Topic Modeling Revisited: A Document Graph-based Neural Network Perspective
Authors: Dazhong Shen, Chuan Qin, Chao Wang, Zheng Dong, Hengshu Zhu, Hui Xiong
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, extensive experiments on four benchmark datasets have clearly demonstrated the effectiveness and interpretability of GNTM compared with state-of-the-art baselines. |
| Researcher Affiliation | Collaboration | Dazhong Shen1,2, Chuan Qin2, Chao Wang1,2, Zheng Dong2, Hengshu Zhu2, , Hui Xiong3, 1School of Computer Science and Technology, University of Science and Technology of China 2Baidu Talent Intelligence Center, Baidu Inc. 3Artificial Intelligence Thrust, The Hong Kong University of Science and Technology |
| Pseudocode | No | The paper does not contain explicit pseudocode or labeled algorithm blocks. Figures 1(b) and 2 are diagrams, not pseudocode. |
| Open Source Code | Yes | Our code and data are available at https://github.com/SmilesDZgk/GNTM. |
| Open Datasets | Yes | Our experiments are conducted on four benchmark datasets with varying sizes, including 20 News Groups (20NG) [27], Tag My News (TMN) [51], the British National Corpus (BNC) [10], Reuters extracted from the Reuters-21578 dataset. The statistics and links of these datasets are shown in Appendix A.4. |
| Dataset Splits | No | The paper uses benchmark datasets but does not explicitly provide the train/validation/test split percentages, sample counts, or refer to specific predefined splits with details for reproduction. It mentions evaluating on 'labeled datasets' for clustering but no split details for this either. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, processor types, or memory amounts) used for running its experiments. |
| Software Dependencies | No | The paper mentions software components like 'Adam optimizer' and 'GloVe word embeddings' but does not provide specific version numbers for any software dependencies needed to replicate the experiment. |
| Experiment Setup | Yes | In practice, we construct document graphs with window size s = 53. We set the Dirichlet prior parameter α = 1. We utilized 300-dimensional Glo Ve word embeddings [41] to fix X (i.e., L = 300) in our model and word vectors in ETM. We also set the size of word vectors µv as 300, i.e., H = 300. The size of transitional vector ak and bk were set as Y = 64. For the optimization, Adam [23] optimizer has been used with the initial learning rate of 0.001 and the linear learning rate decay trick to find the optimal. |