Matching Structure for Dual Learning
Authors: Hao Fei, Shengqiong Wu, Yafeng Ren, Meishan Zhang
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Over 2*14 tasks covering 5 dual learning scenarios, the proposed structure matching method shows its significant effectiveness in enhancing existing dual learning. Our method is verified on numbers of dual applications, including text text, text image and text label, where significant improvements are achieved against the vanilla dual learning. |
| Researcher Affiliation | Collaboration | Hao Fei 1 2 Shengqiong Wu 1 Yafeng Ren 3 Meishan Zhang 4 1School of Computing, National University of Singapore, Singapore 2Sea-NEx T Joint Lab, Singapore 3School of Interpreting and Translation Studies, Guangdong University of Foreign Studies, China 4Institute of Computing and Intelligence, Harbin Institute of Technology (Shenzhen), China. |
| Pseudocode | No | The paper describes algorithmic processes such as 'CKY-based chart parsing' but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Source codes are available at https://github.com/scofield7419/StruMatchDL |
| Open Datasets | Yes | For NMT, we use the WMT14 EN-DE and EN-FR data, and take the Para NMT and QUORA datasets for paraphrase generation. ... For text image, we take the MSCOCO and Flickr30k datasets. For text label we use the Yelp2014 and IMDB datasets. ... A.6. Data and Evaluation Description Here we give a detailed description on the datasets and the evaluation settings we used. ... WMT14(EN-DE) (Bojar et al., 2014) ... MSCOCO (Lin et al., 2014) ... Flickr30k (Plummer et al., 2015) ... CUB (Wah et al., 2011) ... Yelp2014 (Asghar, 2016) ... IMDB data (Maas et al., 2011) ... CIFAR-10 (Krizhevsky et al., 2009) ... CIFAR-100 (Krizhevsky et al., 2009) ... Celeb A-HQ (Karras et al., 2018) ... AFHQ (Choi et al., 2020) ... AGNews (Corso et al., 2005) |
| Dataset Splits | Yes | WMT14(EN-DE) (Bojar et al., 2014) splits the total sentences into training (4.6M), developing (3K) and testing (2K). ... QUORA includes 146K parallel paraphrases, 3K and 30K paraphrase pairs are respectively used for validation and testing, following Miao et al. (2019). |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models, processor types, or memory used for running the experiments. |
| Software Dependencies | No | The paper mentions software components such as 'Stanford NLP', 'N-ary Tree LSTM', 'Transformer-based', and 'BART PLM', but it does not specify concrete version numbers for these or any other software dependencies needed to replicate the experiments. |
| Experiment Setup | Yes | For the NMT, we take the 12-layer configuration, while for paraphrase generation we take the 6-layer of Transformer. Also we use the position embedding. Besides, for the NMT task, we meanwhile implement the sequence-to-sequence baseline, which is a standard attention-based encoder-decoder architecture (Bahdanau et al., 2015), with 3-layer Bi LSTMs as encoder and 2-layer LSTM as decoder. We use beam search with a beam size 5 and length penalty 1.0, so as to yield 5-best generated texts. ... We adopt the Adam optimizer with an initial learning rate of 1e-5 for training the classifier. |