Falsesum: Generating Document-level NLI Examples for Recognizing Factual Inconsistency in Summarization
21210Authors
nan
Abstract
AbstractNeural abstractive summarization models are prone to generate summaries that are factually inconsistent with their source documents. Previous work has introduced the task of recognizing such factual inconsistency as a downstream application of natural language inference (NLI). However, state-of-the-art NLI models perform poorly in this context due to their inability to generalize to the target task. In this work, we show that NLI models can be effective for this task when the training data is augmented with high-quality task-oriented examples. We introduce Falsesum, a data generation pipeline leveraging a controllable text generation model to perturb human-annotated summaries, introducing varying types of factual inconsistencies. Unlike previously introduced document-level NLI datasets, our generated dataset contains examples that are diverse and inconsistent yet plausible. We show that models trained on a Falsesum-augmented NLI dataset improve the state-of-the-art performance across four benchmarks for detecting factual inconsistency in summarization.
Keywords
Full Content
Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 258 - 271 July 10-15, 2022 ©2022 Association for Computational Linguistics GlobEnc: Quantifying Global Token Attribution by Incorporating the Whole Encoder Layer in Transformers Ali Modarressi1⋆ Mohsen Fayyaz2⋆ Yadollah Yaghoobzadeh2 Mohammad Taher Pilehvar3 1 Iran University of Science and Technology, Iran 2 University of Tehran, Iran 3 Tehran Institute for Advanced Studies, Khatam University, Iran m_modarressi@comp.iust.ac.ir {mohsen.fayyaz77, y.yaghoobzadeh}@ut.ac.ir mp792@cam.ac.uk Abstract There has been a growing interest in inter- preting the underlying dynamics of Transform- ers. While self-attention patterns were initially deemed as the primary option, recent studies have shown that integrating other components can yield more accurate explanations. This paper introduces a novel token attribution anal- ysis method that incorporates all the compo- nents in the encoder block and aggregates this throughout layers. Through extensive quanti- tative and qualitative experiments, we demon- strate that our method can produce faithful and meaningful global token attributions. Our ex- periments reveal that incorporating almost ev- ery encoder component results in increasingly more accurate analysis in both local (single layer) and global (the whole model) settings. Our global attribution analysis significantly outperforms previous methods on various tasks regarding correlation with gradient-based saliency scores. Our code is freely available at https://github.com/mohsenfayyaz/GlobEnc. 1 Introduction The stellar performance of Transformers (Vaswani et al., 2017) has garnered a lot of attention to ana- lyzing the reasons behind their effectiveness. The self-attention mechanism has been one of the main areas of focus (Clark et al., 2019; Kovaleva et al., 2019; Reif et al., 2019; Htut et al., 2019). How- ever, there have been debates on whether raw at- tention weights are reliable anchors for explain- ing model’s behavior or not (Wiegreffe and Pinter, 2019; Serrano and Smith, 2019; Jain and Wallace, 2019). Recently, it was shown that incorporating vector norms should be an indispensable part of any attention-based analysis1 (Kobayashi et al., 2020, ⋆ Equal contribution. 1We also have shown the unreliability of weights due to norm disparities in probing studies (Fayyaz et al., 2021). [CLS] a deep and meaningful film . [SEP] 12 11 10 9 8 7 6 5 4 3 2 1 Layer 0.2 0.4 0.6 0.8 1.0 [CLS] big fat waste of time . [SEP] 12 11 10 9 8 7 6 5 4 3 2 1 Layer 0.2 0.4 0.6 0.8 1.0 Figure 1: Aggregated attribution maps (NENC) for the [CLS] token for fine-tuned BERT on SST2 dataset (sen- timent analysis). Our method (GlobEnc) is able to accu- rately quantify the global token attribution of the model. 2021). However, these norm-based studies incor- porate only the attention block into their analysis, whereas Transformer encoder layer is composed of more components. Another limitation of the existing analysis tech- niques is that they are usually constrained to the analysis of single layer attributions. In order to expand the analysis to multi-layered encoder-based models in their entirety, an aggregation technique has to be employed. Abnar and Zuidema …
Content truncated to 500 words
Labeled Text
AbstractNeural abstractive summarization models are prone to generate summaries that are factually inconsistent with their source documents. Previous work has introduced the task of recognizing such factual inconsistency as a downstream application of natural language inference (NLI). However, state-of-the-art NLI models perform poorly in this context due to their inability to generalize to the target task. In this work, we show that NLI models can be effective for this task when the training data is augmented with high-quality task-oriented examples. We introduce Falsesum, a data generation pipeline leveraging a controllable text generation model to perturb human-annotated summaries, introducing varying types of factual inconsistencies. Unlike previously introduced document-level NLI datasets, our generated dataset contains examples that are diverse and inconsistent yet plausible. We show that models trained on a Falsesum-augmented NLI dataset improve the state-of-the-art performance across four benchmarks for detecting factual inconsistency in summarization.
Publication Details
- Paper ID:
- 21210
- Year:
- 2022
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan