1. Easy Questions First? A Case Study on Curriculum Learning for Question Answering
2. Bridging the Gap between Training and Inference for Neural Machine Translation
3. Improving Multi-step Prediction of Learned Time Series Models
4. Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks
5. SEQUENCE LEVEL TRAINING WITH RECURRENT NEURAL NETWORKS
6. Minimum Risk Training for Neural Machine Translation
7. Greedy Search with Probabilistic N-gram Matching for Neural Machine Translation
8. Insertion Transformer: Flexible Sequence Generation via Insertion Operations

## [Easy Questions First? A Case Study on Curriculum Learning for Question Answering]

QA任务我不感兴趣，只讨论CL。

①Greedy Optimal: has the minimum expected effect on the model
Change in Objective causes the smallest increase in the objective.
②Mini-max minimizes the regularized expected risk when including the question with the answer candidate a ij that yields the maximum error.
③Expected Change in Objective the minimum expected effect on the model
④Change in Objective-Expected Change in Objective the minimum value of the difference between the change in objective and the expected change in objective
⑤Correctly Answered is answered by the model M with the minimum cost

## [Bridging the Gap between Training and Inference for Neural Machine Translation]

①假如模型生成了第三个词为abide，为了和ground truth相一致，模型会强制让第四个词生成with，然后with作为输入去生成the rule，但实际上整句是错误的。如cand1就是过度矫正（overcorrection）
②假设模型生成对了by，但也可能因为输入了by而产生了错误的’the law’，假设一种情形，模型记住了with后面一定跟the rule，为了能够生成cand3的，我们应将 with作为输入而不是by，即使我们生成了by。称这种做法为 overcorrection recovery。

### 方法

#### Oracle word selection

##### sentence level

force decoding

①当生成到第j个step时top first的概率是EOS，此时$j \leqslant\left|\mathbf{y}^{*}\right|$，那么避开EOS，选择top second概率的词，使其能够继续生成下去。

②当生成到$\left\{\left|\mathbf{y}^{*}\right|+1\right\}$个时还没生成到EOS，则直接选择EOS作为结尾，强制停止。

e是epoch数。

## [Minimum Risk Training for Neural Machine Translation]

### 方法

（这里就有点像RL了，生成了几个episode，希望reward最大的episode出现的可能性更大。而普通的训练方法则是在每步都优化。）

## [Greedy Search with Probabilistic N-gram Matching for Neural Machine Translation]

### 论文做法

precision和recall则为：

BP为brevity penalty，$w_n$是n-gram的weight。

## [Insertion Transformer: Flexible Sequence Generation via Insertion Operations]

### Insertion Transformer Model

②标准transformer生成n个向量表示，将最后一个表示用于预测下一个词，而在这里需要生成n+1个表示，也即slot representation，每两个词之间共n-1个表示，加上开头与结尾两个。通过加前后两个特殊的标记符，过transformer后获得n+2个向量，并将每两个相邻的向量拼起来获得n+1个向量。

### Training and Loss Functions

#### Uniform

This neutral approach is useful insofar as it forces the model to be aware of all valid actions during each step of decoding, providing a rich learning signal during training and maximizing robustness；
Such an approach also bears resemblance to the principle of maximum entropy, which has successfully been employed for maximum entropy modeling across a number of domains in machine learning

#### Termination

##### Training Differences

①generation step之间没有state的propagation，因为每生成完都要重新计算state，不能复用。
②普通的autoregressive在训练时可以一次性计算完所有的loss，而这里只能一次计算一个。因此更占内存。
③subsampling可能带来variance估计不准的问题。

### 实验

#### Part1

①发现EOS总是过早被生成，因此引入EOS penalty，也即将EOS的概率减掉一个β，当EOS的概率与概率第二高的差距超过β才真正生成EOS，发现确实有不错的提升。

②使用knowledge distillation有显著提升，teacher model是transformer baseline。

#### Part2

parallel decoding的几个例子：