PROJECT 4

Robust (Controlled) Natural Language Generation with Structure‐Aware Equivariance Learning

Project Leader: Muhao Chen, Research Assistant Professor of Computer Science

Abstract: In this project, we will systematically tackle these challenges with a novel NLG framework based on structural equivariance learning. The framework will support a model‐agnostic fine‐tuning process of existing Seq2Seq generation models, in order to enhance their robustness and generalizability by ensuring two important properties: (i) Transformation‐invariance ensures the model to give consistent generation based on semantically equivalent but structurally distinct inputs; (ii) Structure‐awareness is ensured to help the NLG model more precisely describe relations of input data components. We will evaluate our method on a number of controlled NLG task, including table‐to‐text generation, concept‐to‐text generation and event time summarization. Specifically, we will verify the robustness of our method by creating harder versions of benchmarks for the aforementioned NLG tasks where various content‐neutral perturbations are free to be introduced to the input data. We will also compare our method with several techniques based on data augmentation and layout‐agnostic input, in order to show that equivariance learning is more effective to resolve the proposed problems.