DON’TSETTLE FORAVERAGE, GO FOR THEMAX:FUZZYSETS ANDMAX-POOLEDWORDVECTORS

BERTScore leverages the pre-trained contextual embeddings from BERT and matches words in candidate and reference sentences by cosine similarity. It has been shown to correlate with human judgment on sentence-level and system-level evaluation. Moreover, BERTScore computes precision, recall, and F1 measure, which can be useful for evaluating different language generation tasks.

BERTSCORE: EVALUATINGTEXTGENERATION WITHBERT

Simple Evaluation Metrics for Diversified Search Results

Search Result Diversity for Informational Queries

子空间学习法通常假设源域和目标域数据在变换后的子空间中会有着相似的分布。我

领域自适应问题是迁移学习的研究内容之一，它侧重于解决特征空间一致、类别空间一致，仅特征分布不一致的问题。

Ptr-Nets can be used to learn approximate solutions to three challenginggeometric problems – finding planar convex hulls, computing Delaunay triangu-lations, and the planar Travelling Salesman Problem

uij=vTtanh(W1ej+W2di)

PAY LESS ATTENTIONWITHLIGHTWEIGHT ANDDYNAMICCONVOLUTIONS

Hierarchical attention weighting (

Attention weighting for documents (Wang et al, ’16)

The difficulty of generating1In this paper, we use the termsinstance,sample, andexampleinterchangeably.Decision boundaryAdversarial exampleBorderline instanceFigure 1: A high-level illustration of Deep Decision bound-ary Instance Generation (DeepDIG). For a given pre-traineddeep neural network model and two classessandt, Deep-DIG tries to find instances as close as possible to the decisionboundary between the two classessandt.borderline instances stems from the fact that the input space ofDNNs is of high dimension e.g.,R784in the case of simple grayscaleMNIST images, which makes searching for instances close to thedecision boundary a non-trivial and challenging task.