Coreference Graph Guidance for Mind-Map Generation
Authors: Zhuowei Zhang, Mengting Hu, Yinhao Bai, Zhen Zhang
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that our model outperforms all the existing methods. The case study further proves that our model can more accurately and concisely reveal the structure and semantics of a document. Extensive experimental results demonstrate that the proposed method outperforms the state-of-the-art method. |
| Researcher Affiliation | Academia | Zhuowei Zhang, Mengting Hu*, Yinhao Bai, Zhen Zhang College of Software, Nankai University {zhuoweizhang, yinhao, zhangzhen}@mail.nankai.edu.cn, mthu@nankai.edu.cn |
| Pseudocode | Yes | Algorithm 1: Coreference Graph Generation |
| Open Source Code | Yes | Code and data are available at https://github.com/Cyno2232/CMGN. |
| Open Datasets | No | We select 5000 articles from CNN news, where both the length and the number of sentences are no more than 50. We utilize the fine-tuned Distil BERT to annotate their pseudo graphs, which are then used for model training. While the source articles are publicly available, the specific selection and annotated pseudo graphs for this training set are not explicitly linked or stated as publicly available. |
| Dataset Splits | Yes | The benchmark consists of a testing set Dt and a validation set Dv, with 120 and 15 articles respectively. |
| Hardware Specification | Yes | CPU: Intel(R) Xeon(R) Gold 5218 CPU @ 2.30GHz GPU: NVIDIA Tesla V100S PCIe 32 GB |
| Software Dependencies | Yes | System: Ubuntu 22.04.2; Python 3.7; Py Torch 1.12.1; DGL 1.0.2+cu116 (For the implementation of GNNs) |
| Experiment Setup | Yes | Hyperparameter Settings We initialize the word embeddings with 50-dimension Glo VE. The hidden size of Bi LSTM is set to be 25 2. The models are optimized by Adam (Kingma and Ba 2015) with a learning rate of 1e-4. The batch size is 64. For the coreference graph encoder, we employ a 2-layer model for training. For the graph enhancement module, we employ a 5-layer GIN (Xu et al. 2018) model as the base encoder. We set η to 0.2, which adjusts the magnitude of the perturbation of the base encoder, and λ to 0.001, which adjusts the impact of contrastive loss. |