PaperlayerView exact revision on arXiv

Source-linked AI summary

NILE : Natural Language Inference with Faithful Natural Language Explanations

Sawan Kumar, Partha Talukdar

arXiv:2005.12116v1cs.CL

TL;DR

The paper asks whether NLI systems can retain high label accuracy while producing faithful natural-language explanations. NILE generates label-specific explanations and uses them to infer labels, with experiments supporting accurate, testable explanations and sensitivity-based faithfulness evaluation.

  • Problem

    NLI systems are expected to produce explanations that correlate with their decision-making while maintaining high label accuracy.

  • Method

    NILE generates natural-language explanations for each possible label and subsequently processes them to produce the final decision.

  • Results

    NILE demonstrates effectiveness in both label and explanation accuracy and supports accurate systems that provide testable explanations.

  • Takeaways & Limitations

    Faithfulness should be evaluated explicitly through sensitivity analysis, using task-specific probes to measure how explanations relate to model decisions.

  • Takeaways & Limitations

    NILE's design can still allow other biases to be learned, although probes can test for and help address them.

Abstract

from arXiv · show

The recent growth in the popularity and success of deep learning models on NLP classification tasks has accompanied the need for generating some form of natural language explanation of the predicted labels. Such generated natural language (NL) explanations are expected to be faithful, i.e., they should correlate well with the model's internal decision making. In this work, we focus on the task of natural language inference (NLI) and address the following question: can we build NLI systems which produce labels with high accuracy, while also generating faithful explanations of its decisions? We propose Natural-language Inference over Label-specific Explanations (NILE), a novel NLI method which utilizes auto-generated label-specific NL explanations to produce labels along with its faithful explanation. We demonstrate NILE's effectiveness over previously reported methods through automated and human evaluation of the produced labels and explanations. Our evaluation of NILE also supports the claim that accurate systems capable of providing testable explanations of their decisions can be designed. We discuss the faithfulness of NILE's explanations in terms of sensitivity of the decisions to the corresponding explanations. We argue that explicit evaluation of faithfulness, in addition to label and explanation accuracy, is an important step in evaluating model's explanations. Further, we demonstrate that task-specific probes are necessary to establish such sensitivity.

1 Introduction

NILE addresses whether NLI systems can maintain high label accuracy while generating faithful natural-language explanations. It generates label-specific explanations and processes them to infer labels, enabling explanation evaluation through sensitivity probes.

  • 1 Introduction: The paper asks whether NLI systems can produce faithful explanations of predicted labels while maintaining high accuracy.
  • 1 Introduction: NILE generates label-specific natural-language explanations and processes them to infer the final NLI label.The system also uses the premise-hypothesis pair alongside the explanations.
  • 1 Introduction: NILE is evaluated on label accuracy, explanation accuracy, and the sensitivity of predictions to generated explanations.The authors present sensitivity probes as a way to test explanation faithfulness.
  • 1 Introduction: NILE is proposed as an NLI system that naturally provides explanations for its decisions.

2 Related Work

Prior work includes post-hoc natural-language explanations and systems that generate explanations before predicting labels. NILE instead conditions prediction on label-specific explanations to support faithfulness and testability.

  • 2 Related Work: Existing explanation methods for deep learning and NLI are typically categorized as post-hoc explanations.
  • 2 Related Work: Camburu et al. generate a free-form explanation before inferring the NLI label, but the system can produce inconsistent explanations.
  • 2 Related Work: NILE conditions label prediction on generated natural-language explanations rather than treating explanations only as post-hoc outputs.
  • 2 Related Work: The paper frames natural-language explanations as a latent space for capturing compositional task structure and supporting task-specific reasoning.

3 Background

The paper builds NILE using established NLI datasets with human explanations and pretrained transformer models for generation and classification.

  • 3 Background: SNLI contains approximately 570K balanced premise-hypothesis pairs labeled entailment, neutral, or contradiction.
  • 3 Background: e-SNLI extends SNLI with crowdsourced natural-language explanations based on highlighted words in premise-hypothesis pairs.
  • 3 Background: Filtering non-informative explanations produces a training set of approximately 532K examples.
  • 3 Background: NILE uses GPT-2 medium for text generation and RoBERTa base for classification modules.

4 Natural-language Inference over Label-specific Explanations (NILE)

NILE uses label-specific candidate explanations followed by an Explanation Processor that selects the task label, enabling direct probing of whether predictions depend on their explanations. Its design combines a pipelined generator-processor architecture with alternative processing strategies and explicitly addresses biases that could arise from explanation form.

  • Overall architecture: NILE first generates label-specific explanations and then processes them, together with the premise and hypothesis, to produce the task label.The processor can instead access only generated explanations in NILE-PH.
  • Why this architecture: Unlike post-hoc and ExplainThen-Predict architectures, NILE allows prediction sensitivity to generated explanations to be tested by perturbing processor inputs.The pipeline keeps explanations as interpretable inputs rather than an unintelligible intermediate state.
  • Candidate Explanation Generators: Candidate explanation generators are trained separately on human-provided explanations, preventing degenerate explanations from being produced solely to improve task performance.Each label-specific generator is trained on examples carrying the corresponding label and explanation.
  • Explanation Processor architectures: The Aggregate processor produces supporting and opposing intermediate scores so it can handle missing or ambiguous explanations and contradictory evidence.For example, an entailment explanation asserting that a dog is a cat can support contradiction.
  • Processing Premise and Hypothesis: NILE’s premise-hypothesis processing can learn biases from explanation form, so the authors design probes to detect and subsequently fix such biases.The intended test is whether each premise-hypothesis pair and label-specific explanation forms a valid instance-explanation pair.

5 Experiments

The experiments compare NILE with baselines on in-domain and out-of-domain NLI, then test whether predictions respond appropriately to their generated explanations. NILE variants improve explanation quality and provide probes for evaluating faithfulness.

  • Experiment goals: The experiments assess final label performance, explanation accuracy, out-of-domain transfer, and faithfulness to generated explanations.In-domain evaluation uses SNLI, transfer evaluation trains on SNLI and tests on MNLI, and faithfulness is examined through sensitivity probes.
  • In-domain Results: NILE competes with or outperforms the ETPA baseline in explanation accuracy while incurring only a small drop in label accuracy.NILE variants produce more correct explanations than ETPA, while NILE-PH:Append, NILE, and NILE-NS improve label accuracy over ETPA.
  • Transfer to Out-of-domain NLI: All NILE variants improve explanation quality on out-of-domain MNLI, including correct explanations and correct explanations among correct predictions.Training and model selection are performed on SNLI, while evaluation uses matched and mismatched MNLI development sets.
  • Evaluating Faithfulness using Sensitivity Analysis: NILE and NILE-NS use erasure-based comprehensiveness and sufficiency measures to examine how predictions respond to removing instances or explanations.The results suggest comprehensive explanations for both systems, while sufficiency alone may reflect correlations that do not otherwise exist in the system.
  • Evaluating Faithfulness using Sensitivity Analysis: Shuffling same-label explanations reveals a weaker link between NILE-NS predictions and explanations, whereas NILE behaves more expectedly under perturbation.The authors identify this as a task-specific probe of sensitivity and note that NILE’s architecture helps identify and fix associated biases.

6 Conclusion

The conclusion presents NILE as an NLI system that combines label prediction with natural language explanations and evaluates those explanations for faithfulness. The experiments support testable explanations and the need for task-specific sensitivity probes.

  • Conclusion: NILE generates NLI labels together with natural language explanations for the predicted labels.The system is evaluated through both label and explanation accuracy.
  • Conclusion: NILE supports the hypothesis that accurate systems can produce testable natural language explanations of their decisions.Faithfulness is evaluated by examining how explanations correlate with model decision making.
  • Conclusion: The paper argues that faithfulness should be evaluated explicitly alongside label and explanation accuracy.NILE’s faithfulness evaluation uses sensitivity analysis.
  • Conclusion: Task-specific probes are necessary to measure sensitivity between generated explanations and model predictions.The conclusion identifies sensitivity probing as a requirement for evaluating explanation faithfulness.

A Experimental Setup

The experimental setup fine-tunes GPT2-medium for explanation generation and RoBERTa-base models for NLI, using fixed training procedures and multiple random seeds for evaluation.

  • Reported Variability: Table 5 reports mean and standard deviation for NILE-NS:Independent label accuracy on the SNLI Dev set.Excluding the bad initialization with seed 219 yields mean 91.41 and standard deviation 0.20.
  • Model Training: GPT2-medium and RoBERTa-base models are fine-tuned for explanation generation and NLI experiments, respectively.The implementation uses pretrained models and fixed training settings across experiments.
  • Model Selection: RoBERTa label results use model selection across five random seeds based on SNLI Dev label accuracy.The reported label accuracies include mean and standard deviation across the five runs.
  • Hardware: Experiments run on GeForce GTX 1080 Ti GPUs with batch sizes adjusted to fit approximately 12GB of GPU memory.The batch size is selected as the largest multiple of 16 that fits the available memory.
  • Hyper-parameters: GPT2 explanation generation uses a maximum sequence length of 128, batch size 2, learning rate 5e-5, and greedy decoding.The setup also uses Adam epsilon 1e-8, maximum gradient norm 1.0, and seed 42.

B Generated Explanations

The appendix lists label-specific explanations generated for development examples from both SNLI and MNLI.

  • Generated Explanations: Generated label-specific explanations are provided for five development examples from SNLI and MNLI.Each example includes the premise, hypothesis, and gold label alongside the generated explanations.

B.1 SNLI

The SNLI examples pair image-caption premises with hypotheses and generate label-specific explanations for entailment, contradiction, and neutrality. These explanations illustrate how each relation can be justified differently.

  • B.1 SNLI: Neutral explanations emphasize that the premise does not establish the additional claim in the hypothesis.The examples state that embracing does not imply being sisters, and touching a plane propeller does not imply playing with a ball.
  • B.1 SNLI: SNLI examples classify premise–hypothesis pairs as entailment, contradiction, or neutral.The examples include image-caption scenarios such as children washing their hands, amusement-park queues, and bicycle racing.
  • B.1 SNLI: Entailment explanations connect the hypothesis to information explicitly present or directly implied by the premise.For example, “Two young children” is mapped to “two kids,” and “blue jerseys” to “numbered jerseys.”
  • B.1 SNLI: Contradiction explanations identify incompatible actions, events, or meanings between the premise and hypothesis.Examples contrast waiting to ride at an amusement park with watching a movie, and competing in a race with catching fish.

B.2 MNLI

The MNLI examples similarly pair premises with hypotheses and label-specific explanations. They use paraphrase, negation, and underdetermination to distinguish entailment, contradiction, and neutrality.

  • B.2 MNLI: MNLI examples classify premise–hypothesis pairs as entailment, contradiction, or neutral.The examples cover informal statements about rights, military-funded education, welcome statements, supplementation, and apologies.
  • B.2 MNLI: Neutral explanations state that the premise does not warrant the hypothesis or its additional assumptions.Examples note that being told someone is unwelcome does not support the opposite inference, and that new benefits need not be the newest benefits.
  • B.2 MNLI: Entailment explanations treat paraphrases or explicitly implied category relations as support for the hypothesis.Examples equate “supplement” with adding something extra and “friends” with people.
  • B.2 MNLI: Contradiction explanations identify direct opposition between the premise and hypothesis.The examples contrast being sorry with having no apologies and state that military payment conflicts with the hypothesis that it did not occur.
Loading 2005.12116v1…