Towards Improving Faithfulness in Abstractive Summarization

Authors: Xiuying Chen, Mingzhe Li, Xin Gao, Xiangliang Zhang

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on two benchmark summarization datasets, CNN/DM and XSum, demonstrate that our model significantly outperforms strong baselines. The evaluation of factual consistency also shows that our model generates more faithful summaries than baselines.
Researcher Affiliation Collaboration Xiuying Chen1 Mingzhe Li2 Xin Gao1,3 Xiangliang Zhang4,1 1 Computational Bioscience Reseach Center, King Abdullah University of Science and Technology 2 Ant Group 3 SDAIA-KAUST AI 4 University of Notre Dame {xiuying.chen, xin.g ao}@kaust.edu.sa, limingzhe.lmz@antgroup.com, xzhang33@nd.edu
Pseudocode No The paper describes its model architecture and loss functions using mathematical equations and descriptive text, but it does not include a formal 'Pseudocode' or 'Algorithm' block.
Open Source Code Yes 2https://github.com/iriscxy/FES and We include our code implementation in supplemental material.
Open Datasets Yes We validate the effectiveness of our FES model by conducting extensive experiments on public benchmark CNN/DM [12] and XSum [13] datasets.
Dataset Splits Yes For training the summarization model... For validation and test, we use the pairs selected by the extraction model.
Hardware Specification Yes We implement our experiments in Huggingface [42] on 4 NVIDIA A100 GPUs.
Software Dependencies No The paper mentions software components like Huggingface, BART (facebook/bart-large), and PEGASUS (google/pegasus-xsum), but does not specify their version numbers.
Experiment Setup Yes The QA number is set to 8 unless otherwise stated. We use Adam optimizer with ϵ as 1e-8 and β as (0.9, 0.999). The learning rate is set to 3e-5. The warm-up is set to 500 steps for CNN/DM and 125 for XSum. The batch size is set to 8 with gradient accumulation steps of 4. The beam size is set to 6 for CNN/DM and 8 for XSum.