The Natural Language Inference (NLI) task often requires reasoning over
multiple steps to reach the conclusion. While the necessity of generating such
intermediate steps (instead of a summary explanation) has gained popular
support, it is unclear how to generate such steps without complete end-to-end
supervision and how such generated steps can be further utilized. In this work,
we train a sequence-to-sequence model to generate only the next step given an
NLI premise and hypothesis pair (and previous steps); then enhance it with
external knowledge and symbolic search to generate intermediate steps with only
next-step supervision. We show the correctness of such generated steps through
automated and human verification. Furthermore, we show that such generated
steps can help improve end-to-end NLI task performance using simple data
augmentation strategies, across multiple public NLI datasets.

通过序列到序列模型结合外部知识和符号搜索，在仅有下一步监督的情况下，生成了中间步骤，并证明其正确性，而这些生成的步骤可以通过简单的数据扩增策略，提高多个公共 NLI 数据集的综合性能。

利用下一步监督生成自然语言推理中的中间步骤

Generating Intermediate Steps for NLI with Next-Step Supervision

In classical planning, the goal is to derive a course of actions that allows
an intelligent agent to move from any situation it finds itself in to one that
satisfies its goals. Classical planning is considered domain-independent, i.e.,
it is not limited to a particular application and can be used to solve
different types of reasoning problems. In practice, however, some properties of
a planning problem at hand require an expressive extension of the standard
classical planning formalism to capture and model them. Although the importance
of many of these extensions is well known, most planners, especially optimal
planners, do not support these extended planning formalisms. The lack of
support not only limits the use of these planners for certain problems, but
even if it is possible to model the problems without these extensions, it often
leads to increased effort in modeling or makes modeling practically impossible
as the required problem encoding size increases exponentially.
In this thesis, we propose to use symbolic search for cost-optimal planning
for different expressive extensions of classical planning, all capturing
different aspects of the problem. In particular, we study planning with axioms,
planning with state-dependent action costs, oversubscription planning, and
top-k planning. For all formalisms, we present complexity and compilability
results, highlighting that it is desirable and even necessary to natively
support the corresponding features. We analyze symbolic heuristic search and
show that the search performance does not always benefit from the use of a
heuristic and that the search performance can exponentially deteriorate even
under the best possible circumstances, namely the perfect heuristic. This
reinforces that symbolic blind search is the dominant symbolic search strategy
nowadays, on par with other state-of-the-art cost-optimal planning
strategies...

在经典计划中，为了满足智能体的目标，目标是制定一系列行动方案。扩展规划是用于捕获和模拟不同类型的推理问题的一种表达方式。本文中，我们提出使用符号搜索进行成本最优规划，为经典计划的不同表达扩展，包括公理、状态依赖性操作成本、过度订阅计划和前 k 个计划。

符号搜索与表达式扩展的最优规划

Symbolic Search for Optimal Planning with Expressive Extensions

Our goal is to build systems which write code automatically from the kinds of
specifications humans can most easily provide, such as examples and natural
language instruction. The key idea of this work is that a flexible combination
of pattern recognition and explicit reasoning can be used to solve these
complex programming problems. We propose a method for dynamically integrating
these types of information. Our novel intermediate representation and training
algorithm allow a program synthesis system to learn, without direct
supervision, when to rely on pattern recognition and when to perform symbolic
search. Our model matches the memorization and generalization performance of
neural synthesis and symbolic search, respectively, and achieves
state-of-the-art performance on a dataset of simple English description-to-code
programming problems.

本文介绍一种用于程序自动合成的方法，通过结合模式识别和显式推理来解决这些复杂的编程问题，同时使用新颖的中间表示和训练算法，使程序合成系统能够自学，从而在简单的英文描述编程问题数据集上取得了最先进的性能表现。