In a research paper printed on October 24, 2023, a gaggle of researchers from the Shanghai Jiao Tong University and Tencent AI Lab highlighted the significance of word-level auto-completion (WLAC) in computer-assisted translation (CAT) and proposed an efficient method to reinforce its efficiency.
Traditionally, WLAC has performed a “crucial role in computer-assisted translation”, serving as an available device for translators aiming to enhance effectivity. As the researchers famous, “effective auto-completion has the potential to reduce keystrokes by at least 60% during the translation process.”
However, they recognized a vital flaw within the current criterion for figuring out a great auto-completion suggestion. This criterion, based mostly on the utmost probability estimation (MLE) of the goal phrase given the supply context, usually led to suboptimal outcomes.
The major challenge was the impracticality of counting on the reference translation throughout prediction, because it was not available in actual time. This problem prompted the researchers to query the essence of WLAC and discover options that would handle this basic challenge.
In response to that, they launched a novel method to reinforce WLAC techniques. They proposed a relaxed criterion, changing the reference translation with the output from a educated machine translation (MT) system. This adjustment made the criterion extra practical throughout inference, permitting for real-time purposes.
But, the workforce didn’t cease at redefining the criterion. They introduced a joint coaching method between WLAC and MT. By coaching these fashions collectively, they leveraged the mutual advantages, implicitly enhancing the efficiency of each duties. “By jointly training the two models, we enable them to mutually benefit from each other’s knowledge and improve their respective tasks,” mentioned the researchers.
Experimental outcomes on English-Chinese and English-German language pairs showcased exceptional enhancements. The proposed method surpassed state-of-the-art fashions by a big margin, indicating its effectiveness in enhancing WLAC efficiency.“Our joint training method can greatly improve the performance,” they mentioned.
Notably, the joint coaching method not solely demonstrated superior efficiency but additionally exhibited benefits by way of mannequin dimension. The researchers showcased that their method outperformed top-performing techniques submitted to WLAC shared duties in WMT2022 whereas using considerably smaller mannequin sizes.
However, they acknowledge that the generalizability of their findings to different languages might fluctuate and that additional experiments on a number of languages are wanted to realize a complete understanding of the effectiveness of their joint coaching method. They additionally counsel exploring the efficiency of their methodology in low-resource situations as an essential space for future investigation.