English phrases for writing
- 定义术语 We use the term "xx" to refer to a context passage paired with an unanswerable question.
- 给baseline起名字 we refer to these as xxx
- In extractive reading comprehension datasets, a system must extract the correct answer to a question from a context document or paragraph.
- Sentence selection datasets test whether a system acan rank sentences that answer a question higher than sentences that do not.
- sentence selection(QASENT, WikiQA) 和 multiple choice （MCTest, RACE）是不同的任务
- are not able to doing 不能做
- tend to do sth 可能做
- Consequently, we first given a brief introduction to ..., and then investigate .
- be capable of jointly doing sth 能做某事
- considerable performance gains 比较大的性能提升
- The statistics of the xxx dataset is shown in the Table x.
- We use the term xxx to refer to xxx with xxx.
- IN extractive reading comprehension datasets, a system must extract the correct answer to a question from a context document or paragraph.
- user-facing system
- Recognizeing textual entailment requires systems to decide whether a hypothesis is entailed by, contradicted by, or neural with respect to a premise.
- character-derived embeddings
- restrict their applicability in many domains that suffers from a dearth of ... 受限制
- provide a valuable alternatives to doning sth. 提供一种可选方案
- provide a significant performance boost. 提升一个显著的性能提升
- compelling evidence for 有说服力的证据
- doing sth is challenging for two reason. First, xxx. Second, xxx. 做某事是有挑战的因为两个原因，其一。其二。
- This model choice provides us with more memmory for handling .. 模型结构的选择使得我们有更多空间去处理
- compared to alternatives like ... ， resulting in
- Our training proedure consists of two stages. The first stage is ... . This is followed by a xx stage, where we adapt the model to a discriminative task with labels data
- must posses the ability to determine whether ..
- if and only if the associated passage contains the answer
- strong baseline system that use pieline and threshold-based approaches
- Machine comprehension systems mimic the process of reading comprehension by ansewring questions after understanding natural language text.
- neural model for answer span extraction
- Although this does allow us to obtain bidirectional pre-trained model, threr are two main downsides to such an approch.
- semantically equivalent 语义等价
- Given a pair of sentences, the goal is to predict whether the second sentence is an entailment, contradication, or neutral with respect to the first one.
- is more concerned with doing /put more emphasis on N 更加关注
- Types of machine learning 机器学习的累心
- almost all 几乎所有
- preliminaries 前言 background 背景， 作为第二节
- in principle 原则上，大体上
- is free of 免疫 is free of adversarial attck
- is upper bounded by
- For ease of exposition 为了说明简单