Computational Analysis of Historical Poor Law and Workmen’s Compensation Cases (CBR project)

Overview

Background

This project, supported by the Keynes Fund, builds on research initially begun by Simon Deakin and Linda Shuku under the ESRC-funded project on Legal Systems and Artificial Intelligence. Vanessa Cheok, Zain Ali and Ieva Petrukesvicius have assisted in the preparation of the datasets used in the project. 

The project aims to explore the relationship between legal evolution and industrialisation by using computational approaches to the analysis of legal texts. The research will help answer some unresolved questions concerning the nature and direction of legal change during the British industrial revolution, while also advancing the use of statistical methods to explore judicial decision making.

Using natural language processing (NLP) and machine learning (ML) approaches to study the structure of legal texts is a fast-growing field. While much attention has been devoted to the use of these techniques to predict case outcomes under current laws, our approach is different: using historical data, we aim to see how far changes in the content and structure of judicial decisions over time are driven by wider economic factors. Specifically, we take advantage of recently developed NLP and ML methods to test the claim that judges decide cases not solely according to established rules (the doctrine of precedent) but in ways that are systematically shaped by economic shocks, the business cycle, and secular trends associated with technological change and long-run industrial transitions. 

From the early 17th century, the English poor law operated as a proto-welfare state, the first of its kind in Europe and the most extensive and comprehensive in its operation. National legislation required the raising of a local tax (poor rate) in the more than 15,000 parishes and the administration of outdoor relief (cash payments) to the unemployed, long-term ill, and elderly. The legal right to relief, expressed through the concept of the ‘settlement’, could be obtained by various means. One of these was the institution of the ‘yearly hiring’, which conferred a settlement in the parish where an individual had completed a period of continuous service for one year. Critical to the operation of this aspect of the poor law was the legal definition of the status of ‘servant’. This turned on the way in which the work was carried out as well as on marital status, with in general only those unmarried at the start of the service being capable of acquiring a settlement by hiring. It also turned on the juridical understanding of what precisely counted as a full year of service. Centuries before today’s debates about migration, labour flexibility and platform work, the English courts were grappling with the need to adapt legal definitions to a fast-changing labour market. There is evidence that the language used by the courts to define the concept of the settlement by hiring changed significantly over time, becoming more restrictive towards the end of the 18th century and into the 19th. In principle, judges in a common law system are meant to decide like cases alike (the doctrine of precedent), and this understanding of the judicial role was already well established in the 18th century. Then as now, however, the doctrine of precedent gives judges leeway to develop the law in new directions as social conditions changed, ‘distinguishing’ earlier decisions on their facts. The evolution of judicial language in the poor law case law appears to suggest, at least at first sight, that the judges modified the scope of the poor law, making it more restrictive, as time went on. The overall question the project explores is: did changing judicial language systematically reflect or respond to larger changes in the economy and in society during this period?

While the first phase of the project (2024) addressed this question in the context of the poor law, in the second phase (2025) the scope was extended to workmen’s compensation laws of the early 20th century. Workmen’s compensation laws were in force in the United Kingdom between the 1890s and the 1940s, roughly corresponding to the period of the so-called second industrial revolution, when heavy industry and mass transit systems were becoming established. To succeed in a claim, the victim of a workplace accident had to show that they had been employed as a ‘workman’ (a concept functionally similar to the poor law ‘servant’ and modern labour law’s ‘employee’) and that the accident had occurred ‘out of and in the course of [their] employment’. In the same way that the concept of ‘service’ denoted the limits of the right to poor relief under pre-1834 poor law, the concepts of ‘workman’ and ‘course of employment’ operated to determine the scope of the right to compensation for workplace injury and disease in the first half of the 20th century.

We aim to see whether shifts in the meaning of these concepts were associated with developments in the economic environment. If our hypothesis of a litigation-mediated impact on legal language is correct, periods of recession, falling wages and declining strike action should be associated with more restrictive reasoning, limiting the right to compensation, while periods of economic growth, rising wages and increasing incidence of strikes should be associated with more liberal language, expanding it.

Aims and objectives

The project addresses the following questions: 

• in the case of the poor law, did legal decision-making respond to external economic factors in the nature of exogenous shocks (harvest failures), cyclical fluctuations (the business cycle) and secular trends (land enclosure)?
• where there was such a response, was the direction of legal change broadly pro-cyclical (weakening the right to poor relief in times of crisis and downturn) or counter-cyclical (maintaining legal protections in the face of adverse economic conditions)?
• do we see similar trends during the lifespan of workmen’s compensation laws, expressed in the shifting judicial terminology used to define concepts which determined the scope of employers’ liabilities for workplace injuries (‘servant’, ‘workman’ and ‘course of employment’)? 

Methods

For our legal data, we use 2 datasets that we have prepared ourselves. The first is a dataset of c. 250 poor law cases covering the period 1680-1834; the second is a dataset of c. 900 workmen’s compensation cases covering the period 1907 to 1947. The datasets consist of cases heard in the higher courts, and also reported. Thus they while they are a sample of all disputes, they are effectively the population of legally important cases, as only those deemed to have precedential value were reported. We digitised the cases using optical character recognition, and then cleaned and annotated them for analysis. The annotation breaks the judgment text down into various components (including ‘facts’, ‘arguments’, ‘judgment’ and ‘ruling’) and also categorises the outcomes as ‘liberal’ or ‘restrictive’. ‘Liberal’ indicates a more expansive interpretation of the concept of service, in the case of the poor law, and ‘workman’ acting ‘out of and in the course of employment’, in the case of workmen’s compensation laws; ‘restrictive’ indicates a narrow interpretation in each case.

Once the manual coding is completed, we use a number of methods involving NLP and ML applications. The manual coding provides us with the outcome of the case: liberal or restrictive. This is also a binary measure of judicial sentiment. Using automated approaches, we can generate a continuous measure of language evolution. In our new study, we propose to perform sentiment analysis using the dictionary method and BERT model. We will also identify themes emerging in the case law over time, using the BERTopic model.

Our economic data in the case of the poor law analysis take the form of annual data on population and GDP growth, poor relief annual disbursement, the average cost of living and wage growth, both nominal and real, prices for main grains (wheat, barley, and oats)​, and the business cycle. Data for the later period when the workmen’s compensation laws were in force covering wage growth, employment and unemployment, GDP growth, product and retail prices, unionisation, collective bargaining coverage and trends in strike activity, are available from a range of sources. 

In the econometric stage of the analysis, our approach treats economic indicators as the independent variable and seeks to identify how far they predict (in the statistical sense) changes in the linguistic structure of legal texts, taken as proxies for or indicators of the approaches taken by the courts (liberal or restrictive as the case may be) to the definition of poor relief rights. A variety of econometric techniques are used, including OLS regressions, probit models, vector autoregression approaches, and Granger causality tests.

Progress

The first phase of the project, focusing on poor law cases, was completed by the end of 2024. Initial results were written up in a working paper which was subsequently published as an article in a peer-reviewed journal. The dataset of legal cases used in the study was deposited with the UK Data Archive. The second phase, on workmen’s compensation cases, began in January 2025. The cleaning and annotating of the cases has been completed. An initial analysis of results was presented to a conference at Strathclyde University in June 2025.

Findings 

In the first phase of the study, we tested the hypothesis that the trend towards more restrictive judicial decision making in the 18th century poor law was associated with factors in the economy likely to drive up the cost of poor relief. Our hypotheses were broadly supported: the language used by the courts tended to be more liberal at times of rising GDP, rising wages, falling grain prices and falling parish expenditure, and more restrictive when these economic trends were reversed. The same results were obtained when lagged values of the economic variables were used, suggesting that law followed the economy, rather than vice versa.  In the second phase of the study, our exploratory analysis of British and Irish workmen’s compensation cases of the first half of the 20th century suggests that judges decided workmen’s compensation cases pro-cyclically, that is, more restrictively in times of recession and periods of worker militancy. We find evidence that isomorphism in decision making increased over time, implying that the meanings of legal concepts were stabilised over the life cycle of the legislation. Scottish and English judges decided cases in similar ways, but the trend in Irish cases was somewhat distinct from those in English and Scottish cases. Our machine learning and NLP analysis suggests that there is a latent structure to the legal language used in the cases, which predicts, in the sense of being correlated with, case outcomes.