SEAT: Stable and Explainable Attention

Authors: Lijie Hu, Yixin Liu, Ninghao Liu, Mengdi Huai, Lichao Sun, Di Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, through intensive experiments on various datasets, we compare our SEAT with other baseline methods using RNN, Bi LSTM and BERT architectures, with different evaluation metrics on model interpretation, stability and accuracy. Results show that, besides preserving the original explainability and model performance, SEAT is more stable against input perturbations and training randomness, which indicates it is a more faithful explanation.
Researcher Affiliation Academia 1 King Abdullah University of Science and Technology, Thuwal, Saudi Arabia 2 Lehigh University, Bethlehem, Pennsylvania, USA 3 University of Georgia, Athens, Georgia, USA 4 Iowa State University, Ames, Iowa, USA
Pseudocode Yes Algorithm 1 Finding a SEAT
Open Source Code No The paper does not provide an explicit statement or link to the open-source code for the methodology described.
Open Datasets Yes In all experiments, we use four datasets: Stanford Sentiment Treebank (SST) (Socher et al. 2013), Emotion Recognition (Emotion) (Mohammad et al. 2018), Hate (Basile et al. 2019) and Rotten Tomatoes (Rotten T) (Pang and Lee 2005).
Dataset Splits No The paper mentions using datasets for training but does not provide specific percentages or counts for train/validation/test splits, or details about cross-validation setup in the main text.
Hardware Specification No The paper does not specify the hardware used to run the experiments (e.g., GPU models, CPU models, or cloud computing instance types).
Software Dependencies No The paper mentions using certain models (RNN, Bi LSTM, BERT) and a specific loss function (Binary Cross Entropy) but does not provide specific version numbers for software libraries or dependencies (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes For the encoder, we consider three kinds of networks as feature extractors: RNN, Bi LSTM, and BERT. For the decoder, we apply one simple MLP followed by a tanh-attention layer (Bahdanau, Cho, and Bengio 2015) and a softmax layer (Vaswani et al. 2017). [...] We select the Binary Cross Entropy loss as D1 and D2 in (6). [...] The perturbation radius is set as δx=1e-3.