From seafood from Thailand and electronics from Malaysia and China, to textiles from India and wood from Brazil, modern slavery exists in all corners of the planet. It is a multibillion-dollar transnational criminal business that affects us all through trade and consumer choices. In 2016, an estimated 25 million people were forced to work through threats, violence, coercion, deception, or debt bondage. Of these, 16 million were forced to work in the private sector. Given the widespread nature of the problem, governments, corporations, and the general public are increasingly expecting companies to accurately disclose the actions they are taking to tackle modern slavery. Yet, five years on, there are challenges with understanding companies’ compliance under the 2015 UK Modern Slavery Act. It is unclear which companies are failing to report under the MSA, while the quality of these statements often remains poor. Project AIMS (Artificial Intelligence against Modern Slavery) harnesses the power of artificial intelligence (AI) for tackling modern slavery by analyzing modern slavery statements to assess compliance with the UK and Australian Modern Slavery Acts, in order to prompt business action and policy responses. This paper examines the challenges and opportunities for better machine readability of modern slavery statements identified in the initial stages of this project. Machine readability is important to extract data from modern slavery statements to enable analysis using AI techniques. Although extensive technological solutions can be used to extract data from PDFs and HTMLs, establishing transparency and accessibility requirements would reduce the resources required to assess modern slavery reporting and ultimately understand what companies are doing to address modern slavery in their direct operations and supply chains - unlocking this critical ‘AI for Social Good’ use case.
From seafood from Thailand and electronics from Malaysia
and China, to textiles from India, wood from Brazil, and
apparel manufacturing in the United Kingdom,1 modern
slavery exists in all corners of the planet. Modern slavery is a
multi-billion-dollar transnational criminal business that
affects us all through trade and consumer choices. In 2016, an
estimated 25 million people were forced to work through
threats, violence, coercion, deception, or debt bondage. Of
these, 16 million were forced to work in the private sector
The Future Society,3ii in partnership with Walk Free,4iii
launched Project AIMS (Artificial Intelligence against
Modern Sla
Project AIMS builds upon the work of Walk Free,
WikiRate,5iv and Business & Human Rights Resource
Centre (BHRRC)6v to assess the statements produced under the
UK Modern Slavery Act. It draws from the BHRRC Modern
Slavery Registry to develop an AI algorithm to ‘read’ and
assess the statements produced by companies under supply
chain transparency legislation.vi This algorithm will use 18
metrics designed by Walk Free in line with the UK Home
Office guidance
There are four phases to Project AIMS. The first phase of Project AIMS is focused on accessing, gathering and structuring the data from existing company statements, building the largest publicly available text corpus of modern slavery statements.7 In phase two of the project, we will design an automated labeling function through weak supervision tools to increase the amount of available labeled data. Once sufficient data are correctly labeled, the third phase of Project AIMS begins: using supervised machine learning methods to create a document classifier, which can assess modern slavery statements against the 18 metrics. Lastly, in the fourth phase, the results will be published, and the tool will be made publicly available through an open-source API. 3 The Future Society is an independent 501(c)(3) nonprofit think-and-do tank working on advancing the responsible adoption of AI and other emerging technologies for the benefit of humanity. 4 Tackling one of the world’s largest and most complex human rights issues requires serious strategic thinking. Walk Free approaches this challenge by integrating world class research with direct engagement with some of the world’s most influential government, business, and religious leaders. We invest our time and resources in a collaborative manner to drive behavior and legislative change to impact the lives of the estimated 40 million people living in modern slavery today. 5 WikiRate is a nonprofit that hosts an open data platform which allows anyone to systematically gather, analyze and report publicly available information on corporate Environmental, Social and Governance (ESG) This publication addresses the challenges and opportunities identified during the first phase, namely the process of accessing, gathering and structuring the data from the company statements. Data collection and structuring is a key cornerstone in building any successful AI project and thus this publication puts forward a set of lessons learned and recommendations on good practices to facilitate the application of AI for Social Good. This paper adopts the perspective that although extensive technological solutions can be used to extract data from modern slavery statements in PDF and HTML formats, establishing transparency and accessibility requirements would reduce the resources required to do so. It focuses on changes that would enable resource-constrained technical experts to extract data in a more efficient manner than what is technically feasible today.
Following California’s 2010 Transparency in Supply Chains Act, the UK developed the first national legal framework for transparency in supply chains: the 2015 Modern Slavery Act. It includes a provision that requires companies supplying goods or services in the UK with an annual turnover of £36 million or more to publish an annual modern slavery statement indicating the steps they are taking to identify and address modern slavery risks.
Yet, five years on, there are several challenges in
understanding business compliance with the UK Modern Slavery
Act. It is difficult to establish which companies are failing
to report, while the variable quality of the statements
released makes it difficult to understand the actions companies
are taking to address modern slavery. With an estimated 12
000-17 000 UK-based companies having to publish
statements per annum, few studies have attempted to assess these
reports due to the laborious nature of manually analyzing
each statement
To date, access to modern slavery statements has been through company websites8 or the compilation efforts by the BHRRC’s Modern Slavery Registry,vii TISC,viii and WikiRate,ix who have collected, collated, and analyzed these data. Much of this information has been collated manually, with teams of researchers searching for, and systematically reviewing, available statements. Given this is a costly exercise that requires a lot of man-hours, a more centralized and automated approach is desirable. Promising steps in this regard are the development of the UK Home Office registry, and the recent launch of Australia's registry, which will centralize the housing of these statements.x Technological innovations will also reduce the time taken to extract relevant information from these statements. This enables insights into company disclosure of actions to remove modern slavery from their operations and supply chains, and also facilitates the automation of elements of the assessment of these statements.
This is a technically challenging task. However, the
challenges in dealing with large complex structured and
unstructured data sets are not new, and neither is the quest to
harness AI technologies to tackle them
This project aims to use AI to support the achievement of the Sustainable Development Goals (SDG), including SDG 8 aimed at:
Promot[ing] sustained, inclusive and sustainable
economic growth, full and productive employment and
decent work for allxi
8 Examples of best practice under the UK Modern Slavery Act
Based on the creation of the dataset under the first phase of Project AIMS, we set out the following recommendations for policymakers and companies to improve access to modern slavery reporting using technology.
Recommendation 1: Governments with modern slavery reporting requirements should publish an up-to-date, comprehensive list of all companies and their subsidiaries subject to reporting.
Recommendation 2: Governments should keep a single registry where companies must submit their statements. These statements must have consistent formatting to ensure easy retrieval. 2a Ensure all statements are required to disclose the reporting period, are timestamped, and include relevant metadata,10 such as the address of company headquarters, that assist interoperability with other data sources. 2b In addition to housing on the company homepage, require companies to submit their statements to the registry. 2c House historical statements in this same registry. Recommendation 3: Governments should legislate that companies should publish statements in machine readable formats11 to improve comparability and support transparency.
Recommendation 1: Companies subject to modern slavery reporting requirements should endeavor to assist governments with keeping an up-to-date, comprehensive list of all companies and their subsidiaries subject to reporting. Recommendation 2: Companies should place their modern slavery statements on their homepage, with a URL that includes the reporting year. 2a Ensure that all statements disclose the reporting period, are timestamped, and include relevant metadata that assists 10 Based on our research to-date, a good metadata for this purpose would be the address of company headquarters. 11 A machine-readable format is a type of structured format that can be read and processed by a computer. Examples suitable for modern slavery statements include Extensible Markup Language (XML). A machine-readable format does not include PDF, although different PDF formats facilitate readability. interoperability with other data sources, such as the address of company headquarters. 2b In addition to housing on their homepage, submit their statements to the registry, with consistent formatting. 2c Provide records of historical statements on their website and in the registry.
Recommendation 3: Companies should publish in a machine-readable format, with infographics and images comprehensively explained in text that fully summarizes and references all information contained within.
Accessing high-quality, structured, machine readable data
from companies’ Modern Slavery Act statements is a
significant challenge
Access is a significant issue facing anyone who wants to extract data. Data are accessible for AI if it can be identified, extracted, processed, and parsed easily by a computer.
To extract data from relevant reports requires the identification of companies that are subject to mandatory reporting requirements. In the UK, this process currently requires visiting company websites and drawing from existing datasets (such as WikiRate or the Modern Slavery Registry), as there is no centralized government registry yet. This raises a number of issues, including:
To date, there is no publicly available list of companies which are in scope of the UK Modern Slavery Act. This makes it incredibly difficult, if not impossible, to identify which companies are in scope of the Act, and pinpoint which should have reported, but have not yet done so. The AIMS project compares metadata variables (e.g. ‘name’ or ‘URL’) 12 A CAPTCHA is a type of test to determine whether or not a particular user is human. across two separate data sets of modern slavery statements from the WikiRate platform and the Modern Slavery Registry. While these databases have collected statements published by companies, they are inevitably incomplete due to the inherent difficulties of collecting all statements in scope. The goal of this comparison was to conduct a gap analysis and assist with the identification of additional statements. This analysis has revealed the difficulty of analyzing companies with complex structures, often with multiple subsidiaries, inconsistent industry classifications, and companies that span multiple industries, which creates challenges for generating a comprehensive streamlined dataset of companies.
Based on the analysis by Project AIMS, from the approximately 17 000 unique statement URLs stored in the Modern Slavery Registry, just 12 005 could be accessed. In approximately 4 913 cases, errors blocked the scraping process, of which 328 errors were related to HTML stored formats. The remaining 4 585 errors affected those stored in PDF format. More precisely, of the approximately 10 700 URLs containing the statements in PDF format, only 6 212 statements could be accessed.
These errors were caused by a number of issues that make
website scraping complicated, such as shifting webpage
structures, redirects and CAPTCHAS,12 unclear navigation,
and unstructured HTML.13 These issues include:
• Statement missing from homepage. Not all companies
follow government requirements to publish modern slavery
statements in a prominent place on their homepage (Home
Office 2019).
• Shifting webpage structures. Website redesign means that
sections can become more complicated to access.
• Connection issues caused by URL structures. Complicated
URL structures, including multiple query strings and
hashes create significant connection issues.
• Connection issues caused by server connection errors. To
scrape the statements, the computer sends a request for
processing to the web server that hosts the statement, and
the server then sends a response to the computer running
the code. If the server is not connected, it is not possible
to scrape data.
• Unclear links. Some links direct to a page which hosts a
number of links to modern slavery statements instead of
the most recent statement itself. This leads to the text
being extracted from that website instead of the text from
the actual reports.xii While it is helpful for companies to
have a webpage that links to all of the previous modern
slavery statements in one place, it is essential for
13 Unstructured HTML is where the HTML has not been tagged in a
consistent pattern that allows for analysis. Sometimes, unstructured HTML is
simply a consequence of bad programming
The format of modern slavery reporting can also add hurdles for the extraction of data.
A new EU regulation requires all financial statements to be
published in a digital format
Data published in PDF format, which is the form that
modern slavery statements often take, is not easily machine
readable
Figure 2 also provides important information on a company’s modern slavery strategy, but the use of a diagram creates additional difficulties in the process of reading, extracting and structuring data. These diagrams are, however, essential for a number of stakeholder groups to help them understand company modern slavery strategies, which is why we do not recommend removing them, but rather supplementing them with a text-based description.
Structure of Reports • Section Titles. Without clearly demarcated section headings that mirror the government’s sections for reporting, it can be difficult to find relevant information for specific metrics. • Alignment with reporting standards. Use of conventional terminology enables easier extraction, and further analysis would be enhanced if this aligned with globally specific reporting standards or frameworks (e.g. SASB, EU frameworks). • Tagging. Labels by companies to assist machine readability would be very helpful; this is particularly important in areas where we see inconsistent typologies used by companies to describe similar phenomena. This could follow suggested or mandated criteria from governments.15
Using Structured, Machine Readable Formats across Corporate Reports Machine-readable formats would make information contained in modern slavery statements more easily accessible, which would facilitate data retrieval and allow for better comparisons between companies within and across sectors and countries, and show change over time. It would also allow for adjustments that enable access for people with disabilities.
The methodology developed to extract relevant
information through Project AIMS could also be applied to other
reporting frameworks to develop a comprehensive picture of
corporate disclosure and activity. In particular, there is an
opportunity to apply this extraction technology to
Environmental, Social and Governance (ESG) reporting
frameworks. There are currently more than 230 sustainability
reporting frameworks, which ultimately impairs rather than
aids the extraction, comparability, and analysis of the wealth
of information contained within these reports
There is also an opportunity to extend financial reporting
requirements to modern slavery reporting and ESG data to
assist efforts to source and efficiently integrate data into
cross-asset investment decisions and implementation.
Companies’ annual financial reports are made machine-readable
under new European Securities and Markets Authority rules.
Doing the same with modern slavery statements and ESG
data would improve comparability, support transparency
and contribute to increased investor protection
Given the challenges and opportunities for machine readability of modern slavery reporting explored in this paper, we believe that establishing transparency and accessibility requirements would reduce the resources required to assess modern slavery reporting, increase understanding of the actions companies are taking to address modern slavery, and ultimately hold companies accountable for the exploitation that occurs in their direct operations and in their supply chains.