NLP方向英语论文常用字词句精选

编程入门行业动态更新时间:2024-10-20 05:39:05

总结

文章来源：LawsonAbs@CSDN
每天读论文，看英语

英语写作在论文表述中占据非常重要的位置。我的表达时常词不达意，漏洞百出，故总结出平常读论文中见到的一些字词句，也将其分享出来供大家写论文时参考一二。

20221204

data-hungry models

Day 1. 20210611

Despite the recent progress, little is known about the… （这句话理解有点儿别扭，但是正确的。译：关于…仍然知之甚少）
Common methods … （常用的方法…）
the complexity of the task makes it difficult to…
We find that the bias induced by the…
adding contextualized word representations greatly increases performance … （修饰搭配）
yield better results in … （yields）
In contrast, xxx improves xxx, but considerably lowers performance on … (lower)
is one of the largest, most widely used crowdsourced datasets in … (one of 句型)
even with recent advances in … （even with）
models still show a high error rate …
reached a performance ceiling or is there still room for improvement … (ceiling/room 对比)
To answer these questions
the most challenging 5K examples in the development and test sets …
more than 50% of the examples …
are responsible for most of the remaining errors …
have received increasing popularity due to the fact that
Distant supervision based methods for entity and relation extraction…
based method 这种写法。
require light human annotation efforts…
consider the problem of …
propose a joint extraction approach to address this problem by …
To handle noisy instances in a fine-grained manner
gather the wisdom of all agents

day 2. 20210612

To leverage the correlation between these two tasks
the benefits of …
are crucial for downstream tasks, including sentiment classification, opinion summarization,…
aims at extracting aspects and opinion expressions …
To deal with this task.
… mainly consisting of …
Extensive experiments demonstrate that …
separated extraction from web data into different problems based on the modality of the data
take a holistic view toward information extraction
trends within purely textual extraction have shifted toward full-document understanding
这句话可以改写。
considering sentences as independent units.
with this in mind （出于这种考虑）
it is worth considering the information extraction problem as a whole …
Few works in the literature of event extraction have gone beyond individual sentences to make extraction decisions.
事件抽取中少有这样的工作：~
at different levels of granularity （以不同的粒度）
Most existing methods are pipelined framework, which identifies emotions and extracts cause separately, leading to a drawback of error propagation.
Towards this issue
Recent developments in Neural Relation Extraction (NRE) have made significant strides towards Automated Knowledge Base Construction.
在…上取得了重大进展
While much attention has been dedicated towards improvements in accuracy …
for the purpose of analyzing gender bias in relation extraction systems.
Our analysis lays groundwork for future quantifying and mitigating bias in NRE.

day 3.缺

day 4. 20210614

the advent of Deep Learning has radically transformed …
across the entire spectrum from algorithms to circuits.
As myriad(adj.无数的) application domains embrace DL
it has become synonymous with …
general matrix-matrix multiplications (GEMMs)
a natural choice for hardware acceleration to speed up training …
are highly irrugular and sparse…
a flexible and scalable architecture that offers high utilization of all its processing elements(PEs) regardless of kernel shape and sparsity.
performs 5.7× better than…
roughly 3x bettern than state-of-the-art …
demonstrate an instance of …
Deep learning (DL) has emerged as the premier algorithmic technique for analyzing data across multiple domains.
is still an open question, which is the focus of this work.
is extremely compute intensive.
GPUs are currently the most popular acceleration platform in use for training;
built their own specialized training platforms such as …
core component of …
is the primary target for hardware acceleration.
DNN models evolve at a arpid rate, it is impertaive to …
architects need to be cognizant of three trends…
need support for …
fare poorly on …(在…上表现很差)
do not natively support sparsity …
key contributions/novelty
to make the case of accelerating sparse, irregular GEMMs. (提出加速的理由)
dissect state-of-the-art deep learning characteristics.
shows the time break-up of different operations… （展示出了不同操作的时间分割）
the operand matrices vary from tall-skinny (rows dominate over columns) to fat-short (columns dominate over rows)
We measured the compute efficiency…
to boost the efficiency compared to …
are a poster-child for large GEMMs due to … (因为…特征，所以成为了…方面的典型代表)
it is common to … (是很常见的)
in many different scenarios due to two inherent features.

day 5. 20210615

To address this issue.
which makes it extremely prohibitive. (使其费用过高)（这里的prohibitive的含义是：代价较高的。）
One of the key benefits of …
There are a few well recognized …
making it attrative for …
We descibe the details with a walkthrough example in …
enable it to do sth [例句：SIGMA’s flexible substrate enables it to support myriad dataflows. ]
comes at the cost of …
in a pipelined fashion
non-powers of two (非2的次幂)
following the natural binary-tree topology (按照一个自然的二叉树结构)

day 6. 20210616 缺

day 7. 20210617

current benchmark datasets
address real-world situations
make the models fully exploit data
to enhance generalization ability while outperforming indomain performance.
alleviate a class imbalance
a core task of information extraction
In recent years, with a proliferation of
breaking down a holistic perfomance into fine-grained categories
have been highly overestimated
superficial cues
despite such commendable efforts
We refer to the class imbalance problem to the number of negtive examples overwhelmingly larger than that of positive examples.
inherent for the task of NER
a model-agnostic debiasing method （一个与模型无关的去偏差方法）
has reached a plateau
a significant increase
Not surprisingly,

day 8. 20210618

reflects solutions of …
To cope with the class-distribution shift of word between training and test dataset,
how much of our method …
What component can highly support the capacity of …
a wide range of
are prone to （倾向于…）

day 9. 20210619

is especially useful for …
the existing research employed a two-step framework,
It aims at doing sth …
such a pipeline of two steps contains some inherent flaws:
To address these shortcomings,
vast quantities of textual information
faced similar problems
making use of prior knowledge
exploring the commonalities in the challenges and solutions （探索共性）

day 10. 20210620

Most existing methods are pipelined framework, which … , leading to a drawback of error propagation.
presents a neew challenging information extraction task in the domain of …
On all tasks, using BERT embeddings leads to large performance gains, but with increasing task complexity,
serve as competitive baselines
suggests promising research directions (给出有希望的研究方向)

day 11. 20210621

For the convenience to compare our model with previous work
There are two versions of these two datasets … (有两版数据)
to keep in line with…
$ marks results quoted directly from the original papers.
∗ marks results produced with official implementation.
a bonus capbility of …
perform well on …

day 12. 20210622

promising results in …
a large amount of …
significantly deteriorate the model performance …
a simple or weighted combination of … (一种简单的或者是加权的组合)
three essential ingredients …
the ultra-large scale of the weakly labeled data …
annotating a large number of documents can be expensive, time consuming, and prone to human errors.
one of the biggest bottlenecks
To achieve better performance with limited labeled data, researchers resort to large unlabeled data. (resort to)
domain-specific (特定领域)
Throughout the rest of the paper, (在论文的其它部分)

day 13. 20210623

An ultra-large volume of weakly labeled data contains useful domain knowledge. (an volume of)
enormous noise （注意搭配）
This is further corroborated by our empirical observation
in conjunction with
a sequence labeling task
To harness the power of weakly labeled data,
make the fitting to the weak labels more conservative/aggressive.
is closed related to …
the gap between … and … is usually huge.
only marginally improve the performance

day 14. 20210624

we use the grid search to find optimal hyperparameters.
can directly improve the performance.

day 15. 20210628

has become a dominant choice in many domains.
on popular leaderboards
attain excellent results
we mathematically characterize the expressive power of Graphormer
it is still an open question
give an affirmative answer

day 16. 20210629

One notable geometrical property that distinguishes graph-structured data from other structured data is that there does not exist a canonical grid to embed the graph.
Let the feature vector of node v i v_i vi be x i x_i xi
we elaborate(详细阐述) on several key designs in the Graphormer.
An advantage of Transformer is its global receptive field.
has a byproduct problem