[Paper Review] Natural Language Inference with a Human Touch: Using Human Explanations to Guide Model Attention.

Joe Stacey, Yonatan Belinkov+1 more|arXiv (Cornell University)|Apr 16, 2021

Topic Modeling42 references4 citations

TL;DR

This paper proposes using human-generated natural language explanations to guide neural model attention in Natural Language Inference (NLI), improving both in-domain and out-of-domain generalization. By supervising attention weights to focus on explanation-relevant words, the model learns more robust, human-aligned features, reducing reliance on dataset biases.

ABSTRACT

Natural Language Inference (NLI) models are known to learn from biases and artefacts within their training data, impacting how well the models generalise to other unseen datasets. While previous de-biasing approaches focus on preventing models learning from these biases, we instead provide models with information about how a human would approach the task, with the aim of encouraging the model to learn features that will generalise better to out-of-domain datasets. Using natural language explanations, we supervise a model's attention weights to encourage more attention to be paid to the words present in these explanations. For the first time, we show that training with human generated explanations can simultaneously improve performance both in-distribution and out-of-distribution for NLI, whereas most related work on robustness involves a trade-off between the two. Training with the human explanations encourages models to attend more broadly across the sentences, paying more attention to words in the premise and less attention to stop-words and punctuation. The supervised models attend to words humans believe are important, creating more robust and better performing NLI models.

Motivation & Objective

• To address the poor generalization of NLI models caused by dataset biases and artifacts.
• To improve model robustness without sacrificing in-domain performance, overcoming the typical trade-off in robustness research.
• To guide model attention toward words deemed important by humans, using natural language explanations as supervision.
• To encourage broader attention across sentences, reducing over-reliance on stop words and punctuation.
• To create NLI models that learn features aligned with human reasoning for better generalization.

Proposed method

• Human annotators provide natural language explanations for NLI premise-hypothesis pairs, identifying key reasoning elements.
• The model's attention weights are supervised to align with the words present in these human explanations.
• A differentiable attention regularization loss is applied, encouraging attention to focus on explanation-relevant words.
• The model is fine-tuned end-to-end with both NLI classification loss and explanation-guided attention regularization.
• Attention is encouraged to distribute more broadly across sentences, reducing fixation on trivial tokens like stop words.
• The approach integrates human reasoning signals directly into the attention mechanism during training.

Experimental results

Research questions

• RQ1 Can human explanations improve both in-domain and out-of-domain performance in NLI models?
• RQ2 Does attention guidance via human explanations reduce reliance on dataset biases and artifacts?
• RQ3 Can attention be effectively regularized using natural language explanations to improve model robustness?
• RQ4 Does the model attend more to words highlighted in human explanations compared to baseline models?
• RQ5 Does the attention distribution become more human-aligned and less focused on punctuation and stop words?

Key findings

• Training with human explanations improves NLI performance on both in-domain and out-of-domain datasets simultaneously, avoiding the typical robustness-generalization trade-off.
• The model attends more to words present in human explanations, indicating successful alignment with human reasoning patterns.
• Attention is distributed more broadly across sentences, reducing over-concentration on stop words and punctuation.
• The model learns more robust features that generalize better to unseen domains.
• The attention mechanism becomes more interpretable and aligned with human-annotated reasoning cues.
• The approach demonstrates that incorporating human-structured explanations can guide models toward more generalizable and human-like reasoning.