Unsupervised Extractive Summarization with Learnable Length Control Strategies
Authors: Renlong Jie, Xiaojun Meng, Xin Jiang, Qun Liu
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that our unsupervised method largely outperforms the centrality-based baseline using a same sentence encoder. In terms of length control ability, via our trainable knapsack module, the performance consistently outperforms the strong baseline without utilizing end-to-end training. Human evaluation further evidences that our method performs the best among baselines in terms of relevance and consistency. We conduct experiments on CNNDM dataset (Hermann et al. 2015), NYT dateset (Durrett, Berg-Kirkpatrick, and Klein 2016) and CNew Sum dataset (Wang et al. 2021) for news summarization in both English and Chinese. |
| Researcher Affiliation | Collaboration | Renlong Jie1, 2*, Xiaojun Meng1 , Xin Jiang1, Qun Liu1 1Huawei Noah s Ark Lab 2Northwestern Polytechnical University, China jierenlong@nwpu.edu.cn, {xiaojun.meng, Jiang.Xin, qun.liu}@huawei.com |
| Pseudocode | No | The paper describes its methods in prose and through diagrams (Figure 1, Figure 2) but does not include explicit pseudocode blocks or algorithms with structured steps. |
| Open Source Code | No | The paper mentions using external pre-trained models (Sim CSE from Huggingface) and loading a pre-trained knapsack transformer, but it does not state that the source code for *their* proposed methodology is being released or provide a link to it. |
| Open Datasets | Yes | We conduct experiments on CNNDM dataset (Hermann et al. 2015), NYT dateset (Durrett, Berg-Kirkpatrick, and Klein 2016) and CNew Sum dataset (Wang et al. 2021) for news summarization in both English and Chinese. |
| Dataset Splits | Yes | After labelling indices, 95% of samples are used as a train set and the rest as a validation set. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. |
| Software Dependencies | No | The paper mentions using the 'pretrained Sim CSE model (Gao, Yao, and Chen 2021) from Huggingface' and 'Adam optimizers', but it does not specify version numbers for key software dependencies like PyTorch, TensorFlow, Python, or CUDA. |
| Experiment Setup | Yes | The scorer is a 2-layer transformer with 4 attention heads and 768 hidden size, and dimension of FFNN layer is 2048. For the transformer encoder and predictor in contrastive learning module, we use 4-layer transformers with 8 heads and 768 hidden size. The seed is set to be 42. We apply two Adam optimizers with β1 = 0.9, β2 = 0.999 for training randomly initialized parameters and loaded pretrained parameters, respectively. We set the main learning rate lr = 3e 6 for CNNDM and NYT and lr = 1e 6 for CNew Sum after a grid search from set {3e-8, 1e-7, 3e-7, 1e-6, 3e-6, 1e-5}, and the batch size is 64. We apply λ1 = 1 and λ2 = 0.3 for CNNDM/NYT and λ1 = 1 and λ2 = 1 for CNew Sum as in Eq. (8). |