In a research paper published on October 24, 2023, a group of researchers from the Shanghai Jiao Tong University and Tencent AI Lab highlighted the importance of word-level auto-completion (WLAC) in computer-assisted translation (CAT) and proposed an effective approach to enhance its performance.
Traditionally, WLAC has played a “crucial role in computer-assisted translation”, serving as an available tool for translators aiming to improve efficiency. As the researchers noted, “effective auto-completion has the potential to reduce keystrokes by at least 60% during the translation process.”
However, they identified a critical flaw in the existing criterion for determining a good auto-completion suggestion. This criterion, based on the maximum likelihood estimation (MLE) of the target word given the source context, often led to suboptimal results.
The primary issue was the impracticality of relying on the reference translation during prediction, as it was not available in real time. This challenge prompted the researchers to question the essence of WLAC and explore alternatives that could address this fundamental issue.
In response to that, they introduced a novel approach to enhance WLAC systems. They proposed a relaxed criterion, replacing the reference translation with the output from a trained machine translation (MT) system. This adjustment made the criterion more practical during inference, allowing for real-time applications.
But, the team did not stop at redefining the criterion. They presented a joint training approach between WLAC and MT. By training these models together, they leveraged the mutual benefits, implicitly improving the performance of both tasks. “By jointly training the two models, we enable them to mutually benefit from each other’s knowledge and improve their respective tasks,” said the researchers.
Experimental results on English-Chinese and English-German language pairs showcased remarkable improvements. The proposed approach surpassed state-of-the-art models by a significant margin, indicating its effectiveness in enhancing WLAC performance.“Our joint training method can greatly improve the performance,” they said.
Notably, the joint training approach not only demonstrated superior performance but also exhibited advantages in terms of model size. The researchers showcased that their approach outperformed top-performing systems submitted to WLAC shared tasks in WMT2022 while utilizing significantly smaller model sizes.
However, they acknowledge that the generalizability of their findings to other languages may vary and that further experiments on multiple languages are needed to gain a comprehensive understanding of the effectiveness of their joint training approach. They also suggest exploring the performance of their method in low-resource scenarios as an important area for future investigation.