REM-Net: Recursive Erasure Memory Network for Commonsense Evidence Refinement

Authors: Yinya Huang, Meng Fang, Xunlin Zhan, Qingxing Cao, Xiaodan Liang6375-6383

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments on two commonsense question answering datasets, WIQA and Cosmos QA. The results demonstrate the performance of REMNet and show that the refined evidence is explainable.
Researcher Affiliation Collaboration Yinya Huang1, Meng Fang2, Xunlin Zhan1, Qingxing Cao1, Xiaodan Liang1* 1 Shenzhen Campus of Sun Yat-sen University 2 Tencent AI Lab / Robotics X
Pseudocode No The paper includes mathematical equations and descriptions of processes, but no explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code No The paper does not provide any statement about releasing source code or a link to a code repository for the described methodology.
Open Datasets Yes We evaluate REM-Net on two commonsense QA datasets, WIQA (Tandon et al. 2019) and Cosmos QA (Huang et al. 2019). Baselines Majority (2019) predicts the most frequent answer option in the training set.
Dataset Splits Yes We evaluate REM-Net on two commonsense QA datasets, WIQA (Tandon et al. 2019) and Cosmos QA (Huang et al. 2019). The model is trained with 10 epochs and a batch size of 4. Results (accuracy%) on the Cosmos QA dev set.
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., GPU/CPU models, memory) used to run the experiments.
Software Dependencies No The paper mentions using BERT, RoBERTa, Adam, COMET, and TAGME, but does not provide specific version numbers for these or any underlying libraries/frameworks (e.g., PyTorch, TensorFlow, Python).
Experiment Setup Yes The model is optimized by Adam (Kingma and Ba 2015) with a learning rate of 1 10 5. Warmup steps are 1000. We train 25 epochs with batch size 8. For the termination condition of the recursion, we set a fixed recursive step to 2. The upper bound of erased evidence sentences at each recursive step is 50. For Cosmos QA, we use a single REM module to refine the evidence. The model is optimized using the Adam optimizer with a learning rate of 5 10 6 and warmup steps of 1500. The model is trained with 10 epochs and a batch size of 4. The fixed recursive step is 2. The upper bound of erased evidence sentences at each recursive step is 10.