Towards a Holistic Understanding of Mathematical Questions with Contrastive Pre-training

Authors: Yuting Ning, Zhenya Huang, Xin Lin, Enhong Chen, Shiwei Tong, Zheng Gong, Shijin Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments on two real-world mathematical datasets. The experimental results demonstrate the effectiveness of our model.
Researcher Affiliation Collaboration 1 Anhui Province Key Laboratory of Big Data Analysis and Application, School of Computer Science and Technology, University of Science and Technology of China 2 State Key Laboratory of Cognitive Intelligence 3 i FLYTEK AI Research (Central China), i FLYTEK Co., Ltd.
Pseudocode No The paper describes the proposed framework and its components in detail but does not include any pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/bigdata-ustc/QuesCo.
Open Datasets No The paper mentions using two real-world datasets, SYSTEM1 and SYSTEM2, and describes their statistics. However, it does not provide any specific links, DOIs, or citations to publicly access these datasets.
Dataset Splits No The paper states: "For downstream tasks, we randomly partition labeled questions into training/test sets with 80%/20%." It explicitly mentions train and test sets but does not specify a separate validation split or how validation was handled for hyperparameter tuning.
Hardware Specification Yes All experiments are conducted with one Tesla V100 GPU.
Software Dependencies No The paper mentions that the model is "implemented by Py Torch" and uses "BERT-Base-Chinese" as the base encoder, but it does not specify any version numbers for PyTorch or other software dependencies.
Experiment Setup Yes The projector outputs vectors of size 128. The momentum encoder is updated with a momentum term m = 0.999. The size of memory bank is set to 1,600. Each augmentation strategy is applied with the probability of p = 0.3. We adopt the Adam W optimizer with a learning rate of 0.000005. The temperature factors τi of different ranks in RINCE loss are set to {0.1, 0.1, 0.225, 0.35, 0.6} respectively.