Language Models of Code are Few-Shot Commonsense Learners

TL;DR

This paper demonstrates that pre-trained code generation language models excel at structured commonsense reasoning tasks when framed as code generation, outperforming natural language models in few-shot settings.

Contribution

The paper introduces a novel approach of framing structured commonsense reasoning as code generation, leveraging code-trained LMs to improve performance over traditional natural language models.

Findings

Code LMs outperform natural language LMs in structured reasoning tasks.

Framing reasoning tasks as code generation improves few-shot learning performance.

Pre-trained code models generalize well to non-code reasoning tasks.

Abstract

We address the general task of structured commonsense reasoning: given a natural language input, the goal is to generate a graph such as an event -- or a reasoning-graph. To employ large language models (LMs) for this task, existing approaches ``serialize'' the output graph as a flat list of nodes and edges. Although feasible, these serialized graphs strongly deviate from the natural language corpora that LMs were pre-trained on, hindering LMs from generating them correctly. In this paper, we show that when we instead frame structured commonsense reasoning tasks as code generation tasks, pre-trained LMs of code are better structured commonsense reasoners than LMs of natural language, even when the downstream task does not involve source code at all. We demonstrate our approach across three diverse structured commonsense reasoning tasks. In all these natural language tasks, we show that using our approach, a code generation LM (CODEX) outperforms natural-LMs that are fine-tuned on the target task (e.g., T5) and other strong LMs such as GPT-3 in the few-shot setting.