Copy or Rewrite: Hybrid Summarization with Hierarchical Reinforcement Learning
Authors: Liqiang Xiao, Lu Wang, Hao He, Yaohui Jin9306-9313
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Automatic evaluation shows that our approach significantly outperforms the state-of-the-arts on the CNN/Daily Mail corpus. Human evaluation also demonstrates that our generated summaries are more informative and concise than popular models. In this section, we verify our methods on the most popular dataset CNN/Daily Mail, so that we can make a fully comparison with the state-of-the-art works. |
| Researcher Affiliation | Academia | Liqiang Xiao,1 Lu Wang,2 Hao He,1,3 Yaohui Jin1,3 1Mo E Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University 2Khoury College of Computer Sciences, Northeastern University, Boston, MA 02115 3State Key Lab of Advanced Optical Communication System and Network, Shanghai Jiao Tong University |
| Pseudocode | No | The paper describes its methods in prose and diagrams (Figure 2) but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide concrete access to source code for the methodology described, nor does it contain explicit statements about code release or specific repository links for the authors' implementation. |
| Open Datasets | Yes | We evaluate our approach on the summarization corpus CNN/Daily Mail, which is comprised of news stories and is first proposed by Hermann et al. (2015). |
| Dataset Splits | Yes | We follow the processing steps in (See, Liu, and Manning, 2017), obtaining 287,188 training, 13,367 validation and 11, 490 testing samples. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running its experiments. |
| Software Dependencies | No | The paper mentions tools like BERT and Adam optimizer but does not specify versions for core software dependencies such as deep learning frameworks (e.g., PyTorch, TensorFlow) or programming languages and their specific library versions. |
| Experiment Setup | Yes | The word vectors (Mikolov et al., 2013) are pretrained on the whole dataset with 128 dimension. And all the LSTM networks in our framework use 256 hidden units. Adam optimizer (Kingma and Ba, 2014) is applied with a learning rate 0.001 for extractor and 0.0001 for rewriter, and the minibatches size is set to 32. In RL training stage, we set the discount factor λ as 0.95 for return and coefficient β as 0.9 with grid search. During reference, we apply the beam search (Pasunuru and Bansal, 2018) with width 5 on rewriter to avoid trigram repetition. |