[Paper Review] Probing Neural Network Comprehension of Natural Language Arguments

Timothy Niven, Hung‐Yu Kao|arXiv (Cornell University)|Jul 17, 2019

Topic Modeling27 references34 citations

TL;DR

The paper shows that BERT’s strong ARCT performance is largely due to exploitation of spurious cues, and introduces an adversarial dataset where models perform at random, arguing for a robust evaluation standard.

ABSTRACT

We are surprised to find that BERT's peak performance of 77% on the Argument Reasoning Comprehension Task reaches just three points below the average untrained human baseline. However, we show that this result is entirely accounted for by exploitation of spurious statistical cues in the dataset. We analyze the nature of these cues and demonstrate that a range of models all exploit them. This analysis informs the construction of an adversarial dataset on which all models achieve random accuracy. Our adversarial dataset provides a more robust assessment of argument comprehension and should be adopted as the standard in future work.

Motivation & Objective

• Assess whether BERT and other models truly comprehend natural language arguments in ARCT beyond dataset artifacts.
• Identify and quantify spurious statistical cues that models exploit in ARCT.
• Develop an adversarial dataset to provide a robust evaluation and baseline for future work.

Proposed method

• Replicate and extend baseline models (BoV, BiLSTM, GIST, Botschen et al.) and BERT on ARCT.
• Probe model decisions to identify reliance on lexical cues in warrants (e.g., the word 'not').
• Define cue metrics (productivity, coverage) to quantify cue usefulness across data.
• Create an adversarial ARCT dataset by negating claims and flipping labels to mirror cue distributions.
• Evaluate models on adversarial data to measure true argument comprehension beyond cues.

Experimental results

Research questions

• RQ1 Do state-of-the-art models, particularly BERT, rely on spurious cues rather than genuine argument comprehension in ARCT?
• RQ2 Can an adversarial data transformation eliminate cue-driven performance and reveal true reasoning capabilities?
• RQ3 How do different model architectures fare when statistical cues are balanced or mirrored across labels?
• RQ4 What is the impact of training on adversarial data for model robustness on ARCT?

Key findings

• BERT’s peak 77% accuracy is largely explained by exploiting dataset cues, not genuine understanding.
• Probing shows warrants alone yield up to 71% accuracy, with additional reason/claim cues accounting for remaining gains.
• An adversarial dataset mirroring cue distributions around labels reduces BERT performance to random (53% max).
• Across models, performance on original data deteriorates to near random levels on adversarial data, indicating reliance on spurious cues.
• Adversarial training and evaluation reveal that robust argument comprehension remains unrealized by current models.