Adversarial NLI: A New Benchmark for Natural Language Understanding

By Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Aaron Edward Weston, Douwe Kiela. 2019

1,244 citations recorded by the source corpus.

We introduce a new large-scale NLI benchmark dataset, collected via an iterative, adversarial human-and-model-in-the-loop procedure. We show that training models on this new dataset leads to state-of-the-art performance on a variety of popular NLI benchmarks, while posing a more difficult challenge with its new test set. Our analysis sheds light on the shortcomings of current state-of-the-art models, and shows that non-expert annotators are successful at finding their weaknesses. The data collection method can be applied in a never-ending learning scenario, becoming a moving target for NLU, rather than a static benchmark that will quickly saturate.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Natural Language Processing, Evaluation & Benchmarks

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Natural Language Processing

In Evaluation & Benchmarks

Citations in the corpus

How Can We Accelerate Progress Towards Human-like Linguistic Generalization? (2020, background): “approaches such as adversarial filtering that may limit crowdworkers’ ability to resort to shortcuts (Zellers et al., 2018; Nie et al., 2019).”
Adversarial Training for Large Neural Language Models (2020, methodology): “In this subsection, we assess the impact of adversarial pre-training on the model’s robustness against adversarial attacks, using three standard adversarial NLP benchmarks: ANLI (Nie et al., 2019), HELLASWAG (Zellers et al., 2019) and adversarial SQuAD (Jia and Liang, 2017).”
Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model (2022, methodology): “The ANLI [46] dataset is an adversarially mined NLI dataset that aims to create a difﬁcult set of NLI problems.”
Self-Consistency Improves Chain of Thought Reasoning in Language Models (2022, background): “, 2018), ANLI (Nie et al., 2020) and RTE (Dagan et al.”
Repairing the Cracked Foundation: A Survey of Obstacles in Evaluation Practices for Generated Text (2022, background): “, 2018) and subsequently has been used to explore whether adversarial data augmentation techniques are useful for evaluation (Nie et al., 2020; Phang et al., 2021).”
Finetuned Language Models Are Zero-Shot Learners (2021, background): “0 -20 20 40 60 Percent of Data Clean in Dataset Percent change in performance of FLAN (accuracy, F1, or BLEU) eval on only clean data did better eval on all data (including dirty) did better ANLI R2 SQuAD v2 ANLI R1 DROP ReCoRD PIQA 25 500 75 100 Figure 9: Like GPT-3, we also measured performance on”
Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus (2021, methodology): “Our observations support dynamically collecting data with the human-in-the-loop approach (Nie et al., 2020; Kiela et al., 2021) that might reduce contamination of future benchmarks since (i) pretaining data is infrequently collected, and (ii) annotator-written examples for a given task are less…”
Are NLP Models really able to Solve Simple Math Word Problems? (2021, methodology): “Our work differs from adversarial data collection methods such as Adversarial NLI (Nie et al., 2020) in that these methods create examples depending on the failure of a particular model while we create examples without referring to any specific model.”
HateCheck: Functional Tests for Hate Speech Detection Models (2020, background): “Addressing this will require the compilation of “live” datasets, such as dynamic adversarial benchmarks (Nie et al., 2020).”
GPT-NeoX-20B: An Open-Source Autoregressive Language Model (2022, methodology): “Due to space constraints a representative subset of the results are shown here, with the rest in Appendix E. Natural Language Tasks We evaluate our model on a diverse collection of standard language model evaluation datasets: ANLI (Nie et al., 2020), ARC (Clark et al., 2018), HeadQA (English) (Vilar”
The RefinedWeb Dataset for Falcon LLM: Outperforming Curated Corpora with Web Data, and Web Data Only (2023, background): “Tasks Type Random small core main ext HellaSwag (Zellers et al., 2019) Sentence completion 25.0 ✓ ✓ ✓ ✓ LAMBADA (Paperno et al., 2016) Sentence completion 0.0 ✓ ✓ ✓ Winogrande (Sakaguchi et al., 2021) Coreference resolution 50.0 ✓ ✓ ✓ ✓ PIQA (Bisk et al., 2020) Multiple-choice question answering 50.”
Unnatural Instructions: Tuning Language Models with (Almost) No Human Labor (2022, methodology): “In another human-in-the-loop scenario, models trained on Unnatural Instructions can be queried by humans to find examples on which these models fail, thus collecting harder examples (Nie et al., 2020).”
Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned (2022, background): “Others use humans in the loop to continuously and dynamically build, break, and ﬁx [20] models in order to continuously make them more robust to failure modes [40, 32, 55, 61].”
Anticipating Safety Issues in E2E Conversational AI: Framework and Tooling (2021, background): “Evolving benchmarks, such as Dynabench (Kiela et al., 2021), or other adversarial iterative procedures (Dinan et al., 2019b; Nie et al., 2019; Xu et al., 2020) can provide the required adaptability: our societal standards and expectations change, and we would not tolerate models that do not reﬂect…”
Recipes for Safety in Open-domain Chatbots (2020, background): “Dinan et al. (2019b); Nie et al. (2019) are examples of such evolving benchmarks 4 .”
Open-Domain Conversational Agents: Current Progress, Open Problems, and Future Directions (2020, background): “…resources have also been invested in developing much larger and more natural crowdsourced benchmarks such as natural language inference (NLI) tasks (Bowman et al., 2015; Williams, Nangia, and Bowman, 2018; Nie et al., 2020) and commonsense reasoning tasks (Zellers et al., 2019; Qin et al., 2019a).”