Retrieve, Program, Repeat: Complex Knowledge Base Question Answering via Alternate Meta-learning

Authors: Yuncheng Hua, Yuan-Fang Li, Gholamreza Haffari, Guilin Qi, Wei Wu

IJCAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluated our MARL model on the CQA dataset [Saha et al., 2018]. CQA is a large-scale complex question answering dataset containing 944K/100K/156K question-answer pairs for training/validation/test... Our full model MARL achieves the best overall performance of 66.96% and 77.71% for macro and micro F1, respectively, outperforming all the baseline models KVmem, CIPITR-All, and CIPITR-Sep. (Tables 1 and 2 also present experimental results and an ablation study).
Researcher Affiliation Academia 1School of Computer Science and Engineering, Southeast University, Nanjing, China 2Faculty of Information Technology, Monash University, Melbourne, Australia 3Southeast University-Monash University Joint Research Institute, Suzhou, China 4Key Laboratory of Computer Network and Information Integration, Southeast University, China
Pseudocode Yes Algorithm 1: The MARL algorithm
Open Source Code Yes We have released our code at https://github.com/Devin Jake/MARL.
Open Datasets Yes We evaluated our MARL model on the CQA dataset [Saha et al., 2018]. CQA is a large-scale complex question answering dataset containing 944K/100K/156K question-answer pairs for training/validation/test.
Dataset Splits Yes CQA is a large-scale complex question answering dataset containing 944K/100K/156K question-answer pairs for training/validation/test.
Hardware Specification No The paper mentions 'We implemented the MARL model in Py Torch' but does not provide any specific hardware details such as GPU models, CPU types, or memory specifications used for running experiments.
Software Dependencies No The paper states 'We implemented the MARL model in Py Torch' and 'Adam optimizer was applied in RL' but does not specify version numbers for PyTorch, Adam, or any other software dependencies.
Experiment Setup Yes We set η1 = 1e-4 when adapting the model to each new task, and set η2 = 0.1 to optimize θ with the gradient update derived from the meta-test data. The reward that the adaptive programmer gained was used to update the retriever parameter φ through the Ada Bound optimizer [Luo et al., 2019] in which the learning rate η3 was initially set to 1e-3 and the final (SGD) learning rate was set to 0.1. When finding the top-N support set, we set N = 5. ... We trained the MARL model with the batch size of 1 and stopped training when the accuracy on the validation set converged (at around 30 epochs).