Social Commonsense Reasoning with Multi-Head Knowledge Attention

TL;DR

This paper introduces a multi-head knowledge attention model that enhances social commonsense reasoning by integrating semi-structured inference rules into transformer models, improving performance on reasoning tasks like abductive inference and counterfactual prediction.

Contribution

The paper presents a novel multi-head knowledge attention mechanism that encodes inference rules and improves reasoning capabilities in transformer-based models, especially for social commonsense tasks.

Findings

Model outperforms RoBERTa on reasoning tasks

Counterfactual reasoning improves abductive inference

Robustness validated through knowledge perturbation

Abstract

Social Commonsense Reasoning requires understanding of text, knowledge about social events and their pragmatic implications, as well as commonsense reasoning skills. In this work we propose a novel multi-head knowledge attention model that encodes semi-structured commonsense inference rules and learns to incorporate them in a transformer-based reasoning cell. We assess the model's performance on two tasks that require different reasoning skills: Abductive Natural Language Inference and Counterfactual Invariance Prediction as a new task. We show that our proposed model improves performance over strong state-of-the-art models (i.e., RoBERTa) across both reasoning tasks. Notably we are, to the best of our knowledge, the first to demonstrate that a model that learns to perform counterfactual reasoning helps predicting the best explanation in an abductive reasoning task. We validate the robustness of the model's reasoning capabilities by perturbing the knowledge and provide qualitative analysis on the model's knowledge incorporation capabilities.