Adversarial NLI for Factual Correctness in Text Summarisation Models
arXiv:2005.11739
Abstract
We apply the Adversarial NLI dataset to train the NLI model and show that the model has the potential to enhance factual correctness in abstract summarization. We follow the work of Falke et al. (2019), which rank multiple generated summaries based on the entailment probabilities between an source document and summaries and select the summary that has the highest entailment probability. The authors' earlier study concluded that current NLI models are not sufficiently accurate for the ranking task. We show that the Transformer models fine-tuned on the new dataset achieve significantly higher accuracy and have the potential of selecting a coherent summary.
References in corpus (7)
- Towards A Rigorous Science of Interpretable Machine Learning
- Shortcut Learning in Deep Neural Networks
- XLNet: Generalized Autoregressive Pretraining for Language Understanding
- TinyBERT: Distilling BERT for Natural Language Understanding
- WT5?! Training Text-to-Text Models to Explain their Predictions
- Adversarial NLI: A New Benchmark for Natural Language Understanding
- Learning by Semantic Similarity Makes Abstractive Summarization Better