Leveraging Patterns and Interesting Anomalies through Scalable Semantic Analysis

智能手机、汽车、医疗设备、银行系统、电网：软件运行着我们世界的许多方面，而且它的影响范围每天都在扩大。从表面上看，我们可以看到这些应用程序之间存在差异，但是这些差异有多深呢？它的功能有很多共性吗？人们在构建时所犯的错误是否存在模式？是否有人以创新的方式解决了常见问题？这些问题超出了我们目前的科学和工程能力，无法大规模回答。找到这些共性、模式和精华将帮助我们从经验中学习，支持好的方法，并警告坏的方法，基于证据而不是一厢情愿的想法。这将允许创建更好的软件，最终使依赖于这些软件的社会受益。******实现这一目标并非易事。在现实世界中，已经收集了很多关于软件的数据。不幸的是，这样的数据是基于文本细节的：虽然基于文本的数据足够好，可以让谷歌这样的搜索引擎回答常见问题（比如找到当地餐馆的列表），但基于文本的数据对文本的潜在含义一无所知。软件程序的行为是由计算机执行一套用软件编程语言编写的详细指令来实现的。人类语言是复杂而模糊的，而软件语言必然是明确的：计算机哑而忠实，完全按照指令行事。原则上，我们可以通过分析编写软件的文本（称为源代码）来理解软件可能的行为（称为其语义）。但我们面临着挑战：工业软件往往非常庞大；有很多；而且我们没有可靠的方法来比较一个软件程序与另一个软件程序的语义。******我们将通过解决两个关键问题来解决这些挑战：我们必须大规模地分析软件语义，克服限制传统方法有用性的基本技术问题；为了发现软件语义中的模式和异常，我们必须在大规模上比较和对比这些分析的结果。******我们将开发允许对大量软件源代码进行语义探索的技术。我们将开始探索，以展示我们的新技术的价值。我们的发现将导致对软件本质的更深层次的理解，丰富我们的科学知识。对好的和坏的实践有更深的了解，通过避免坏的实践而加强好的实践，将导致更好的软件。越来越依赖于软件的社会，将会因为以更低的成本生产出更好的软件而受益

Smartphones, automobiles, medical devices, the banking system, the electrical grid: software runs many aspects of our world, and its reach increases every day. Superficially, we can see that there are differences between each of these applications, but how deep do those differences run? Is there a lot of commonality in its functionality? Are there patterns to the mistakes that people make when building it? Are there hidden gems in which someone has solved common problems in an innovative way? Such questions are beyond our current scientific and engineering abilities to answer on a large scale. Finding those commonalities, patterns, and gems would help us to learn from experience, support good approaches, and warn of bad approaches, based on evidence and not on wishful thinking. This would allow better software to be created, ultimately benefitting our society which depends on that software.******Achieving this goal is not straightforward. Much data is already collected about software in the real world. Unfortunately, such data is based on textual details: while text-based data is good enough to allow search engines like Google to answer common questions (like finding a list of local restaurants), text-based data says nothing about the underlying meaning of the text. The behaviour of a software program is achieved by a computer performing a detailed set of instructions, written in a software programming language. While human languages are complex and ambiguous, software languages are necessarily unambiguous: the computer is dumb but faithful, doing exactly as it is told. In principle, we can understand the possible behaviours of software (called its semantics) by analyzing the text in which it is written (called source code). But we face challenges: industrial software tends to be really big; there is a lot of it; and we have no reliable means of comparing the semantics of one software program with those of another.******We will address these challenges by tackling two key problems: we must analyze software semantics on a large-scale, overcoming the fundamental technical problems that limit the usefulness of traditional approaches; and we must compare and contrast the results of such analyses, also on a large-scale, in order to discover patterns and anomalies in the software semantics. ******We will develop techniques to permit the semantic exploration of huge amounts of software source code. We will start that exploration to demonstrate the value of our novel techniques. Our discoveries will lead to deeper understanding of the nature of software, enriching our scientific knowledge. Deeper knowledge of good and bad practices will lead to better software, by avoiding bad practices but enforcing good practices. Society, which depends ever more heavily on software, will benefit as a result of better software, which is produced at lower costs.**