Stochastic Answer Networks for Natural Language Inference
Xiaodong Liu†, Kevin Duh‡ and Jianfeng Gao† † Microsoft Research, Redmond, WA, USA ‡ Johns Hopkins University, Baltimore, MD, USA
Abstract
We utilize a stochastic answer network (SAN) to explore multi-step inference strategies in Natural Language Inference. Rather than directly predicting the results given the inputs, the model maintains a state and iteratively refines its predictions. This can potentially model more complex inferences than the existing single-step inference methods. Our experiments show that SAN achieves state-of-the-art results on four benchmarks: Stanford Natural Language Inference (SNLI), MultiGenre Natural Language Inference (MultiNLI), SciTail, and Quora Question Pairs datasets.
中文速览
自然语言推理（Natural Language Inference，NLI）要求模型判断两段文字之间的逻辑关系（如蕴含、矛盾、中立），难点在于有些推理需要多个步骤才能完成，而现有方法几乎都是一步到位地给出答案。研究者将机器阅读理解领域的随机答案网络（Stochastic Answer Network，SAN）引入NLI任务，让模型维护一个内部状态并在前提和假设之间反复迭代、逐步修正预测，而非直接输出结论；训练时还加入随机预测dropout，最终通过多步输出投票得到最终答案。在SNLI、MultiNLI、SciTail和Quora Question Pairs四个基准上，SAN均达到当时最优或极具竞争力的水平，尤其在"反义词"和"长句"等困难类别上提升显著。这项工作证明了多步推理机制能有效增强模型处理复杂语言逻辑的能力，也为后续结合预训练语言模型进一步提升提供了可扩展的框架。
原文 arXiv:1804.07888;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1804.07888v2