[Writing_4]一些表达

State-of-the-art deep learning methods have shown a remarkable capacity to model complex data domains, but struggle with geospatial data.

We propose to enhance spatial representation beyond mere spatial coordinates, by conditioning each data point on feature vectors of its spatial neighbours, thus allowing for a more flexible representation of the spatial structure

MixMatch targets all the properties at once which we find leads to the following benefits:

A common underlying assumption in many semi-supervised learning methods is that

we propose an efficient training scheme训练方法，scheme体系 to learn meta-networks

We employ使用 multiple多个 LM objectives to pretrain
UNILM in an unsupervised manner.

The problem is that the budget for annotation is limited.

are beneficial for text classification

In practice在实际中

Using NMT in a multilingual setting exacerbate使加剧，使恶化 the problem by the fact that given k language

In this work, we take a different approach采取了不同的措施 and aim to improve

compares favorably (up to +2.4 BLEU) to other approaches in the literature is competitive with pivoting

Another family of approaches is based on distillation.Along these lines Firat et al. (2016b) proposed to fine tune

it is attractive to 干什么是有吸引力的 have MT systems that are guaranteed to exhibit zero-shot generalization since the access to parallel data is always limited and training is computationally expensive

Similar to the style transfer works discussed above做状语, it also disentangled the semantics and the sentiment of sentences using a neutralization module and an emotionalization module respectively.

Several techniques have been proposed for addressing
the problem of domain shifting.

Despite their promising results, these works
share two major limitations.

We also demonstrate through a series of analysis that
the proposed method benefits greatly from incorporating
unlabeled target data via semi-supervised
learning, which is consistent with our motivation

Neural Machine Translation (NMT) performance degrades sharply when parallel training data is
limited

The majority of current systems for end-to-end dialog generation focus on response quality without an explicit control over the affective content of the responses.

While these methods showed encouraging results,

Various solutions have been proposed to mitigate this issue

In this work, we show for the first time that one can align word embedding spaces without any cross-lingual supervision,i.e.即, solely based on unaligned datasets of each language

. This performance is on par with平起平坐 supervised approaches

This paper aims to extend previous studies on “style transfer” along three axes.

we seek to试图 gain a better understanding of what is necessary to make things work

We will open-source our code and release the new benchmark datasets used in this work, as well as our pre-trained classifiers and language models for reproducibility.

For instance例如, the latter requires methods such as REINFORCE

However, a classifier that is separately trained on the resulting encoder representations has an easy time recovering很轻松的做某事 the sentiment.

So far, the model is the same as the model used for unsupervised machine translation by Lample et al. (2018),albeit with虽然 a different interpretation of its inner workings,

we use a combination of什么的组合 multiple automatic evaluation criteria informed by our desiderata.

Unless stated otherwise 除非另有声明, we suppose that we have N monolingual corpora fCigi=1:::N , and we denote by ni the number of sentences

The motivating intuition is that

Finally, we denote by Ps->t and Pt->s the translation models from source to target and vice versa.

still possess拥有 significant amounts of monolingual data

This set up设定 is interesting for a twofold reason.

This procedure is then iteratively repeated, giving rise to 产生 translation models of increasing quality

We then present展现 experimental results in section.

Let us denote by WS the set of words使用Ws代表单词集合 in the source domain associated with the (learned) words embeddings ZS = (zs 1; ::::; zs jWSj), Z being the set of all the embeddings

which is also an LSTM, takes as input 将什么什么当成输入 the previous hidden state, the current word and a context vector given by a weighted sum over the encoder states.

�D are the parameters of the discriminator, �enc are the parameters of the encoder, and Z are the encoder word embeddings.

we propose the surrogate criterion替代标准

the coefficient is in average 0.75

Since WMT yields 可以当成have来理解 a very large-scale monolingual dataset

Without the auto-encoding loss (when �auto = 0), the model only obtains 20.02, which is 8.05 BLEU points below the method using all components.

Finally, performance is greatly degraded also when the corruption process of the input sentences is removed.

Our approach is also reminiscent of the Fader Networks architecture

it would not be hard for us to imagine what state change may happen to the apple.

we intentionally有意 frame the action as讲什么构造成 a language expression

Such ability is central torobots which not only perceive
from the environment

However, a concomitant伴随的 defect is that

The motivation behind背后的动机 is twofold双重的.

In the presence of

a language model with access to拥有 information available in a
KB.

Our Knowledge-Language Model (KALM) continues this line of work by augmenting a traditional model with a KB.

The proposed model does not require parallel text-summary pairs, achieving 结果状语 promising results in unsupervised sentence compression on benchmark datasets.

The LM prior incentivizes刺激 C to produce human-readable summaries。

Therefore it is not comparable, as it is semi-supervised.as they were obtained on a different, not publicly available test set.

Following previous work, we report the average F1 of ROUGE-
1, ROUGE-2, ROUGE-L.

If we remove the LM prior, performance drops, esp. in ROUGE-2 and ROUGEL. This makes sense, since连词 the pretrained LM rewards correct word order.

A possible workaround might be to modify SEQ so that以便 the first encoder-decoder pair would turn the inputs to longer sequence.

We demonstrate that significant gains can be realized by applying
adaptive convolutions to baseline CNNs.

Our adaptive convolutions improve performance of all the baseline CNNs as much as up to 2.6 percentage point, without any exception毫无例外的, in seven text classification benchmark datasets.

Our work is different from them in that we focus on the convolution operation.

An intriguing有意思的 theoretical property of our method is that it provides an effective mechanism to encourage diversity of word embedding vectors,

We side-step避开 these difficulties by completely avoiding the
need for example summaries

the entire model was trained from scratch 从头开始训练的

In contrast to this line of interesting work.

For our problem对于我们遇到的问题（我们着手解决的问题）

Our findings align with the behavior reported by Gu.

we attain获得 within 0:4% of the performance of full fine-tuning

It is widely known that众所周知 neural network training is sensitive to the loss that is minimized

This paper tries to shed light upon 阐明… behavior of neural networks trained with label smoothing.

We demonstrate that label smoothing implicitly calibrates隐式的校准了 learned models

Before describing our findings, we provide a mathematical description of label smoothing

NMT models can be immensely极大的 brittle to small perturbations applied to the inputs

Our method advances existing explanation methods by addressing issues in coherency and
generality.

However, in contrast to the high discrimination power, the interpretability of DNNs has been considered an Achilles’ heel软肋，弱点 for decades.

hindering further development and application of deep learning.

Specifically, this study aims to具体而言这一研究旨在 answer the following research questions:

For all models other than CNN对于除过CNN的所有模型

or the language/claims in the paper should be softened。

Some minor grammatical mistakes/typos打印错误 (nitpicking):