Towards Incremental NER Data Augmentation via Syntactic-aware Insertion Transformer

Authors: Wenjun Ke, Zongkai Tian, Qi Liu, Peng Wang, Jinhua Gao, Rui Qi

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on two benchmark datasets, i.e., Ontonotes and Wikiann, demonstrate the comparable performance of SAINT over the state-of-the-art baselines.Extensive experiments are conducted on the Wikiann and Onto Notes datasets to verify the effectiveness of our method and demonstrate the significance of introducing syntactic features.
Researcher Affiliation Collaboration Wenjun Ke1 , Zongkai Tian2 , Qi Liu2 , Peng Wang1 , Jinhua Gao3 , Rui Qi4 1School of Computer Science and Engineering, Southeast University 2Beijing Institute of Computer Technology and Application 3Institute of Computing Technology, Chinese Academy of Sciences 4China Life Property Casualty Insurance Company Limited
Pseudocode No The paper describes the methods in text and uses flow diagrams (Figure 2) but does not include any pseudocode or algorithm blocks.
Open Source Code No The paper does not provide an explicit statement about releasing source code or a link to a repository for the described methodology.
Open Datasets Yes We conduct experiments on the two NER datasets of Wikiann [Pan et al., 2017] and Onto Notes [Hovy et al., 2006].
Dataset Splits Yes In the experiment, we sampled 100, 200, 300, 400, and 500 items of data as training data under low resource constraints. The model with the best validation set performance in 20 epochs is chosen as the final model for testing.
Hardware Specification Yes Our method is implemented with Python 3.7.12, Py Torch 1.8.0, and the NVIDIA Tesla V100 16G platform.
Software Dependencies Yes Our method is implemented with Python 3.7.12, Py Torch 1.8.0, and the NVIDIA Tesla V100 16G platform.
Experiment Setup Yes In the experiment, SAINT utilized the Bert-base model as the pretrained language model and adopted the Adam W optimizer with a learning rate of 1e-5. The training batch size is 20, and the model is trained for 5 epochs. The inference process is performed in three steps. The parameter λs is adjusted to 0.2. The NER model s optimizer in the experiment is the Adam W optimizer, the learning rate is 2e-5, and the batch size is 8. The model with the best validation set performance in 20 epochs is chosen as the final model for testing.