KATG: Keyword-Bias-Aware Adversarial Text Generation for Text Classification

Authors: Lingfeng Shen, Shoushan Li, Ying Chen11294-11302

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments demonstrate that adversarial sentences generated by our KATG model can strengthen the generalization and the robustness of text classification models. In this section, we carry out three sets of experiments to comprehensively analyze our KATG model.
Researcher Affiliation Academia 1College of Science, China Agricultural University, China 2Natural Language Processing Lab, Soochow University, China 3College of Information and Electrical Engineering, China Agricultural University, China
Pseudocode Yes Algorithm 1: Adversarial training with KATG
Open Source Code No The paper does not provide any statement or link indicating that the source code for the described methodology is publicly available.
Open Datasets Yes Datasets (1) AGNews: It consists of news in four categories. Each category contains 30,000 training examples and 1,900 test examples. (2) Amazon: It is a binary sentiment classification dataset containing Amazon reviews. We take 60,000 reviews per class. (3) SST-2: It is built on movie reviews for binary sentiment classification (Socher et al. 2013). We use its standard split 6,920 (training)-872 (dev)-1,821 (test) . (4) IMDb: It is a large dataset for binary sentiment classification, containing 25,000 movie reviews for training and 25,000 for testing.
Dataset Splits Yes SST-2: It is built on movie reviews for binary sentiment classification (Socher et al. 2013). We use its standard split 6,920 (training)-872 (dev)-1,821 (test) .
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU models, or memory specifications used for running the experiments.
Software Dependencies No The paper mentions software components like Bi-LSTM, Adam, Text CNN, LSTM, and BERT, but it does not specify version numbers for any libraries or frameworks (e.g., PyTorch, TensorFlow, scikit-learn).
Experiment Setup Yes In KATG, the hidden state size m in Bi-LSTM, the chain length T in the generative flow model and the number of prior sentences d are 256, 10 and 6. During the training of KATG, the batch size is set to 16 and the learning rate of Adam is 0.0005. Moreover, we implement three widely-used text classification networks, Text CNN (Kim 2014), LSTM (Hochreiter and Schmidhuber 1997) and BERT (Devlin et al. 2019). For BERT, we choose the default settings of BERT-base-uncase.