Enhancing Natural Language Representation with Large-Scale Out-of-Domain Commonsense

TL;DR

This paper introduces OK-Transformer, a method that effectively incorporates large-scale out-of-domain commonsense knowledge into language models to improve text understanding without catastrophic forgetting.

Contribution

The paper proposes OK-Transformer, a novel approach that seamlessly integrates out-of-domain commonsense into existing Transformer models without additional pre-training.

Findings

Improves performance in commonsense reasoning tasks.

Enhances general text classification accuracy.

Effective in low-resource commonsense scenarios.

Abstract

We study how to enhance text representation via textual commonsense. We point out that commonsense has the nature of domain discrepancy. Namely, commonsense has different data formats and is domain-independent from the downstream task. This nature brings challenges to introducing commonsense in general text understanding tasks. A typical method of introducing textual knowledge is continuing pre-training over the commonsense corpus. However, it will cause catastrophic forgetting to the downstream task due to the domain discrepancy. In addition, previous methods of directly using textual descriptions as extra input information cannot apply to large-scale commonsense. In this paper, we propose to use large-scale out-of-domain commonsense to enhance text representation. In order to effectively incorporate the commonsense, we proposed OK-Transformer (\underline{O}ut-of-domain \underline{K}nowledge enhanced \underline{Transformer}). OK-Transformer effectively integrates commonsense descriptions and enhances them to the target text representation. In addition, OK-Transformer can adapt to the Transformer-based language models (e.g. BERT, RoBERTa) for free, without pre-training on large-scale unsupervised corpora. We have verified the effectiveness of OK-Transformer in multiple applications such as commonsense reasoning, general text classification, and low-resource commonsense settings.