# Analyzing Compositionality-Sensitivity of NLI Models

arXiv: https://arxiv.org/abs/1811.07033 · HTML: https://mlsanity.com/paper/1811.07033/

## Key facts

- Cited by 26 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 16 Nov 2018.

## Built on by (most cited first, 26 of 26)

- [Dynabench: Rethinking Benchmarking in NLP](https://mlsanity.com/paper/2104.14337/) (2021)
- [Robustness Gym: Unifying the NLP Evaluation Landscape](https://mlsanity.com/paper/2101.04840/) (2021)
- [A fine-grained comparison of pragmatic language understanding in humans and language models](https://mlsanity.com/paper/2212.06801/) (2022)
- [Generating Data to Mitigate Spurious Correlations in Natural Language Inference Datasets](https://mlsanity.com/paper/2203.12942/) (2022)
- [Text-CRS: A Generalized Certified Robustness Framework against Textual Adversarial Attacks](https://mlsanity.com/paper/2307.16630/) (2023)
- [DISCO: Distilling Counterfactuals with Large Language Models](https://mlsanity.com/paper/2212.10534/) (2022)
- [Avoiding Inference Heuristics in Few-shot Prompt-based Finetuning](https://mlsanity.com/paper/2109.04144/) (2021)
- [ReaSCAN: Compositional Reasoning in Language Grounding](https://mlsanity.com/paper/2109.08994/) (2021)
- [Exploring Transitivity in Neural NLI Models through Veridicality](https://mlsanity.com/paper/2101.10713/) (2021)
- [Metaethical Perspectives on 'Benchmarking' AI Ethics](https://mlsanity.com/paper/2204.05151/) (2022)
- [Shaking Syntactic Trees on the Sesame Street: Multilingual Probing with Controllable Perturbations](https://mlsanity.com/paper/2109.14017/) (2021)
- [Is My Model Using The Right Evidence? Systematic Probes for Examining Evidence-Based Tabular Reasoning](https://mlsanity.com/paper/2108.00578/) (2021)
- [longhorns at DADC 2022: How many linguists does it take to fool a Question Answering model? A systematic approach to adversarial attacks](https://mlsanity.com/paper/2206.14729/) (2022)
- [SyGNS: A Systematic Generalization Testbed Based on Natural Language Semantics](https://mlsanity.com/paper/2106.01077/) (2021)
- [Is your benchmark truly adversarial? AdvScore: Evaluating Human-Grounded Adversarialness](https://mlsanity.com/paper/2406.16342/) (2024)
- [Decorrelate Irrelevant, Purify Relevant: Overcome Textual Spurious Correlations from a Feature Perspective](https://mlsanity.com/paper/2202.08048/) (2022)
- [Realistic Data Augmentation Framework for Enhancing Tabular Reasoning](https://mlsanity.com/paper/2210.12795/) (2022)
- [On Measuring the Intrinsic Few-Shot Hardness of Datasets](https://mlsanity.com/paper/2211.09113/) (2022)
- [Pathologies of Pre-trained Language Models in Few-shot Fine-tuning](https://mlsanity.com/paper/2204.08039/) (2022)
- [Methods for Estimating and Improving Robustness of Language Models](https://mlsanity.com/paper/2206.08446/) (2022)
- [Enhancing Tabular Reasoning with Pattern Exploiting Training](https://mlsanity.com/paper/2210.12259/) (2022)
- [A Survey of Neural Network Robustness Assessment in Image Recognition](https://mlsanity.com/paper/2404.08285/) (2024)
- [SMoA: Sparse Mixture of Adapters to Mitigate Multiple Dataset Biases](https://mlsanity.com/paper/2302.14413/) (2023)
- [Using Natural Language Explanations to Improve Robustness of In-context Learning](https://mlsanity.com/paper/2311.07556/) (2023)
- [How the Advent of Ubiquitous Large Language Models both Stymie and Turbocharge Dynamic Adversarial Question Generation](https://mlsanity.com/paper/2401.11185/) (2024)
- [Can current NLI systems handle German word order? Investigating language model performance on a new German challenge set of minimal pairs](https://mlsanity.com/paper/2306.04523/) (2023)

## Abstract (arXiv)

Success in natural language inference (NLI) should require a model to understand both lexical and compositional semantics. However, through adversarial evaluation, we find that several state-of-the-art models with diverse architectures are over-relying on the former and fail to use the latter. Further, this compositionality unawareness is not reflected via standard evaluation on current datasets. We show that removing RNNs in existing models or shuffling input words during training does not induce large performance loss despite the explicit removal of compositional information. Therefore, we propose a compositionality-sensitivity testing setup that analyzes models on natural examples from existing datasets that cannot be solved via lexical features alone (i.e., on which a bag-of-words model gives a high probability to one wrong label), hence revealing the models' actual compositionality awareness. We show that this setup not only highlights the limited compositional ability of current NLI models, but also differentiates model performance based on design, e.g., separating shallow bag-of-words models from deeper, linguistically-grounded tree-based models. Our evaluation setup is an important analysis tool: complementing currently existing adversarial and linguistically driven diagnostic evaluations, and exposing opportunities for future work on evaluating models' compositional understanding.