Source-linked AI summary
Probing Neural Network Comprehension of Natural Language Arguments
Timothy Niven, Hung-Yu Kao
TL;DR
ARCT evaluates whether models can connect claims and reasons using implicit warrants, yet BERT nearly matches untrained humans. The paper probes this performance, finds it fully attributable to spurious statistical cues, and constructs an adversarial dataset where models return to random accuracy. It recommends the adversarial dataset for more robust ARCT evaluation.
Problem
ARCT requires world knowledge to connect claims, reasons, and implicit warrants, making BERT’s near-human performance difficult to interpret as argument comprehension.
Method
The paper probes models using cue-isolation experiments and constructs adversarial examples by negating claims and inverting labels to mirror cue distributions.
Results
BERT’s 77% peak performance is entirely accounted for by spurious cues, while its adversarial-test peak falls to 53% and mean and median reach 50%.
Takeaways & Limitations
The adversarial dataset provides a more robust assessment of argument comprehension and should become the standard for future ARCT work.
Takeaways & Limitations
Training on negated data can still produce above-random performance by exploiting claim-warrant statistics, so the reported setup uses differing training and test heuristic distributions.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper introduces ARCT as warrant-based argument inference and investigates why BERT performs unexpectedly well despite the task requiring implicit world knowledge. It attributes this performance to spurious statistical cues and proposes an adversarial dataset that removes them.
- Argumentation mining identifies claims and reasons that support or attack them, but determining argumentative relations can be difficult even for humans.
- Warrants provide implicit world knowledge that licenses inferences between reasons and claims, so systems must reason with and discover them.
- ARCT supplies a claim and reason, then asks models to choose the correct warrant over an alternative designed so R ∧A →¬C.
- All but one SemEval system failed to exceed 60% binary accuracy, underscoring that ARCT requires additional world knowledge beyond the supplied warrants.
- BERT reached 77% test accuracy, only three points below the average untrained human baseline, motivating analysis of what it learned.
- Probing found that BERT exploits warrant cue words, while claim-negation and label inversion mirror cue distributions and reduce adversarial-test performance to 53%.The adversarial dataset is proposed as a more robust standard for future ARCT evaluation.
2 Task Description and Baselines
The experiments classify which of two candidate warrants is correct for an argument using shared model parameters, independently computed logits, and a softmax decision.
- Each dataset point contains a claim, reason, and two candidate warrants randomly assigned binary labels, with the correct warrant indicated by y(i).
- Shared parameters θ classify each warrant independently with the argument to produce logits for the two candidates.
- The two logits are concatenated, converted by softmax into warrant probabilities, and resolved by selecting the highest-probability candidate.
- Baselines include BoV, BiLSTM, GIST, Botschen et al.’s model, and human performance.
- BERT pairs the claim and reason as one text segment with each warrant and passes the final CLS vector through a linear layer to obtain logits.
3 BERT
BERT was fine-tuned for warrant classification, but its results varied substantially across random seeds because the small ARCT training set sometimes produced degenerate training runs.
- BERT fine-tunes an architecture that pairs the claim-reason segment independently with each warrant and derives logits from the final CLS vector.
- BERT encountered near-random validation and test accuracy in 5/20 runs, with training accuracy also below 80%.
- After removing degenerate runs, BERT’s mean was 71.6 ± 0.04 and its median was 71.2%, but the authors discard these results as not meaningful.
- The analysis therefore focuses on BERT’s peak performance of 77% to argue that its apparent success is explained by spurious cues.
4 Statistical Cues
ARCT contains warrant-level linguistic artifacts that correlate unevenly with labels, creating exploitable statistical cues. The paper characterizes these cues using productivity and coverage.
- Cue Metrics: The paper measures each cue’s applicability, productivity, and coverage to quantify its usefulness and prevalence.Applicability counts cases where a cue occurs with one label but not the other; productivity measures correct predictions among applicable cases, while coverage is α_k/n.
- Cue Findings: The warrant cue “not” predicts the correct label 61% of the time and applies to 64% of the dataset.These figures describe its productivity and coverage across the whole dataset.
- Cue Metrics: A cue is useful in an m-label task when its productivity exceeds 1/m.Under this condition, a learner can benefit from using the cue’s presence.
- Cue Findings: Other productive cues include frequent unigrams such as “is,” “do,” and “are,” plus bigrams such as “will not” and “cannot.”The strongest cue is less productive on the test set but is one of many cues available to learners.
5 Probing Experiments
Probing experiments test whether BERT’s performance depends on argument structure or on distributional cues available from partial inputs. The results attribute its peak accuracy entirely to spurious cues.
- Experimental Design: The experiments train models with warrants alone, or with warrants paired with reasons or claims, to isolate distributional information.These reduced-input settings no longer provide a complete argument to match with a warrant.
- Results: 71% accuracy is achieved by BERT using warrants alone, leaving six points below its 77% peak.Adding reasons accounts for four points and adding claims accounts for two more.
- Results: The combined gains from reason- and claim-based cues account for the remaining six percentage points of BERT’s peak performance.The reported gains are four points for (R, W) over (W) and two points for (C, W).
6 Adversarial Test Set
The adversarial test set mirrors warrant-level cues across labels by pairing each example with a claim-negated, label-inverted copy. On this set, BERT’s performance approaches random accuracy.
- Construction: Each adversarial example negates the claim and inverts the label, then combines the copy with the original data.Because R ∧ A → ¬C, this transformation mirrors cue distributions around both labels.
- Results: 53% is BERT’s peak accuracy on the adversarial test set, while its mean and median are both 50%.Training used the original data, with adversarial validation and testing; comparable results were obtained without training augmentation.
- Caveat: Training on negated data can still produce above-random performance by exploiting statistics shared between claims and warrants.The experimental setup breaks these heuristics by using different distributions in training and testing.
- Construction: The transformation removes the major warrant-based statistical signal by mirroring linguistic artifacts around the labels.The original and adversarial points retain the same label assignment for the correct and alternative warrants.
7 Related Work
Prior ARCT systems used transfer learning from NLI datasets and, in one case, FrameNet knowledge. The paper argues that adversarial transformation should support more robust evaluation across the dataset.
- Prior ARCT Systems: Successful prior ARCT systems transferred learning from NLI datasets using models such as ESIM and InferSent.Adding FrameNet knowledge produced modest performance gains in later work.
- Dataset Artifacts: The paper identifies statistical-cue controls as relevant to broader work on dataset artifacts in vision and NLP.Related studies also probe whether models can classify examples from restricted inputs such as hypotheses, reasons, or warrants.
- Evaluation: The authors argue that the adversarial transformation must be applied to the entire dataset to eliminate solutions based on spurious cues.They distinguish this from earlier use of a similar transformation that yielded random accuracy on the original test set.
8 Conclusion
The authors show that BERT’s apparent argument-comprehension ability is explained by spurious statistical cues, while BERT remains a strong learner of subtle distributional information. They argue that adversarial evaluation should become standard for ARCT and motivate broader investigation of such cues in NLP.
- 8 Conclusion: BERT’s maximum performance fell from three points below the average untrained human baseline to essentially random on the adversarial dataset.This shows that the original performance was not evidence of argument comprehension.
- 8 Conclusion: BERT relied less on the strongest cue word than BoV and BiLSTM, indicating that it exploited subtler joint distributional information.The authors characterize BERT as a very strong learner despite its reliance on spurious statistics.
- 8 Conclusion: The adversarial dataset should become the standard for future ARCT work to provide a more robust evaluation and spur more productive research.The authors also call for further research into the extent of spurious-statistics exploitation in NLP.