Exploring the Potential of Natural Language Processing Techniques in Criminal Justice Agencies

全文：探索自然语言处理技术在刑事司法机构中的潜力：对假释委员会释放决定中的种族差异的调查。拉米审查（2017）强调了在获取统计数据以探索英格兰和威尔士刑事司法（CJ）系统中的种族差异方面的重大困难，并敦促CJ机构纠正这一问题。在某些情况下，解决方案只涉及发布现有数据集的匿名版本。在许多其他情况下，所需的数据并不存在。采用新的数据收集程序往往是不现实的，因为它需要在过去十年中预算耗尽的机构的额外投资。为了应对这些限制，这个跨学科项目将探索另一种具有成本效益的途径，通过将自然语言处理技术应用于存储在现有行政CJ记录中的大量自由文本数据，来生成能够突出CJ系统中潜在差异的新数据。所有CJ机构都会产生大量记录，记录所处理案件的特点和所通过的决定。使用内容分析对这些记录进行手工处理，将相关文本信息编码到统计数据集中，从而实现定量分析的传统由来已久（例如Hood, 1992； Myers和Talarico， 1987）。这些技术的关键问题在于它们的可伸缩性。手工处理记录的成本/时间与要处理的案例数量成正比，这使得样本要么太小，要么太贵（Pina-Sánchez等人，In Press）。随着数据科学和人工智能领域的进步，我们建议使用文本挖掘技术来自动进行这种编码过程。最近，Pina-Sánchez等人（2019）展示了这种技术在处理句子记录方面的潜力。然而，这种概念证明仍然受到有关文档有效性和所提出的算法复杂性的重要限制的影响。在先前工作的基础上，并与假释委员会（一个重要的国家CJ机构）合作，该项目将推动这一关键研究领域的方法学前沿。假释委员会进行风险评估，以决定是否可以安全地将囚犯释放到社区。在最近的年度演习中，假释委员会（2018a）处理了16,436次“书面听证会”，定期记录简短的结构化文本摘要（大约两页长）。这些“听证摘要”抓住了案件的主要特征，以及囚犯的人口统计因素，包括他们的种族。分析这些“听力摘要”的重要样本，项目目标是：i)开发能够处理“听力摘要”的文本挖掘算法；Ii)评估这些算法生成的数据的可靠性；以及iii)分析他们产生的数据，以探索假释委员会决定中潜在的种族差异。总之，拟议的跨学科项目及其利用的战略伙伴关系将为博士生提供一个独特的机会，探索数据科学方法如何在刑事司法结果的背景下应用于公共利益。通过传播统计数据和所产生的方法，该项目还将为CJ的研究提供创新的新思路，并希望促进政府部门的CJ分析师采用新的尖端方法（例如，处理判决前报告、句子抄本以及几乎任何其他CJ记录）。这样做，该项目有一个非常现实的潜力，提供一个直接的现实潜力，提供对David Lammy的数据集的请求的直接回应，揭示CJ系统中任何潜在的差异。

Full title: Exploring the Potential of Natural Language Processing Techniques in Criminal Justice Agencies: An Investigation of Racial Disparities in Release Decisions from the Parole BoardThe Lammy Review (2017) highlights significant difficulties in accessing statistics with which to explore ethnic disparities in the Criminal Justice (CJ) system in England and Wales, and urges CJ agencies to redress this issue. In some instances, the solution simply involves publishing anonymised versions of existing datasets. In many other cases, the required data do not exist. Adopting new data collection processes is often unrealistic since it requires additional investments from agencies that have seen their budgets depleted over the last decade. In response to these constraints, this interdisciplinary project will explore an alternative cost-effective route to generate new data capable of highlighting potential disparities in the CJ system - by applying natural language processing techniques to the large volumes of free-text data stored in existing administrative CJ records. All CJ agencies generate large numbers of records documenting the characteristics of cases processed and decisions adopted. There is a long tradition of manually processing such records using content analysis to code relevant text information into statistical datasets, thus enabling quantitative analyses (e.g. Hood, 1992; Myers and Talarico, 1987). The key problem with these techniques lies in their scalability. The cost/time of processing records manually is directly proportional to the number of cases to be processed, which renders samples either too small or too expensive (Pina-Sánchez et al., In Press). Following advances in the field of Data Science and Artificial Intelligence, we propose the use of text-mining techniques to undertake such coding process automatically. Recently, Pina-Sánchez et al. (2019) demonstrated the potential of such techniques for the processing of sentence records. Yet this proof of concept is still affected by important limitations regarding document validity and the sophistication of algorithms presented. Building on this previous work, and partnering with The Parole Board - a key national CJ agency - the project will push the methodological frontier in this crucial area of research. The Parole Board carries out risk assessments to decide whether prisoners can be safely released into the community. In their latest annual exercise The Parole Board (2018a) processed 16,436 'paper-hearings' for which short structured textual summaries (roughly two-pages long) were routinely recorded. These 'hearing summaries' capture the main characteristics of the case, together with demographic factors of the prisoner including their ethnicity. Analysing a significant sample of these 'hearing summaries' the project objectives are: i) to develop text-mining algorithms capable of processing 'hearing summaries'; ii) to assess the reliability of the data these algorithms generate; and iii) to analyse the data they produce to explore potential racial disparities in Parole Board decisions.In conclusion, the proposed interdisciplinary project and the strategic partnership it leverages will provide the PhD student with a unique opportunity to explore how data science methods can be applied for public good in the context of criminal justice outcomes. By disseminating the statistical data and methods generated, the project will also enable innovative new lines of CJ research, and, it is hoped, facilitate the adoption of new cutting edge methods by CJ analysts across a range of government agencies (e.g. processing of pre-sentence reports, sentence transcripts, and virtually any other CJ records). Doing so the project has a very realistic potential of providing a direct realistic potential of providing a direct response to David Lammy's request for datasets shedding new light on any potential disparities in the CJ system.