Example Sentences:

• Hefei Qiu: We propose a method of automatic machine translation evaluation that is quick, inexpensive, and language-independent, that correlates highly with human evaluation, and that has little marginal cost per run. .source

• Zihan Li: These methods have the stability and reliability of trust-region methods but are much simpler to implement, requiring only few lines of code change to a vanilla policy gradient implementation, applicable in more general settings (for example, when using a joint architecture for the policy and value function), and have better overall performance.source

• Wei Ding: Information is interpreted more easily and more uniformly if it is placed where most readers expect to find it. These needs and expectations of readers affect the interpretation not only of tables and illustrations but also of prose itself. Readers have relatively fixed expectations about where in the structure of prose they will encounter particular items of its substance. If writers can become consciously aware of these locations, they can better control the degrees of recognition and emphasis a reader will give to the various pieces of information being presented. Good writers are intuitively aware of these expectations; that is why their prose has what we call “shape.” source

Before & After:

• Hefei Qiu

  • before: Different from previous methods of using data augmentation to construct positive pairs in contrastive learning, we use semantic similarity or entailment relation of two sentences to build positive pairs.

  • after: In the way of constructing positive pairs, different from previous methods of using data augmentation which are usually sophisticated or uninterpretable, we apply semantic similarity or entailment relation of two sentences which is easy to implement and easy to understand.

• Tianyu Kang

  • before: This idea comes from the natural structure of the neural network. Since we borrow the structure from humans’ real neural networks in the brain, we think we can borrow more prosperity from the human learning process.

  • after: It’s an analogy that goes back to the dawn of the artificial neural network: ever since we discovered that humans could solve problems by neural networks in the brain, we’ve wondered if the machine might work in a similar fashion.

• Zihan Li

  • before: To diagnose and consider relationships between numerous features sequentially, we use conditional independent test, which follows the Bayesian rule, a powerful non-parametric method detecting and revealing causal relationships between variables, but considers the effects from previous variables instead of traditional independent test.

  • after: To diagnose and consider relationships between numerous features sequentially, we develop a framework based on the conditional Bayesian independent test with two advantaged factors: detecting and revealing causal relationships between variables, considering the effects from previous variables instead of traditional independent tests.