Contrastive Triple Extraction with Generative Transformer

Authors: Hongbin Ye, Ningyu Zhang, Shumin Deng, Mosha Chen, Chuanqi Tan, Fei Huang, Huajun Chen14257-14265

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on three datasets (i.e., NYT, Web NLG, and MIE) show that our approach achieves better performance than that of baselines.
Researcher Affiliation Collaboration Hongbin Ye1,2*, Ningyu Zhang1,2 , Shumin Deng1,2, Mosha Chen3, Chuanqi Tan3, Fei Huang3, Huajun Chen1,2 1 Zhejiang University 2 AZFT Joint Lab for Knowledge Engine 3 Alibaba Group
Pseudocode Yes Algorithm 1 Triplet Contrastive Learning; Algorithm 2 Batch-wise Dynamic Attention Masking
Open Source Code No The paper links to pre-trained models (UniLM) and datasets, but does not provide a link or explicit statement for the open-sourcing of their own CGT model's implementation code.
Open Datasets Yes We conducted experiments on three benchmark datasets: New York Times (NYT), Web NLG1, and MIE2. The NYT dataset is produced using a distant supervision method and is widely used for triplet extraction (Riedel, Yao, and Mc Callum 2010). [...] The Web NLG dataset (Gardent et al. 2017) was used for natural language generation, but was later used for triplet extraction (Zeng et al. 2018a). [...] MIE (Zhang et al. 2020d) is a large-scale Chinese dialogue information extraction dataset for the medical domain.
Dataset Splits Yes It contains 56,195 sentences for training, 5,000 sentences for validation, and 5,000 sentences for test. The Web NLG dataset [...] It consists of 5,019/500/703 instances for training, validation, and testing, respectively. MIE [...] It contains 800 instances for training, 160 instances for validation, and 160 instances for testing.
Hardware Specification Yes We utilized Pytorch to implement our CGT model and conducted experiments using four Nvidia 1080-Ti graphical processing units.
Software Dependencies No The paper mentions "Uni LM-base-uncased" and "Pytorch" but does not specify their version numbers for reproducibility.
Experiment Setup Yes The initial learning rate was set to 2e-5, and we reduced the rate by 20% at every eight epochs. The batch size was 64 for English and 32 for Chinese, and the total number of epochs was 50 for all datasets. The beam size was set to 4, α was set to 0.1, γ was set to 0.2, and θ was set to 0.6.