Papperoni

2019

Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference

R. T. Mccoy, Ellie Pavlick, Tal Linzen

Open PDF Google Scholar

citations

Cite Score

47

AI summary

This paper introduces the HANS dataset to evaluate syntactic heuristics in NLI models, finding that models like BERT rely on fallible heuristics, leading to poor performance on HANS; augmenting training data improves performance, indicating room for progress in NLI systems by addressing these biases.

Main Contributions

Introduces the HANS dataset, an NLI evaluation set designed to test specific hypotheses about invalid heuristics.
Demonstrates shortcomings in state-of-the-art models trained on MNLI using the HANS dataset.
Shows that these shortcomings can be mitigated by augmenting a model's training set with HANS-like examples.
Finds that four existing NLI models perform very poorly on HANS, suggesting that their high accuracies on NLI test sets may be due to the exploitation of invalid heuristics.
Shows models performed significantly better on both HANS and on a separate structure-dependent dataset when their training data was augmented with HANS-like examples.

Abstract

A machine learning system can score well on a given test set by relying on heuristics that are effective for frequent example types but break down in more challenging cases. We study this issue within natural language inference (NLI), the task of determining whether one sentence entails another. We hypothesize that statistical NLI models may adopt three fallible syntactic heuristics: the lexical overlap heuristic, the subsequence heuristic, and the constituent heuristic. To determine whether models have adopted these heuristics, we introduce a controlled evaluation set called HANS (Heuristic Analysis for NLI Systems), which contains many examples where the heuristics fail. We find that models trained on MNLI, including BERT, a state-of-the-art model, perform very poorly on HANS, suggesting that they have indeed adopted these heuristics. We conclude that there is substantial room for improvement in NLI systems, and that the HANS dataset can motivate and measure progress in this area.

Citation Graph

Loading graph...

References [46]

Sort:

Filter:

[1]BERT: Pre-Training of Deep Bidirectional Transformers for Language Understanding

Jacob Devlin, M. W. Chang, K. Lee, Kristina Toutanova - 2018

39 papers in library cite

Simply amazing. It's very impressive how they make a leap vs. existing stuff (you can see from the references, pretty much no one is doing what they are doing, other than GPT)

[2]A Large Annotated Corpus for Learning Natural Language Inference

Samuel R. Bowman, G. Angeli, Christopher Potts, Christopher D. Manning - 2015

25 papers in library cite

Dataset collection is ok. The model that they create seems very low effort.

[3]A Broad-Coverage Challenge Corpus for Sentence Understanding Through Inference

A. Williams, Nikita Nangia, S. Bowman - 2018

19 papers in library cite

Very nice paper and cool dataset - good thing they expanded SNLI. Also, they at least tried to have a good baseline, and comparisons of domains are nice.

[4]The PASCAL Recognising Textual Entailment Challenge

Ido Dagan, O. Glickman, Bernardo Magnini - 2005

19 papers in library cite

It's very nice how they had the foresight to create a challenge that became relevant like 10 years later.

[5]A Decomposable Attention Model for Natural Language Inference

A. P. Parikh, O. Tackstrom, Dipanjan Das, Jakob Uszkoreit - 2016

11 papers in library cite

Very nice alternative to the common LSTM encoder-decoder architecture! Seems similar o the Transformers arch in the sense that they don't use RNNs. Nice that they analyze computational complexity as well.

[6]Annotation Artifacts in Natural Language Inference Data

Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Richard Schwartz, S. Bowman, Noah A. Smith - 2018

6 papers in library cite

I love papers that show flaws in current methodology/datasets. Very nice for pointing it out!

[7]Enhanced LSTM for Natural Language Inference

Qinlang Chen, X. Zhu, Z. H. Ling, S. Wei, H. Jiang, D. Inkpen - 2017

5 papers in library cite

SotA NLI

[8]Allennlp: A Deep Semantic Natural Language Processing Platform

Matt Gardner, J. Grus, M. Neumann, Oyvind Tafjord, P. Dasigi, N. Liu, M. Peters, M. Schmitz, Luke Zettlemoyer - 2018

5 papers in library cite

Framework from alenai

[9]What You Can Cram Into a Single Vector: Probing Sentence Embeddings for Linguistic Properties

Alexis Conneau, German Kruszewski, G. Lample, L. Barrault, M. Baroni - 2018

2 papers in library cite

[10]Hypothesis Only Baselines in Natural Language Inference

A. Poliak, J. Naradowsky, A. Haldar, R. Rudinger, B. V. Durme - 2018

5 papers in library cite

Using only the hypothesis can solve the NLI task (it shouldn'!)

[11]Stress Test Evaluation for Natural Language Inference

A. Naik, A. Ravichander, N. M. Sadeh, C. P. Rose, Graham Neubig - 2018

4 papers in library cite

A bit flopped but I like the idea - they test if the system is truly making inferences.

[12]Human vs. Muppet: A Conservative Estimate of Human Performance on the GLUE Benchmark

Nikita Nangia, Samuel R. Bowman - 2019

3 papers in library cite

I like the premise but seems a bit flopped

[13]Breaking Nli Systems With Sentences That Require Simple Lexical Inferences

M. Glockner, V. Shwartz, Y. Goldberg - 2018

3 papers in library cite

[14]Analyzing the Behavior of Visual Question Answering Models

A. Agrawal, D. Batra, D. Parikh - 2016

2 papers in library cite

[15]Accurate Unlexicalized Parsing

Dan Klein, Christopher D. Manning - 2003

7 papers in library cite

[16]A Fast Unified Model for Parsing and Sentence Understanding

S. Bowman, J. Gauthier, Abhinav Rastogi, R. Gupta, C. Manning, Christopher Potts - 2016

5 papers in library cite

[17]Assessing the Ability of lstms to Learn Syntax-Sensitive Dependencies

Tal Linzen, E. Dupoux, Y. Goldberg - 2016

5 papers in library cite

[18]Entailment, Intensionality and Text Understanding

C. Condoravdi, D. Crouch, V. D. Paiva, R. Stolle, D. G. Bobrow - 2003

5 papers in library cite

[19]Assessing bert's Syntactic Abilities

Y. Goldberg - 2019

4 papers in library cite

[20]Collecting Diverse Natural Language Inference Problems for Sentence Representation Evaluation

A. Poliak, A. Haldar, R. Rudinger, J. E. Hu, Ellie Pavlick, A. S. White, B. V. Durme - 2018

4 papers in library cite

[21]Fine-Grained Analysis of Sentence Embeddings Using Auxiliary Prediction Tasks

Y. Adi, E. Kermany, Yonatan Belinkov, O. Lavi, Y. Goldberg - 2016

4 papers in library cite

[22]Colorless Green Recurrent Networks Dream Hierarchically

K. Gulordava, Piotr Bojanowski, E. Grave, Tal Linzen, M. Baroni - 2018

3 papers in library cite

[23]Non-Entailed Subsequences as a Challenge for Natural Language Inference

R. T. Mccoy, Tal Linzen - 2019

3 papers in library cite

[24]Targeted Syntactic Evaluation of Language Models

R. Marvin, Tal Linzen - 2018

3 papers in library cite

[25]Evaluating Compositionality in Sentence Embeddings

I. Dasgupta, Daniel Guo, Andreas Stuhlmuller, S. J. Gershman, N. D. Goodman - 2018

2 papers in library cite

[26]Inference Is Everything: Recasting Semantic Resources Into a Unified Evaluation Framework

A. White, P. Rastogi, K. Duh, B. Durme - 2017

2 papers in library cite

[27]Lexicosyntactic Inference in Neural Models

A. S. White, R. Rudinger, K. Rawlins, B. V. Durme - 2018

2 papers in library cite

[28]Revisiting the Poverty of the Stimulus: Hierarchical Generalization Without a Hierarchical Bias in Recurrent Neural Networks

R. T. Mccoy, Robert Frank, Tal Linzen - 2018

2 papers in library cite

[29]Analyzing Compositionality-Sensitivity of NLI Models

Y. Nie, Yuzhi Wang, Mohit Bansal - 2018

1 paper in library cites

[30]Assessing Composition in Sentence Vector Representations

A. Ettinger, A. Elgohary, C. Phillips, P. Resnik - 2018

1 paper in library cites

[31]Behavior Analysis of NLI Models: Uncovering the Influence of Three Factors on Robustness

I. Sanchez, J. Mitchell, Sebastian Riedel - 2018

1 paper in library cites

[32]Cambridge: Parser Evaluation Using Textual Entailment by Grammatical Relation Comparison

L. Rimell, S. Clark - 2010

1 paper in library cites

[33]Do Latent Tree Learning Models Identify Meaningful Structure in Sentences?

A. Williams, A. Drozdov, Samuel R. Bowman - 2018

1 paper in library cites

[34]Effects of Merely Local Syntactic Coherence on Sentence Processing

W. Tabor, B. Galantucci, D. Richardson - 2004

1 paper in library cites

[35]Making and Correcting Errors During Sentence Comprehension: Eye Movements in the Analysis of Structurally Ambiguous Sentences

L. Frazier, K. Rayner - 1982

1 paper in library cites

Missing year

[36]Natural Language Inference

B. Maccartney, Christopher D. Manning

1 paper in library cites

[37]Neural Models of Factuality

R. Rudinger, A. S. White, B. V. Durme - 2018

1 paper in library cites

[38]RNNs Implicitly Implement Tensor-Product Representations

R. T. Mccoy, Tal Linzen, E. Dunbar, P. Smolensky - 2019

1 paper in library cites

[39]Stress-Testing Neural Models of Natural Language Inference With Multiply-Quantified Sentences

A. Geiger, I. Cases, L. Karttunen, Christopher Potts - 2018

1 paper in library cites

[40]Syntactic/Semantic Structures for Textual Entailment Recognition

Y. Mehdad, A. Moschitti, F. M. Zanzotto - 2010

1 paper in library cites

[41]Teaching Syntax by Adversarial Distraction

Jeremy Kim, C. Malon, A. Kadav - 2018

1 paper in library cites

[42]Tense Manages to Predict Implicative Behavior in verbs

Ellie Pavlick, Chris Callison Burch - 2016

1 paper in library cites

[43]The Cognitive Basis for Linguistic Structures

T. G. Bever - 1970

1 paper in library cites

[44]The Fine Line Between Linguistic Generalization and Failure in Seq2seq-Attention Models

N. Weber, L. Shekhar, N. Balasubramanian - 2018

1 paper in library cites

[45]Thematic Roles Assigned Along the Garden Path Linger

K. Christianson, A. Hollingworth, J. F. Halliwell, F. Ferreira - 2001

1 paper in library cites

[46]Visual Concepts and Compositional Voting

J. Wang, Zhengyou Zhang, C. Xie, Y. Zhou, V. Premachandran, Jiacheng Zhu, L. Xie, A. Yuille - 2018

1 paper in library cites

Cited by

5

papers in your library

Cites

14

papers in your library

Read

on December 29, 2025

TBH I expected more. It's a very shallow analysis of the heuristics (but ok, props to them for making it simple). I think that ultimately they don't solve the problem.

Tags

Paper Aliases

No aliases