TDv2: A Novel Tree-Structured Decoder for Offline Mathematical Expression Recognition
Authors: Changjie Wu, Jun Du, Yunqing Li, Jianshu Zhang, Chen Yang, Bo Ren, Yiqing Hu2694-2702
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | On the authoritative CROHME 14/16/19 datasets, our method achieves the state-of-the-art results. |
| Researcher Affiliation | Collaboration | 1 NEL-SLIP Lab, University of Science and Technology of China 2 i FLYTEK Research 3 Tencent Youtu |
| Pseudocode | No | The paper describes the proposed method in detail with figures illustrating the architecture and process, but it does not include formal pseudocode or algorithm blocks. |
| Open Source Code | Yes | The source codes of our proposed TDv2 are available at https://github.com/yqingli123/TDv2.git. |
| Open Datasets | Yes | We verified the proposed model on CROHME benchmark (Mouchere et al. 2016; Mouch ere et al. 2016; Mahdavi et al. 2019), which is the most widely used public dataset for HMER. |
| Dataset Splits | No | The training set has 8836 mathematical expressions, including 101 mathematical symbol classes and 9 mathematical spatial relations. We verified the proposed model on CROHME benchmark... CROHME 2014/2016/2019 test sets contain 986/1147/1199 handwritten mathematical expressions respectively. While training and test set sizes are specified, explicit details about a validation set split or its size are not provided in the paper. |
| Hardware Specification | Yes | We implement TDv2 using Py Torch and conduct experiments on a Nvidia Tesla V100 with 12GB RAM. |
| Software Dependencies | No | The paper mentions using 'Py Torch' for implementation but does not specify its version number or any other software dependencies with their specific version numbers. |
| Experiment Setup | Yes | In our experiment, we set λ1 = λ2 = 1, λ3 = 0.5. The encoder uses a Dense Net... We set the depth of each Dense Block to 22, the channel of each layer to 24. Both the node classification module and branch prediction module adopt 2 GRU layers. The dimension of the GRU unit, the dimension of the word embedding layer and the dimension of the relation embedding layer are set to 256. The dimension of the node attention mechanism and the branch attention mechanism is set to 512. The optimization algorithm is Adadelta (Zeiler 2012) with gradient clipping, and the learning rate is set to 1. |