面向文本的可解释预测方法研究

构建可解释的预测方法是当前深度学习领域中的热点和难点问题之一。可解释性能帮助人们理解，信任并有效控制预测模型，提高模型的泛化能力。本项目“面向文本的可解释预测方法”的目标正是建立同时具有预测能力和可解释性的深度学习方法，其主要存在三个方面的挑战：1）模型构造，建立本身具有可解释性的预测模型；2）知识融合，引入外部知识提升预测模型的解释能力；3）事后归因，事后对已经训练好的模型进行解释。针对这些挑战，本项目提出：1）结合粗粒度标注数据的自解释预测方法，利用易获取的粗粒度标注数据（如文档类别标签）训练模型，使其能够从输入中抽取关键证据解释预测结果；2）结合知识图谱的解释生成方法，通过引入外部知识图谱帮助模型生成合理且易于理解的解释理由；3）基于知识蒸馏的预训练模型解释方法，利用知识蒸馏实现知识从预训练模型到可解释模型的迁移，从而间接解释预训练模型的预测结果。

Developing explainable prediction methods is one of the significant issues in the field of deep learning. Explainability can help people understand, trust and effectively control the prediction models, and improve their generalization ability. The project "text-oriented explainable prediction method" aims to build deep learning models that have both predictive power and explainability. The main challenges lie in three aspects: 1) building prediction models that are inherently explainable; 2) improving the explainability of prediction models by introducing external knowledge; 3) providing post hoc explanations for the pre-trained models. To address these challenges, this project proposes: 1) self-explaining models that only rely on coarse-grained annotations (such as document labels) to train, and can extract key evidence from the inputs to explain the predictions; 2) a rationale generation model that combines external knowledge graph to produce reasonable and easy-to-understand explanations for the predictions; 3) a knowledge distillation based model for explaining pre-trained models, which transfers knowledge from the pre-trained models to an explainable model, so as to explain their predictions indirectly.