Adversity-aware Few-shot Named Entity Recognition via Augmentation Learning
Authors: Li Huang, Haowen Liu, Qiang Gao, Jiajing Yu, Guisong Liu, Xueqin Chen
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on both certain and uncertain datasets, including few-shot and cross-domain conditions, demonstrate the superiority and robustness of the proposed AAL compared to state-of-the-art baselines. Performance on Certainty Condition. Table 1 and Table 2 present the performance comparison of our AAL against baselines on SNIPS and Cross-Dataset under certain conditions. AAL achieves an average enhancement of 6.27% and 3.04% in overall results for the 1-shot and 5-shot scenarios, outperforming the robust baseline MANNER. Performance on Uncertainty Condition. Table 3 and Table 4 report the performance of our AAL alongside baselines on SNIPS and Cross-Dataset following the application of Bert-Attack adversarial algorithm to the target domain data. |
| Researcher Affiliation | Academia | 1School of Computing and Artificial Intelligence, Southwestern University of Finance and Economics, Chengdu, China 2Engineering Research Center of Intelligent Finance, Ministry of Education, Chengdu, China 3Kash Institute of Electronics and Information Industry, Kashgar, China EMAIL, EMAIL, EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes the methodology using mathematical formulations and textual descriptions of steps, but it does not contain a clearly labeled 'Pseudocode' or 'Algorithm' block. |
| Open Source Code | Yes | For reproduction, the source code is released at https://github. com/swufe-Nice Lab-Geo Text/AAL.git. |
| Open Datasets | Yes | To align with previous studies, we conduct experiments on the following datasets: (1) SNIPS (Coucke et al. 2018): It has 7 domains with different label sets and a small number of samples, with a relatively even number of samples per domain per label set, which makes it easy to simulate a small number of samples. (2) Cross-Dataset (Hou et al. 2020): It is constructed from datasets from four different domains: Co NLL-2003 (Tjong Kim Sang 2002), GUM (Zeldes 2017), WNUT-2017 (Derczynski et al. 2017), and Ontonotes (Pradhan et al. 2013). |
| Dataset Splits | Yes | Few-shot NER on Episode Learning. Given the source domain Ds = {(Ss, Qs)}, the task of few-shot NER should adapt to the target domain of Dt = {(St, Qt)}. Under episode learning, each episode consists of a support set Ss/t = {(x(i) s/t, y(i) s/t)}N K i=1 for adaption, and a query set Qs/t = {(x(j) s/t, y(j) s/t)}N K j=1 for evaluation. Here, N denotes the number of entity types in an episode, K and K denote the number of examples per entity type in support set and query set, respectively, commonly referred N-way K-shot setting (Ding et al. 2021). Typically, K is very small, often K = 1 or 5. |
| Hardware Specification | Yes | We employ the Adam W optimizer for AAL, accelerated on an NVIDIA A100 GPU. |
| Software Dependencies | No | The paper mentions using 'Uncased-Bert base model as the PLM' and 'Adam W optimizer' but does not specify version numbers for any software libraries or dependencies like Python, PyTorch, TensorFlow, etc. |
| Experiment Setup | Yes | Implementation Details. We exploit the Uncased-Bert base model as the PLM. {dm, dz, ne, α, γ} and dropout are set to {768, 128, 5, 0.5, 0.5, 0.1}. We employ the Adam W optimizer for AAL, accelerated on an NVIDIA A100 GPU. |