Interactive Model with Structural Loss for Language-based Abductive Reasoning
This work addresses the problem of improving reasoning in natural language inference for AI systems, but it is incremental as it builds on existing pretrained models with a novel loss and interaction mechanism.
The paper tackles the abductive natural language inference task by proposing an interactive model with a structural loss that groups hypotheses rather than ranking them, achieving about 1% higher accuracy and 5% higher AUC compared to previous methods.
The abductive natural language inference task ($α$NLI) is proposed to infer the most plausible explanation between the cause and the event. In the $α$NLI task, two observations are given, and the most plausible hypothesis is asked to pick out from the candidates. Existing methods model the relation between each candidate hypothesis separately and penalize the inference network uniformly. In this paper, we argue that it is unnecessary to distinguish the reasoning abilities among correct hypotheses; and similarly, all wrong hypotheses contribute the same when explaining the reasons of the observations. Therefore, we propose to group instead of ranking the hypotheses and design a structural loss called ``joint softmax focal loss'' in this paper. Based on the observation that the hypotheses are generally semantically related, we have designed a novel interactive language model aiming at exploiting the rich interaction among competing hypotheses. We name this new model for $α$NLI: Interactive Model with Structural Loss (IMSL). The experimental results show that our IMSL has achieved the highest performance on the RoBERTa-large pretrained model, with ACC and AUC results increased by about 1\% and 5\% respectively.