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. This is an updated version of our ACL paper.

No abstract

In this paper, we describe our entry in the gendered pronoun resolution competition which achieved fourth place without data augmentation. Our method is an ensemble system of BERTs which resolves co-reference in an interaction space. We report four insights from our work: BERT's representations involve significant redundancy; modeling interaction effects similar to natural language inference models is useful for this task; there is an optimal BERT layer to extract representations for pronoun resolution; and the difference between the attention weights from the pronoun to the candidate entities was highly correlated with the correct label, with interesting implications for future work.

The Argument Reasoning Comprehension Task requires significant language understanding and complex reasoning over world knowledge. We focus on transfer of a sentence encoder to bootstrap more complicated models given the small size of the dataset. Our best model uses a pre-trained BiLSTM to encode input sentences, learns task-specific features for the argument and warrants, then performs independent argument-warrant matching. This model achieves mean test set accuracy of 64.43%. Encoder transfer yields a significant gain to our best model over random initialization. Independent warrant matching effectively doubles the size of the dataset and provides additional regularization. We demonstrate that regularization comes from ignoring statistical correlations between warrant features and position. We also report an experiment with our best model that only matches warrants to reasons, ignoring claims. Relatively low performance degradation suggests that our model is not necessarily learning the intended task.

We report the results of preliminary investigations into the relationship between linguistic alignment and dialogical argumentation at the level of discourse acts. We annotated a proof of concept dataset with illocutions and transitions at the comment level based on Inference Anchoring Theory. We estimated linguistic alignment across discourse acts and found significant variation. Alignment features calculated at the dyad level are found to be useful for detecting a range of argumentative discourse acts.