TL;DR
Synslator is a new computer-aided translation (CAT) tool that turns machine translation into a real-time collaboration. By integrating two distinct "online learning" models, it allows translation systems to adapt instantly to a human’s edits and specific domain terminology, resulting in a 13% boost in human translation efficiency.
What Problem Does This Paper Address?
Traditional Machine Translation (MT) is often a "one-and-done" process: the machine provides a draft, and the human fixes it. However, the machine usually doesn't learn from those fixes in real-time. If a translator corrects a specific technical term in sentence one, the machine might make the same mistake again in sentence ten.
While "Interactive Machine Translation" (IMT) exists, most systems struggle to balance two conflicting deployment needs:
- Private Environments: Where the system is allowed to retrain or fine-tune on user data.
- Public Cloud Environments: Where privacy or policy constraints prevent the system from "learning" (training) on user data, requiring it to use "plug-in" memories during the inference phase instead.
Synslator aims to bridge this gap with a user-friendly interface that supports both scenarios while providing "autocomplete-like" suggestions at the subword level.
Methods
The researchers developed two primary engines to handle different data-usage permissions:
- adaptive-TM-MT: Designed for scenarios where training is permitted. It uses a Transformer-based architecture that incorporates retrieved "Translation Memories" (TMs) directly into the model layers. It essentially "fine-tunes" its understanding of the domain by looking at similar past translations.
- simplified-kNN-MT: Designed for public cloud settings where training isn't authorized. Instead of retraining, it uses a $k$-nearest-neighbor approach. During translation, it searches a "datastore" of existing translations and interpolates the machine's prediction with the most likely words found in the retrieved examples.
- Subword-Prefix Decoding: This is the "secret sauce" for interactivity. Most MT models predict full words. Synslator uses a "Hit Vector" to match whatever characters the human has typed against its vocabulary. This allows the model to complete a word or sentence even if the human has only typed the first two letters (e.g., typing "fo" might trigger the model to suggest "for the purpose of...").
- GPT-based Language Model: A separate language model runs in the background to suggest fluent, stylistically consistent phrases that complement the core translation engine.
Datasets
The authors evaluated Synslator using three primary domains:
- In-house IT Domain: A proprietary Chinese-English dataset used to train and test the
adaptive-TM-MTmodel. - Open-sourced Law Domain: Used to evaluate the
simplified-kNN-MTmodel. - Open-sourced Subtitle Domain: Also used for the kNN-based evaluation to test performance on less formal text.
Key Insights and Ideas
- The "Autocomplete" for Translators: The most significant insight is the move from sentence-level interaction to subword-level interaction. By using the subword-prefix decoding algorithm, the machine acts less like a static dictionary and more like an intelligent co-writer that anticipates the translator's intent.
- Deployment Flexibility: The paper recognizes that "one size fits all" doesn't work for AI deployment. By offering both a fine-tuning model and a retrieval-based (kNN) model, they ensure the tool is useful whether the user is a high-security legal firm or a casual cloud user.
- Confidence Highlighting: Synslator uses color-coding to show the model's confidence. If the machine is highly certain about its next few words (probability > 0.6), it highlights them, allowing the translator to simply hit "TAB" to accept the suggestion, drastically reducing keystrokes.
Key Results
- Efficiency Gains: In real-world testing with ten professional translators over three weeks, Synslator improved post-editing efficiency by 13% compared to traditional static post-editing.
- Translation Quality: The
adaptive-TM-MTmodel achieved a BLEU score of 29.05 (up from 25.47 for vanilla NMT) in the IT domain. - Prediction Accuracy: The inclusion of the GPT-based language model and the "next 3-best" suggestions significantly improved "N-gram Accuracy," meaning the machine's guesses were much more likely to match what the human actually wanted to write.
Limitations and Future Work
While the paper demonstrates clear efficiency gains, a few areas remain for exploration:
- Computational Overhead: The $k$-nearest-neighbor (kNN) approach, while effective for privacy, can be computationally expensive during inference because it requires searching a large datastore for every word predicted. The authors "simplified" this by limiting the search to 16 neighbors, but scaling this to massive datasets remains a challenge.
- Language Support: The current evaluation focuses on Chinese-to-English. The effectiveness of subword-prefix decoding might vary in morphologically rich languages (like Finnish or Turkish) where word endings change frequently.
- User Learning Curve: While the tool is "user-friendly," the paper notes that efficiency gains improved over the three-week test period, suggesting that translators need time to learn how to best "dance" with the interactive suggestions.
Summary updated