Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Supervising Model Attention with Human Explanations for Robust Natural Language Inference

Joe Stacey, Yonatan Belinkov, Marek Rei

TL;DR

Natural Language Inference models often exploit dataset biases, hindering generalization. This paper proposes supervision of attention using human explanations from e-SNLI to guide the model toward explanation-relevant words, aiming to improve robustness. By evaluating several strategies, including supervising existing self-attention heads and adding an extra attention layer, and combining free-text explanations with highlighted tokens, the authors demonstrate substantial in-distribution and out-of-distribution gains, achieving a new SNLI state-of-the-art with DeBERTa. The results show attention shifts toward premise words and away from punctuation/stop-words, improving interpretability and robustness in NLI. Overall, explanation-guided attention offers a simple yet effective route to more robust and interpretable NLI systems.

Abstract

Natural Language Inference (NLI) models are known to learn from biases and artefacts within their training data, impacting how well they generalise to other unseen datasets. Existing de-biasing approaches focus on preventing the models from learning these biases, which can result in restrictive models and lower performance. We instead investigate teaching the model how a human would approach the NLI task, in order to learn features that will generalise better to previously unseen examples. Using natural language explanations, we supervise the model's attention weights to encourage more attention to be paid to the words present in the explanations, significantly improving model performance. Our experiments show that the in-distribution improvements of this method are also accompanied by out-of-distribution improvements, with the supervised models learning from features that generalise better to other NLI datasets. Analysis of the model indicates that human explanations encourage increased attention on the important words, with more attention paid to words in the premise and less attention paid to punctuation and stop-words.

Supervising Model Attention with Human Explanations for Robust Natural Language Inference

TL;DR

Natural Language Inference models often exploit dataset biases, hindering generalization. This paper proposes supervision of attention using human explanations from e-SNLI to guide the model toward explanation-relevant words, aiming to improve robustness. By evaluating several strategies, including supervising existing self-attention heads and adding an extra attention layer, and combining free-text explanations with highlighted tokens, the authors demonstrate substantial in-distribution and out-of-distribution gains, achieving a new SNLI state-of-the-art with DeBERTa. The results show attention shifts toward premise words and away from punctuation/stop-words, improving interpretability and robustness in NLI. Overall, explanation-guided attention offers a simple yet effective route to more robust and interpretable NLI systems.

Abstract

Natural Language Inference (NLI) models are known to learn from biases and artefacts within their training data, impacting how well they generalise to other unseen datasets. Existing de-biasing approaches focus on preventing the models from learning these biases, which can result in restrictive models and lower performance. We instead investigate teaching the model how a human would approach the NLI task, in order to learn features that will generalise better to previously unseen examples. Using natural language explanations, we supervise the model's attention weights to encourage more attention to be paid to the words present in the explanations, significantly improving model performance. Our experiments show that the in-distribution improvements of this method are also accompanied by out-of-distribution improvements, with the supervised models learning from features that generalise better to other NLI datasets. Analysis of the model indicates that human explanations encourage increased attention on the important words, with more attention paid to words in the premise and less attention paid to punctuation and stop-words.

Paper Structure

This paper contains 21 sections, 6 equations, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Training NLI Models with Explanations
  4. Training with Explanations Beyond NLI
  5. Creating More Robust NLI Models
  6. Attention Supervision Method
  7. Supervising Self-Attention Layers
  8. Selecting Attention Heads for Supervision
  9. Supervising an Additional Attention Layer
  10. Experimental Setup and Evaluation
  11. Experiments
  12. Performance in and out of Distribution
  13. Experiments with DeBERTa
  14. Comparing Results with Prior Work
  15. Choosing Which Explanations to Use and Which Heads to Supervise
  16. ...and 6 more sections

Figures (4)

  • Figure 1: An example of using a free text explanation to identify important words in the premise and hypothesis. In this case the words dog, sleeping and swims have been identified from the explanation.
  • Figure 2: An example of how the attention loss is calculated when supervising an existing self-attention layer.
  • Figure 3: Accuracy when supervising each of the attention heads in turn, compared to the baseline with no supervision, supervising all heads and supervising the top 3 heads.
  • Figure 4: Average attention from the [CLS] token in the baseline and when we are supervising each attention head. Both models incorrectly predicted the first example as being neutral. The second example was correctly labeled by the supervised model (neutral), while the baseline model incorrectly predicted contradiction. The e-SNLI free-text explanations for the sentences include: 'One must be happy in order to have a big grin' and 'Just because it is a person does not mean it is a child'.