Denoising Pre-training for Machine Translation Quality Estimation with Curriculum Learning

Authors: Xiang Geng, Yu Zhang, Jiahuan Li, Shujian Huang, Hao Yang, Shimin Tao, Yimeng Chen, Ning Xie, Jiajun Chen

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on various benchmarks reveal that CLQE outperforms Direct QE and other strong baselines.
Researcher Affiliation Collaboration 1 National Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China 2 Huawei Translation Services Center, Beijing, China {gx, zhangy, lijh}@smail.nju.edu.cn, {huangsj, chenjj}@nju.edu.cn, {yanghao30, taoshimin, chenyimeng, nicolas.xie}@huawei.com
Pseudocode Yes Algorithm 1: Denoising pre-training for machine translation quality estimation.
Open Source Code Yes We make our CLQE code available (https://github.com/ NJUNLP/njuqe). ... We provide our implementation online.5
Open Datasets Yes We employ WMT19 and WMT20/WMT214 QE dataset for English-German (EN-DE) and English-Chinese (EN-ZH) direction respectively. ... https://www.statmt.org/wmt##, ## can be 19, 20, 21.
Dataset Splits Yes The size of training, development, and test sets are 13K/1K/1K, 7K/1K/1K, and 8K/1K/1K for WMT19, 20, and 21 QE tasks, respectively. ...the pre-trained model is selected with the pseudo validation set for further fine-tuning.
Hardware Specification Yes All experiments are performed on NVIDIA V100 GPUs.
Software Dependencies No The paper mentions using Fairseq(-py) but does not provide a specific version number. Other mentioned software or models (e.g., XLM-R, GPT-2) do not include version details for their implementations.
Experiment Setup Yes We set the initial competence c0 = 0.05 and total duration of curriculum learning T = 5 epochs. Other details can be found in supplementary materials.