

2. Simplified-kNN-MT

For public cloud settings where training on user data is restricted, Synslator employs a $k$-nearest-neighbor (kNN) approach. This method does not require retraining. Instead, it queries a datastore of existing translations during inference and interpolates the NMT model's probability distribution with the distribution of target tokens found in the retrieved examples.

The workflow of the simplified-kNN-MT engine, illustrating the retrieval of examples and distribution interpolation.

Interactive Functionalities

Beyond sentence-level translation, Synslator introduces "Subword-Prefix Decoding." Unlike standard models that predict full words, Synslator uses a "Hit Vector" to match user input against its vocabulary at the character level. This enables an autocomplete-like experience where the model suggests completions as the translator types.

To further enhance fluency, a separate GPT-based language model runs in parallel, suggesting stylistically consistent phrases. The tool also employs color-coded confidence highlighting; when the model's prediction probability exceeds 0.6, the suggestion is highlighted, allowing the user to accept it instantly via the "TAB" key.

Evaluation and Results

The system was evaluated across three domains: IT (proprietary), Law (open-source), and Subtitles (open-source).

Key findings include:

  • Efficiency: A three-week study with ten professional translators showed a 13% increase in post-editing efficiency compared to traditional methods.
  • Quality: The adaptive-TM-MT model improved BLEU scores significantly in the IT domain (from 25.47 to 29.05).
  • Interactivity: The combination of subword-prefix decoding and GPT-based suggestions improved N-gram accuracy, aligning machine predictions more closely with human intent.

Conclusion and Future Work

Synslator demonstrates that real-time collaboration between humans and MT systems can significantly boost productivity. While the kNN approach offers privacy benefits, its computational overhead remains a challenge for scaling. Future work will focus on optimizing retrieval speeds and expanding subword-prefix decoding to morphologically rich languages.

Summary updated

Abstract

Interactive machine translation (IMT) has emerged as a progression of the computer-aided translation paradigm, where the machine translation system and the human translator collaborate to produce high-quality translations. This paper introduces Synslator, a user-friendly computer-aided translation (CAT) tool that not only supports IMT, but is adept at online learning with real-time translation memories. To accommodate various deployment environments for CAT services, Synslator integrates two different neural translation models to handle translation memories for online learning. Additionally, the system employs a language model to enhance the fluency of translations in an interactive mode. In evaluation, we have confirmed the effectiveness of online learning through the translation models, and have observed a 13% increase in post-editing efficiency with the interactive functionalities of Synslator. A tutorial video is available at:https://youtu.be/K0vRsb2lTt8.
BibTeX
@inproceedings{wang2024synslator,
  author = {Jiayi Wang and Ke Wang and Fengming Zhou and Chengyu Wang and Zhiyong Fu and Zeyu Feng and Yu Zhao and Yuqi Zhang},
  title = {{Synslator: An Interactive Machine Translation Tool with Online Learning}},
  booktitle = {The Web Conference},
  year = {2024},
  doi = {10.1145/3589335.3651240},
}
Figure