• Publications
  • Influence
Towards Faithful Neural Table-to-Text Generation with Content-Matching Constraints
TLDR
Automatic and human evaluations show that the novel Transformer-based generation framework proposed can significantly outperform state-of-the-art by a large margin.
Sectorization and Configuration Transition in Airspace Design
TLDR
An undirected graph cut-based approach that employs a memetic local search-embedded constrained evolution algorithm, NSGA-II, to generate nondominated airspace configurations and a new concave hull-based method to automatically depict sector boundaries is proposed.
Repulsive Attention: Rethinking Multi-head Attention as Bayesian Inference
TLDR
This paper proposes a non-parametric approach that explicitly improves the repulsiveness in multi-head attention and consequently strengthens model's expressiveness and provides theoretical inspirations on the not-well-understood questions: why and how one uses multi- head attention.
Adaptive Transfer Learning on Graph Neural Networks
TLDR
This work proposes a new transfer learning paradigm on GNNs which could effectively leverage self-supervised tasks as auxiliary tasks to help the target task and significantly improve the performance compared to state-of-the-art methods.
Repulsive Bayesian Sampling for Diversified Attention Modeling
TLDR
For the first time, a novel understanding of multi-head attention from a Bayesian-sampling perspective is provided, based on particle-optimization sampling methods, and non-parametric approaches that explicitly improve the diversity of Multi-head Attention are proposed.
Deep Semantic Compliance Advisor for Unstructured Document Compliance Checking
TLDR
Deep Semantic Compliance Advisor (DSCA) is an unstructured document compliance checking platform which provides multi-level semantic comparison by deep learning algorithms and a Graph Neural Network based syntactic sentence encoder is proposed to capture the complicate syntactic and semantic clues of the statement sentences.
A smart airspace sectorization approach based on spectral clustering and NSGA-II
TLDR
A smart sectorization approach which is based on spectral clustering and NSGA-II is proposed, which can obtain better solutions with less number of sectors, no constraint violation and more reasonable workload allocation.
Understanding the Generalization Benefit of Model Invariance from a Data Perspective
TLDR
The generalization benefit of model invariance is studied by introducing the sample cover induced by transformations, i.e., a representative subset of a dataset that can approximately recover the whole dataset using transformations.
Guess First to Enable Better Compression and Adversarial Robustness
TLDR
This paper proposes a bio-inspired classification framework in which model inference is conditioned on label hypothesis and provides a class of training objectives and an information bottleneck regularizer which utilizes the advantage that label information can be discarded during inference.