Analysing the language of mechanisation in nineteenth-century British newspapers

Barbara McGillivray, Nilo Pedrazzini, Arianna Ciula, Jon Lawrence, Tiffany Ong, Mia Ridge, Miguel Vieira

1. Introduction: mechanisation in 19th-century Britain and the English lexicon

Mechanisation in nineteenth-century Britain was far more than a series of technological advancements, and it deeply influenced and shaped the nation’s socio-cultural fabric. The effects of these shifts had strong repercussions on various aspects of Late Modern English.^[1] The English lexicon evolved to accommodate a myriad of new terms, expressions, and nuances associated with the growing industrial landscape and cityscape, reflecting the profound societal shifts, and the perceptions, cultural ethos and anxieties surrounding technology at the time.^[2] The focus of this chapter is on the linguistic analysis of nineteenth-century newspaper sources to trace the semantic evolution of some English words related to the domain of mechanisation. This was the main objective of the “Language of mechanisation” work package, one of the main components of the “Living with Machines” project.^[3]

Several studies have analysed the portrayal of machines and technology in Victorian fiction and culture, focussing on literary landscapes.^[4] While some attention has been devoted to the role of periodicals, the historical linguistic analysis of newspaper sources is still underexplored.^[5] As non-canonical sources, newspapers offer a kaleidoscopic view of the era, capturing the many voices that shaped the discourse of the time and offering an opportunity to uncover perspectives that often remain marginalised in traditional historical narratives. Our analysis sheds light into the linguistic mechanisms affecting the lexicon of nineteenth-century English, but also into the social changes of the time and the cultural role of mechanisation.

The nineteenth century marked a significant period of transformation in education and literacy, characterised by gradual but consistent improvements in educational opportunities and by the proliferation of printed media. Prior to the introduction of compulsory elementary education in 1870, education was fragmented across various institutions, and adult education played a crucial role in expanding literacy rates.^[6] The rise of popular novels and newspapers, which saw a rapid increase in number and distribution during this period, exerted a profound influence on readers and contributed to the spread of literacy. The mid-nineteenth century witnessed a remarkable surge in the production of books focused on literary instruction and grammar, reflecting societal developments and the growing public interest in education. A concerted effort was made for the first time to educate the labouring poor in rural areas and industrial towns, facilitated by various economic developments such as the introduction of fast and affordable printing techniques, the repeal of taxes on newspapers and paper, and the establishment of circulating libraries and working-men’s reading clubs.^[7]

The expansion of literacy had significant implications for the dissemination of Standard English in written form and contributed to the democratisation of styles in literature and journalism. Moreover, the standardisation and codification of English accelerated in the nineteenth century, with dictionaries and linguistic scholars endeavouring to capture the breadth and depth of the language. The publication of dictionaries, most notably the Oxford English Dictionary (OED), provided a comprehensive record of the language, charting its historical trajectory and documenting the various influences shaping its evolution. Technological advances, social changes, geographical expansion, and contact with other languages led to a particularly dynamic phase in the history of the English language, during which the vocabulary underwent rapid changes.^[8]

One of the salient features of the nineteenth-century English lexicon is the influx of new words stemming from the rapid advancements in science, technology, and industry, such as ‘telegraph’ or ‘telephone’. Another significant aspect of the nineteenth-century English lexicon is the semantic evolution and expansion of existing words via the acquisition of new meanings in relation to their application to emerging technical advancements and shifts in meaning as new connotations or evaluations became associated with them.^[9]

This chapter explores the semantic changes affecting the lexical field of mechanisation in the English lexicon during the nineteenth century and relates them to the profound technological, cultural, and social transformations of the time. We base our analysis on a portion of the British Newspaper Archive (BNA), which (at the time of writing) contains over 80 million pages of newspapers dating from the 1700s. The large scale of this dataset offers us a unique opportunity to gather textual evidence on the actual usage of words, allowing for a nuanced interpretation of the historical linguistic phenomena observed, which complements our knowledge from more traditional linguistic resources.

Historical dictionaries such as the Oxford English Dictionary play a critical role in the analysis of diachronic lexical semantics (i.e. changing word meanings/usage over time) by offering a chronological record of word usage, and have traditionally supported qualitative methods of historical semantic analysis. Tracing the progression of meanings over time, they provide contextual examples of word usage from diverse sources across time. What they lack, however, is the quantitative dimension. Historical dictionaries do not provide information about the number of times a certain word usage is attested in a particular historical period. This can limit our analysis of language within specific communicative contexts. Quantitative insights allow us to identify and nuance lexical trends and map the rise or fall of usage of specific words over time. Thanks to digitisation efforts such as those that led to the creation of the BNA, access to large-scale textual collections can significantly enrich the evidence base for historical linguistics research. Large digital corpora provide researchers with rich and extensive datasets for analysing lexical and semantic changes over time, identifying patterns, and uncovering previously unnoticed trends or phenomena.^[10] Additionally, advanced computational methods from natural language processing allow us to perform quantitative assessments of linguistic data. Last but not least, visualisation techniques can offer opportunities to explore these data from a variety of perspectives, aiding interpretability and inspiring further analysis.

This chapter presents the results of different research initiatives focussed on the semantic analysis of the English lexicon of mechanisation in the nineteenth century. This research took place as part of the Language of Mechanisation work package within the Living with Machines project, and involved interdisciplinary teams of linguists, historians, research software engineers, curators and other library professionals, and research software engineers.

First, we examined the temporal changes in the meaning of the words related to mechanisation and the variations in their usage across different geographical regions at scale. Section 2 describes the work we have done in developing methods for the automatic detection of lexical semantic change, i.e. the phenomenon by which the meaning of words changed over time. We were able to identify when each word started being used in different contexts and with different meanings by training diachronic word embedding models on 4.6 billion tokens (word instances) from digitised nineteenth-century British newspaper articles and applying changepoint detection methods. Word embeddings are numerical representations of words that reflect their evolving meanings and usage patterns across different time periods. They are widely used in computational linguistics and Natural Language Processing (NLP) to analyse changes in words’ meaning over time.^[11] Comparing the results of this computational analysis with the content of the Oxford English Dictionary highlighted a number of interesting new insights into the subtle shifts of words’ semantics. We also report on our work looking at diatopic (geographic) variation in the lexicon of mechanisation, i.e. how the words were used differently in different regions covered by our newspaper corpus. Our results highlighted certain geographical areas where words changed earlier compared to other regions. Interpreting these linguistic insights within the broader societal and cultural milieu of the time allowed us to see the connection between the linguistic phenomena and the historical context in which they occurred.

Moving to a smaller scale and human annotation methods, section 3 describes the work we have done in designing, implementing, and optimising the semantic annotation of a portion of our newspaper corpus via voluntary crowdsourcing. The multi-disciplinary collaboration in this sub-project and other crowdsourcing projects led to the creation of novel tasks that asked volunteers to closely read texts and select the most appropriate datasets containing high-quality human annotations of historical texts. These datasets, where specific words are annotated with their meanings in context, are very rare. By engaging members of the public in close-reading articles from nineteenth-century newspapers, we published the first annotated sense datasets for historical English.^[12] These datasets are relevant to historical linguists working on the evolution of the English lexicon in the nineteenth century. They also provided insights into the evolving language associated with specific types of machines during this transformative period.

Finally, section 4 describes our work in visualising and analysing the crowdsourced annotation data. We developed an interactive Notebook which allows the user to explore the dataset of source texts and annotations. The Notebook also makes the code used to process and visualise the data available for examination. The Notebook allowed us to examine patterns, trends, and variations in the lexical semantics related to mechanisation in the newspaper corpus. The Notebook interface is designed to enable users to filter and sort the data based on various criteria, such as geographical regions, specific lexical items, or political leaning of the newspaper titles. These functionalities allow for targeted investigations and comparisons, offering opportunities to develop a nuanced understanding of how language evolved in response to mechanisation across different contexts. The development of this interactive tool enhanced the accessibility and usability of the crowdsourced annotation data, and to deepen our analysis of semantic change associated with mechanisation.

2. Automatic detection of meaning change in the lexicon of mechanisation

The mechanisation of work that unfolded following the Industrial Revolution was not a uniform, linear phenomenon. Historians have long recognised that handcraft skills and mechanisation continued to exist side-by-side throughout Britain’s gradual rise to industrial pre-eminence.^[13] Linguistic variation across registers, social class, and geography, in addition to change over time, are complexities that require a deep engagement with historical scholarship, since it is a particularly complex topic to analyse quantitatively. In our investigation, we sought to explore the evolution of word meanings over the nineteenth century by focussing on words which previous historical and lexicological scholarship had identified as belonging to the key vocabulary related to mechanisation and as having registered significant meaning change across the period.

Our main point of departure was Manfred Gorlach’s discussion of language change in nineteenth-century English.^[14] We selected words mentioned as examples of meaning change that occurred as a result of mechanisation, specifically ‘bulb’, ‘coach’, ‘gear’, ‘matches’, ‘railway’, ‘stamp’, ‘stock’, ‘traffic’, ‘train’, ‘trade’, and ‘wheel’.

The corpus of historical British newspapers used to carry out our computational analysis comprises around 4.6 billion words and spans the period between 1801 and 1920. It includes titles digitised by the British Library in a project focusing on newspapers in poor condition that were mostly printed in London but also distributed outside London (around 2.3 billion words)^[15] and the titles specifically selected and digitised from the British Library newspaper collection for the Living with Machines project (a further 2.3 billion words).^[16]

As the flowchart in Figure 5.1 shows, we split the newspaper corpus into decade-long ‘time slices’, namely separate subcorpora containing only articles published within a certain year range, which could be achieved thanks to the associated newspaper metadata, including the year of publication. We decided to work with 10-year time slices to be able to track potential changes in the meaning of words from one decade (e.g. the 1840s, i.e. 1840-1849) to the next (e.g. the 1850s). Because of the large size of the corpus, we only applied minimal preprocessing, i.e. lowercasing and punctuation removal, but no lemmatisation (i.e. standardising of words to their base form, e.g. ‘making’ and ‘made’ to ‘make’). We then trained word vector representations (or ‘word embeddings’) with Word2Vec^[17] for each decade. Word embeddings represent words into a numerical form as points in a multidimensional space. The location of these points is determined by the linguistic contexts in which the words appear. Words that frequently appear in similar contexts (e.g. ‘coffee’ and ‘tea’) are positioned in close proximity because they share usage patterns (i.e. drinking, breakfast, hot beverages, etc.). This spatial arrangement of words allows for the modelling of the vocabulary as a landscape where semantic relationships can be visually and mathematically discerned. For instance, analogies between words can be captured by adding or subtracting the vectors of the words in question, as in the now classic example ‘king’ - ‘man’ + ‘woman’ ~ ‘queen’. By visualising the respective points, the semantic relations between these words will also be reflected in their position relative to each other, as in Figure 5.2.

Following an established practice in semantic change detection research,^[18] we ensured that the vectors for the same words across decades were comparable by ‘aligning’ all the embedding models to the model for the most recent decade available (1910s).^[19] The result was one embedding model for each decade in the nineteenth century, where every word occurring in the subcorpus for the respective decade is represented by a vector (i.e., a sequence of numbers which can be visualised as a point in a geometrical space), as shown in the schematised entries taken from the 1860s and 1870s models for ‘England’, ‘house’, ‘machine’, and ‘train’ in Figure 5.3.

At this point, we have aligned, hence comparable, representation of the same words across decades, namely, diachronic word embeddings. Working with digitised historical texts, however, often means dealing with Optical Character Recognition (OCR) errors, inevitably created during the newspaper digitisation process. A possible, though computationally expensive, way of partially counteracting the noise created in the embeddings trained on OCR'd texts is to attempt error post-correction. Doing so directly on the text sources would be computationally infeasible for very large data, such as our newspaper corpus, so we opted for updating the trained embeddings of spelling variants likely due to OCR errors, by averaging the respective vectors and keeping only the likely correct entry in our corpus.^[20] For example, the method found the entries ‘eugland’, ‘ngland’, and ‘rngland’ to be orthographic variants of the entry ‘england’ due to OCR errors. Therefore, the respective vector will be merged with the vector for ‘england’ by averaging them:

$v_{england} = \frac{v_{england} + v_{eugland} + v_{ngland} + v_{rngland}}{4}$

OCR errors can dramatically impact the usability of word embeddings for the linguistic analysis of historical data.^[21] Figure 5.4 shows the ten words with the most similar vector (i.e. the ten ‘nearest neighbours’) to the vector for ‘machine’ in the 1860s before and after merging the most likely orthographic variants of the word resulting from OCR errors. It is evident that, prior to merging OCR variants, nearly all the ten nearest neighbours of ‘machine’ are clearly OCR errors of the same word, whereas after OCR post-correction we mainly obtain types of ‘machine’.

However, the fact that OCR errors of a word are represented by very similar word vectors can be seen in itself as indirect evidence that the word vectors are capturing the semantic properties of the words in our corpus, despite the great amount of noise in the data. Spelling variants of a word, if not in themselves indicative of any differences in usage, are expected to have similar vector representations, which are generated using the distribution of the word form in the texts (i.e. their occurrence in the proximity of other particular words) as a reflection of, or proxy for, their meaning. That is, ‘maohines’ and ‘machines’ appear in roughly the same location in the model.

2.1 Computational exploration

In our exploration of the linguistic change catalysed by the Industrial Revolution,^[22] we compared the vector representation of selected words in the most recent decade (the 1910s) with each of the previous time slices. We used a well-established algorithm for detecting the number (if any) of significant changepoints in time-series data, to recognise potentially major and/or sudden changes in the semantics of the selected words at any point in the 19th century.^[23]

As shown in Figure 5.5, not only do the vector representation of the words ‘train’ and ‘railway’ become progressively more similar to the one in the 1910s (which is generally to be expected), but we also see a much more sudden change after the 1830s for ‘train’ and after the 1860s for ‘railway’. This corresponds to what we know of these words’ semantics from external, curated knowledge bases such as the Oxford English Dictionary, which record several new senses around the respective decades. As shown in Figure 5.6, other words, such as ‘gear’, show a more uneven trajectory, with several peaks and dips (or ‘troughs’, in time-series terminology), and an overall more gradual, stepwise change, rather than one major semantic-shifting decade.

Overarching comparisons between the paces with which words changed can also be obtained by clustering the semantic trajectories, such as those represented in Figures 5.4 and 5.5, using time-series analysis techniques. This can provide a way of making data-driven decisions about the next steps in the analyses, such as the kinds of variables the linguist, or historian, may want to focus on next. The observations made about the different types of semantic trajectories of ‘train’, ‘railway’, and ‘gear’ (as more abrupt versus more gradual), can be made at scale on several more words by applying a well-known time-series clustering technique called k-means with Dynamic Time Warping^[24], which takes the semantic trajectory of words (as cosine similarity scores) as input and clusters them based on this trajectory. In Figure 5.7, Cluster 1 corresponds to a more gradual semantic change trajectory, such as that of ‘gear’ shown above, whereas Cluster 2 corresponds to a trajectory with at least one major abrupt shift in the semantics of the word, as is the case with ‘train’ and ‘railway’. Together with ‘gear’, in Cluster 1 we find words such as ‘bike’, ‘bulb’, ‘car’, ‘match’, ‘machine’, ‘stock’, ‘trolley’, and ‘trade’; in Cluster 2, next to ‘train’ and ‘railway’, we see ‘coach’, ‘stamp’, ‘traffic’, and ‘wheel’.

2.2 Close-reading interpretation

The automatic changepoint detection method can be complemented with a close-reading interpretation of the nature of the semantic changes by comparing the sets of nearest neighbours of the relevant word before and after any detected change point. The changepoints detected for words such as ‘coach’, ‘gear’, ‘railway’, ‘train’, and ‘traffic’, for instance, are clearly connected to fast technological changes and their roles in the burgeoning transport system. Most often, the linguistic change manifests itself as an expansion of earlier sets of meanings, showing a transition from pre- to post-industrial usages.

However, not all meaning changes occurred at the same rate, as already mentioned. The semantic journey of ‘train’ from its initial association with long pieces of fabric, such as robes, skirts, or dresses, to its identification with railway transportation showed a rapid, almost complete overtaking by the latter usage of the former in the span of just two decades (see Figure 5.8).

Meanwhile, the word ‘railway’, whose change is closely associated with that of ‘train’, showed a somewhat more gradual overtaking by the new meaning, which remained as widespread in newspapers as the earlier usage (more generically as ‘any roadway along which wheels may run to facilitate the transport of heavy loads’) for at least three more decades after the sense specifically associated with trains became predominant around the 1860s (Figure 5.9).

We also captured much more gradual and complex changes. The trajectory of ‘gear’ (Figure 5.10) shows an incremental introduction of specific senses related to new mechanical advances throughout the nineteenth century. This, in turn, indicates that its semantic change has to do with different associations with other words becoming predominant at various points in time (e.g. ‘ironing gear’, ‘caulking gear’, ‘winnowing gear’, etc.).

This method also allowed us to highlight potential discrepancies with existing linguistic resources, such as the Oxford English Dictionary (OED) and its Thesaurus, in the nature and timing of semantic changes. For instance, the evolution of ‘coach’ (Figure 5.11), which emerged from automatically detecting changepoints, challenged the OED's classification of its 19th-century use as a ‘railway carriage’ as predominantly American English. Its semantic shift appears to be attested in England as well and may have occurred earlier than previously documented, offering a potential refinement to historical linguistic narratives.

2.3 The potential of newspaper metadata: preliminary insights on regional semantic shifts

Our analysis is only one possible way of exploring the multi-billion-word newspaper corpus to study semantic change related to the mechanisation process. In particular, we have only begun to uncover the full potential offered by the extensive metadata that can now be linked to each newspaper title thanks to the digitisation of historic reference works.^[25] This enhanced metadata includes diverse variables such as the geographical region, the declared political orientation of the newspapers, and their price at different points in time, all of which can provide useful insights into the complex interrelations between semantic change, technological innovation, and language use.

In a preliminary experiment, we used the place and year of publication of each newspaper to divide the corpus into two subcorpora containing articles published in two broad geographical regions, North and South England, historically corresponding to one of the main socio-political divides in Britain, before further splitting each geographical subcorpus into temporal subcorpora. After automatically detecting potential changepoints for words such as ‘bulb’, ‘coach’, ‘gear’, ‘matches’, ‘railway’, ‘stamp’, ‘trade’, ‘traffic’, ‘train’, and ‘wheel’, we compared the results from the two subcorpora against each other to assess whether their semantic shift occurred virtually simultaneously across the two regions or whether some degree of diatopic (i.e. spatial) variation could be posited. As an exploratory step, we also considered potential differences in usage between singular and plural forms of the same words.

Our preliminary results showed, in fact, differences between the North and South of England. While for some words, like ‘match’ and ‘stamp’, a changepoint was detected for both regions, for other words, such as ‘machines’ and ‘stock’, a changepoint was only detected for the South, and for others, like ‘trade’, ‘bulbs’, and ‘cars’, only for the North. Moreover, even for words with a potential changepoint in both regions, the decade in which the shift occurred may differ: for ‘match’, for instance, a changepoint was detected later in the century for the North than for the South. ‘Stamp’ appears to have had a sudden change in usage in North England from its older generic meaning of ‘marking’, ‘engraving’, or ‘impression’ to its newer philatelic sense, as becomes clear after inspecting which words are most closely associated with stamp before and after the detected changepoint. Similarly the word cars seems to have had a sudden change in the North from its older existing meaning of a wheeled, usually horse-drawn conveyance to that associated with railway carriages or wagons. In the South, the newer meanings of both words are attested but do not induce a changepoint detectable by the algorithm, suggesting a less abrupt and less intense change in usage. Figures 5.12 and 5.13 compare the change in the nearest neighbours of the word ‘cars’ in the North and South of England, respectively.

For some words, a changepoint was detected in a region either only in the singular or the plural form, as in ‘match’ (but not ‘matches’) in the South or ‘stamp’ (but not ‘stamps’) in the North of England; in other cases, a changepoint was detected earlier for one of the two forms. These subtle differences, which warrant further exploration, may have to do with the different usages and therefore different triggers for semantic change associated with some words in the singular and plural, for example when they are used as plural generics (e.g. ‘cars are an amazing invention’), as opposed to referring to multiple instances of a concrete object (e.g. ‘there are cars on the street’).^[26]

These are, however, preliminary hypotheses and no systematic attempt was made to validate and interpret the results. Only further more rigorous analyses, such as those implemented by Pedrazzini and McGillivray and outlined in Sections 2 and 2.1, will allow us to establish what factors most strongly drove diatopic variation in the process of semantic change.^[27]

3. Designing sense disambiguation tasks for volunteers

How did the words trolley, car, coach and cycle change over time? The early results of the changepoint detection work described above fed into our work engaging the public with computational linguistic work by designing semantic annotation tasks on the Zooniverse citizen science platform. We invited the public to contribute to our data science research by annotating individual instances of specific target words with historical senses from the OED.^[28]

As discussed further in chapter 6, unlike commercial, Mechanical Turk-style projects with financial rewards, volunteer participants in crowdsourcing and citizen science are motivated by taking part in an enjoyable and meaningful task with interesting collections.^[29]This creates a challenge in that tasks should make a clear contribution to a wider issue or research area while completing them should be both clear and interesting enough to motivate ongoing contributions. Our work on the 'language of mechanisation' workflows built on the lessons learnt from earlier collaboration on the 'What was a machine?' workflows described in chapter 6, which included our first semantic annotation task and a text summarisation / transcription task. It also provided an opportunity to explore the potential of voluntary crowdsourcing to provide data for Digital Humanities methods such as the development of computational linguistics algorithms.

Our changepoint detection analysis provided a list of potential target keywords of interest. We reviewed the historical senses listed for each in the OED, looking for senses that were distinct enough to make it easy for volunteers to choose the right sense for a given word in context. While we considered a range of words, ultimately we designed tasks around the words ‘trolley’, ‘car’, ‘coach’ and ‘cycle’ (‘bicycle’ or ‘motorcycle’). We set up four 'workflows' (a group of sequential tasks on a specific image from a historical newspaper), each focused on a specific target keyword. See Figure 5.14 for an example (for further details see Appendix 1).

Having established relationships with volunteers through earlier projects on Zooniverse, we were able to share our work in progress and invite feedback from them. Their feedback helped refine our design of novel tasks that required volunteers to read historical texts closely. We undertook a range of outreach and marketing activities when the workflows were ready to 'go live'.^[30] Volunteer contributions via our workflows created a rich set of annotations. The annotated articles are available for download and re-use in the British Library's research repository, and have been extensively documented in readme files and a data paper.^[31]

4 Visualisations and analysis

Following the crowdsourcing work described above, we shared the data with King’s Digital Lab (KDL, King’s College London) so that they could help analyse and visualise the corpus. The KDL team collaborated with the Living with Machines project team in the late summer of 2023 over roughly two months to produce three products in the form of Notebooks which were used to inform the analysis that follows. The design and implementation of the Notebooks is detailed in Appendix 2.

4.1 The evolution of trolley: from hand-powered cart to fuel-powered waggon

According to the Oxford English Dictionary (OED), early examples of the word ‘trolley’ in English usage represent local, vernacular words for a low, narrow cart, particularly one designed for use in constricted spaces.^[32] The OED editors cite Edward Moor’s book Suffolk Words and Phrases (1823) as the first attested usage, underscoring the word’s regional dimension in the early nineteenth century. Our analysis appeared to throw up three examples before this date, but manual inspection suggests that these are all false positives. Our first confirmed uses come from 1850, when we have two cases. In both cases the local newspaper felt the need to explain the word’s meaning, suggesting that ‘trolley’ was still viewed as a neologism at this point (both instances explained it was a type of waggon, one specifying what type: ‘in the trolley, or low waggon’). But, as the OED records, by the 1850s ‘trolley’ (or sometimes ‘trolly’ or ‘trawley’) was also being used more specifically to refer to low waggons designed to run on rails or other types of track, either on the railways or in and around factories and docks. Interestingly, the present-day dominant meaning, where ‘trolley’ is used to denote a hand-powered means of conveyance on small wheels or castors (as in tea-trolley or supermarket-trolley), arose much later, beyond the period covered in our analysis (the OED’s first attested use of this sense is an account of someone purchasing a tea-trolley in 1937). Although both early senses exist in parallel from the 1850s, our analysis shows that the balance of usage shifted across the period 1850-1920.

In our analysis we therefore focus on the shifting balance between uses of ‘trolley’ that refer to a small waggon pulled by human or animal power, and those referring to a steam or electric-powered vehicle, usually mounted on rails.

As Figure 5.15 extracted from the project Observable Notebook shows, the number of annotations available for the early years of our corpus (before 1860) is very low, and this holds across the board for all annotation categories identified, i.e. the pre-industrial sense of ‘trolley’, the mechanisation sense, and other, as well as the cases in which the annotators were not able to make a decision. The increase in annotations after 1860 corresponds to the larger size of the more recent portion of our corpus, reflecting the growth of the press and its evolving approach to vernacular usage.

When we consider the relative frequencies of the two main senses considered we can assess their complementary distribution in the corpus. By factoring out the growing size of our corpus we can focus on the trajectory of the pre-mechanisation and post-mechanisation meanings. We excluded the “Other” label from this analysis as this category grouped together annotations which displayed a level of ambiguity in the contexts. Figure 5.16 shows the number of annotations for each meaning divided by the total number of annotations available for ‘trolley’ in each year. It clearly shows how the sense “a trolley-car powered by steam or electricity” overtook the original sense in the late nineteenth century.

In Figures 5.16 and 5.17 the curves in the plots are created using a monotone-x curve, i.e. a curve interpolation method that ensures a smooth increasing or decreasing curve when visualising data. The trend lines (straight) in the plots represent the general direction or pattern of the annotation counts for that word per year and are generated using linear regression. The accuracy of these trend lines is affected by the limited size of the dataset before 1860, and therefore should be interpreted with caution. Nevertheless the main trends confirm our expectations, with the sense related to mechanisation gradually taking over the original one.

The shaded area in the plot represents the potential variability or imprecision in the data, which explains why they show values above 100 per cent or below per cent. The range is calculated using confidence intervals for proportions.

The meaning of ‘trolley’ as cart powered by people or animals is mainly associated with a North American use; however, the annotations data under examination seem to show it as dominant in British sources too; this might be due to sources that refer to reports from North America so a more detailed analysis would be needed to substantiate this claim. It should be noted that we have relatively few data points for ‘trolley’ with a total of 561 articles with an agreement rate greater or equal to 65 per cent.

Given the relatively small number of confirmed uses we are dealing with for the word ‘trolley’, any overall conclusions must be tentative. In Figure 5.18 our statistical analysis identifies two key moments of semantic change, 1860 and 1890. The change point detection algorithm is a method used to detect abrupt changes or shifts in time-series data. It is useful when dealing with data that exhibits constant behaviour, where the data remains relatively constant between certain points in time but undergoes sudden changes at other points. The first point, in 1860, seems to be associated with the widespread adoption of the regional vernacular word ‘trolley’ as a word to describe a low waggon used for moving people or goods. The second, in 1890, appears to be associated with the shift towards meanings associated with powered vehicles, even though Britain did not adopt the American term ‘trolley-car’, preferring to call its powered public transit vehicles ‘trams’. Manual inspection suggests two reasons for this. First, British newspapers carried extensive reports of American news which tended to retain the vernacular usage, perhaps for added colour or authenticity. Second, specific parts of the tram, notably the connecting rod that picked up electric current, adopted the American usage; we find frequent reports about incidents involving the ‘trolley poll’ of British municipal trams. Together, these two influences, the allure of American news and the adoption of the American term for parts of British trams appear to explain the meaning shift about 1890.

4.2 Mechanisation and meaning-shift in other transport words: bike, coach, and car

We repeated this analysis with three additional words connected with transport that are known to have undergone significant semantic change, and where mechanisation was thought likely to have played its part in these linguistic processes: ‘bike’, ‘coach’, and ‘car’. Of these three, only ‘bike’ was a neologism, being an Anglo-Saxon shortening of ‘bicycle’, which the OED records was itself borrowed from French in the 1860s at the height of the first cycling craze (the first English use of ‘bicycle’ cited by the OED is a Daily News story from 1868).^[33] The earliest recorded use of the shortening ‘bike’ is from a cycling trade magazine in 1880, a few years before the great fin-de-siecle cycling boom generated by the invention first of the safety bicycle and then the pneumatic tyre in the 1880s.^[34] The OED’s first attested usage of ‘bike’ to mean motorcycle is from Automobile Topics, a specialist American motoring magazine, in 1903; their first British example comes from a letter written by T.E. Lawrence to George Bernard Shaw in 1924. Once again mechanisation and Americanisation appear to be deeply intertwined in this process of semantic change.

The same cannot be said for the other two transport words we analysed: ‘coach’ and ‘car’. Both have deep roots in British English as nouns describing vehicles for conveying people and/or goods. The OED’s earliest examples of ‘coach’ meaning a large, horse-drawn carriage date from Elizabethan times (as ‘coche’ and ‘cosche’); their first usage with the spelling ‘coach’ is from the diary of Samuel Pepys in 1665.^[35] The origins of the word ‘car’ are even older, with the first attested uses (with various spellings) dating from the fourteenth and fifteenth centuries. Like ‘coach’, it was a word for a horse-drawn conveyance, and could be used to refer to anything from a humble cart to a royal carriage.^[36] Both words registered significant semantic change in the age of mechanisation, with North America again playing a part in this story. In British English, ‘coach’ was initially adopted to denote the covered, especially first-class, railway carriages on early trains. Later in the nineteenth-century, it became used more generally to denote any railway carriage, but phrases like ‘Pullman coach’ retained the earlier sense of luxury. The predominant present-day meaning (i.e. a comfortably equipped motor bus designed for long journeys) appears to have had its origins in pre-1914 America, the first British usage cited by the OED dates from 1930.^[37] In British English, the OED records that the meaning of ‘car’ was extended to denote the passenger compartment of various new modes of transport from the late eighteenth century, including air balloons and cable cars, but unlike in America, the word was not widely used to denote railway passenger compartments (instead ‘coach’ or ‘carriage’ became the preferred terms). The OED’s earliest citations for the dominant present-day usage of ‘car’, both in Britain and North America (i.e. a motorised, private road vehicle for people), date from 1896 and 1900, when the motor car first began to grip the public imagination in industrialised societies.^[38] Crucially, it appears that all three words continued to be used interchangeably to denote both mechanised and non-mechanised vehicles, with lexical context the only way to signal the intended meaning. Our aim was to use diachronic analysis to map how the balance of semantic meaning shifted in local newspapers across the long nineteenth century.

Some of the Notebook graphs are helpful to offer insights and spur reflection on the decades when changes of meanings occur. For example, in the case of ‘bike’, the annotations data seem to support a claim that its meaning as motorbike, powered by an engine, occurs quite late around 1910, while the meaning of bike powered by human legs seems to take off from the 1880s. It is indeed only during the interwar period that motorbikes (briefly) became the predominant means of private motorised road transport.

According to the annotations data at hand, new meanings for the word ‘coach’ seem to occur in the decade starting in 1840; however, these clearly do not take over the horse-drawn carriage meaning; indeed, other words for motor bus were in use, such as the more common ‘charabanc’. An onomasiological analysis which looks at the various words used to express this concept would be needed to substantiate these insights; the motor bus meaning would come into use when long journeys became possible, in the interwar period.

With respect to ‘car’, not surprisingly, the road vehicle meaning started to take off in the early 1900s. In Figure 5.23 the parameters have been tweaked to detect further change points.

A first change point occurs around 1850 where the two new meanings (electric tramcar and road vehicle, not yet the private motor car) appear and begin to rise; these remain subdued, however, until the second change point around the 1900s, where the motorised road vehicle meaning takes off and other meanings diminish or remain at a low level. This change point marks the explosive growth of the private motor car as a presence not just on British roads, but in the national psyche. This is the moment that saw Kenneth Grahame create Mr Toad as a demonic, car-obsessed road hog in The Wind and the Willows (1908), and road safety concerns fed the call for drivers’ licences, speed limits and measures to control the speeding driver.^[39]

5 Conclusion

In this chapter we reported on multifaceted exploration of semantic change during the era of mechanisation. Semantic change is a highly multifaceted phenomenon, where external factors, such as cultural influences and the adoption of terminology from different regions, significantly impact the semantic trajectory of words. At the same time, words do not change meaning in isolation of each other, and connected changes add to this complexity. Our analysis leveraged computational algorithms, manual annotated data, and close reading and added a new quantitative dimension to our notion of semantic change.

Our computational approach, based on diachronic word embeddings, enabled us to identify both gradual and abrupt semantic shifts in key mechanisation-related terms over the nineteenth century such as ‘train’, ‘railway’, and ‘gear’, shedding light on the nuanced trajectories of linguistic change catalysed by industrialisation. The employment of time series analysis techniques allowed us to detect significant changepoints, offering insights into the temporal dynamics of semantic evolution. Furthermore, close-reading interpretation complemented automated methods, providing rich contextual understanding of the semantic shifts. These insights partially align with historical accounts of the change in the English lexicon in response to advancements in transportation and machinery, reinforcing the notion of language as a dynamic reflection of socio-cultural and technological change. For example, the semantic trajectory of ‘train’ from its earlier association with fabric to its current usage in railway transportation mirrored historical narratives of the transformative impact of steam locomotion on transportation systems. Similarly, the semantic evolution of ‘gear’ demonstrated incremental changes in meaning corresponding to advancements in mechanical engineering and industrial production.

At the same time, our analysis revealed semantic shifts not extensively documented in existing linguistic resources and dictionaries. This is particularly evident in the case of ‘coach’, where we showed that its nineteenth-century usage as a ‘railway carriage’, attributed predominantly to American English in the OED, had significance in English discourse as well, potentially occurring earlier than previously recognized. Our investigation extended beyond a diachronic analysis to incorporate regional variations and explore diatopic shifts in meaning. Our preliminary findings identified potential differences in the semantic trajectories between North and South England, pointing to the interplay between geography, language use, and technological innovation during the mechanisation era, and stressing the importance of considering socio-geographic factors in the study of linguistic change across different linguistic communities.

The voluntary crowdsourcing annotation process leveraged the collective efforts of multiple annotators to annotate individual instances of trolley, coach, car, and bike in nineteenth-century British newspapers with historical senses from the OED. The design of these tasks involved collaboration between researchers and volunteers to develop the first experiment in crowdsourced semantic annotation of historical texts that we are aware of. The crowdsourced annotation data provided the starting point for a series of interactive visualisations that allowed us to examine words potentially changing meaning during the nineteenth century. The process involved several iterations, integrating user research, project requirements, and technical solutions. The visualisations found evidence of the semantic shifts that the words of interest underwent, and highlighted some of the complexities inherent in any empirical analysis of semantic change. At the same time, the low number of annotated instances, particularly in the early years of our corpus, calls for some caution in the interpretation of the data, while the growing size of our corpus in the second half of the nineteenth century reflects the broader expansion of the press during that period and possibly its opening up to vernacular usage.

The analysis of ‘trolley’ found evidence of the newer meaning ‘a trolley-car powered by steam of electricity’ overtaking the earlier meaning in the late nineteenth century, with two key moments of semantic change identified around 1860 and 1890, corresponding with shifts in regional vernacular usage and influences from American English. A similar connection between semantic change and Americanisation affected the neologism ‘bike’, which originally referred to human-powered bicycles but later extended to encompass motorbikes, particularly during the interwar period when motorbikes gained popularity. On the other hand, ‘coach’ had its origins in large, horse-drawn carriages from Elizabethan times, and over time was used to refer to covered railway carriages on early trains and eventually evolved to denote motor buses for long journeys. Similarly, ‘car’ initially described various horse-drawn conveyances, later evolving to include passenger compartments of new transport modes. The shifts in meaning of the three terms ‘bike’, ‘coach’, and ‘car’ correlated with the rise of motorised road vehicles and the popularisation of the private motor car, although this transition was not linear or uniform, as all were employed interchangeably to refer to both motorised and non-motorised vehicles, relying on contextual clues to convey the intended sense. For these reasons, our diachronic analysis aimed to chart the evolution of semantic nuances in the newspapers throughout the extended nineteenth century.

While our analysis has provided valuable insights into the semantic evolution of mechanisation-related vocabulary during the nineteenth century and has helped paint a more nuanced picture than a mere binary shift from an older meaning to a newer one, it is essential to acknowledge the limitations of our approach. Despite efforts to mitigate noise, the presence of OCR errors in historical texts posed a significant challenge, potentially impacting the accuracy and reliability of our semantic analyses. Although we implemented post-correction techniques to address OCR inaccuracies automatically, the extent to which these corrections influence the semantic trajectories of words remains a point of uncertainty, underscoring the need for further methodological refinement. Moreover, our preliminary exploration of regional semantic shifts highlighted the complexities of diatopic variation but did not systematically validate or interpret the results. As such, a more rigorous interrogation of regional differences in language use, informed by historical, sociolinguistic, and geographical insights, would be needed to contextualise our findings accurately.

Manual annotation and close reading analysis has highlighted the importance of considering the context, as words were often used interchangeably to denote both mechanised and non-mechanised vehicles, requiring careful interpretation based on lexical context. Given the considerable time resources needed to conduct manual semantic annotation and to ensure high agreement between annotators, its limited extent posed a notable constraint on our results. The limited amount of manual annotation restricted the robustness of the insights we were able to derive from it. Expanding the scope of manual annotation in the future would enhance the richness and accuracy of the results. Additionally, a more extensive manual annotation process would facilitate the validation and refinement of automated algorithms, underscoring the importance of investing in more extensive manual semantic annotation efforts to augment the depth and rigour of semantic change research.

As pointed out in our methodological and process-related discussion, several further valuable lessons have emerged, guiding future work at the intersection between semantic change and historical research. In line with the overarching ethos of the broader “Living with machines” project, we stress the significance of interdisciplinary collaboration, without which this work package would have not succeeded. The synergy among computational linguists, historians, research software engineers, and library curators has been instrumental to this work, and made it possible to integrate diverse methodological approaches, and manual and automated processes. The result has been a richer and more multifaceted understanding of the evolution of historical language and, with it, of complex historical phenomena.

References

Armaselu, Florentina, Apostol, Elena-Simona, Khan, Anas Fahad, Liebeskind, Chaya, McGillivray, Barbara, Truică, Ciprian-Octavian, Utka, Andrius, Valūnaitė Oleškevičienė, Giedrė, van Erp, Marieke, ‘LL(O)D and NLP perspectives on semantic change for humanities research’, Semantic Web, Volume 13, Issue 6, September 2022, Pages 1051-1080, https://10.3233/SW-222848.

Beelen, Kaspar, Jon Lawrence, Daniel C S Wilson, David Beavan, ‘Bias and representativeness in digitized newspaper collections: Introducing the environmental scan’, Digital Scholarship in the Humanities, Volume 38, Issue 1, April 2023, Pages 1–22, https://doi.org/10.1093/llc/fqac037

Bergs, Alexander and Brinton, Laurel J. (Eds) (2012). English Historical Linguistics. De Gruyter Mouton, Berlin, Boston.

Filgueira Vicente, Rosa, Michael Jackson, Anna Roubickova, Amrey Krause, Tessa Hauswedell, Julianne Nyhan, David Beavan, et al. ‘Defoe: A Spark-Based Toolbox for Analysing Digital Historical Textual Data’. In 2019 IEEE 15th International Conference on E-Science (e-Science), 2019. https://www.research.ed.ac.uk/portal/en/publications/defoe-a-sparkbased-toolbox-for-analysing-digital-historical-textual-data(a9682a06-d159-4d00-9f92-f9a27f0bb385).html.

Gorlach, M. (1999). English in Nineteenth-Century England. Cambridge: Cambridge University Press.

Hamilton, William L., Jure Leskovec, and Dan Jurafsky. 2016. Diachronic word embeddings reveal statistical laws of semantic change. In Proceedings of the 54th Annual Meeting of the Association for Compuational Linguistics (Volume 1: Long Papers), pages 1489–1501, Berlin, Germany. Association for Computational Linguistics.

Kay, C. and Allan, K. L. (2015). English Historical Semantics. Edinburgh : Edinburgh University Press

Ketabgian, Tamara (2011). The Lives of Machines: The Industrial Imaginary in Victorian Literature and Culture. University of Michigan Press.

Killick, Roberta, Paul Fearnhead, and Idris A. Eckley. 2012. ‘Optimal detection of changepoints with a linear computational cost’. Journal of the American Statistical Association, 107, 500 (2012): 1590–1598.

Leslie, Sarah J., Sangeet Khemlan, Sandeep Prasada and Sam Glucksberg. “Conceptual and linguistic distinctions between singular and plural generics,” Proceedings of the 31st Annual Cognitive Science Society, (2009): 479-484.

Manlove, Colin (1993). Charles Kingsley, H. G. Wells, and the Machine in Victorian Fiction. Nineteenth-Century Literature 48 (2): 212–239.

Mari, Alda, Claire Beyssade and Fabio Del Prete. Genericity. Oxford: Oxford University Press, 2012.

McGillivray, B. (2022). How to Use Word Embeddings for Natural Language Processing. SAGE Research Methods: Doing Research Online. https://doi.org/10.4135/9781529609578

Mikolov, Tomas, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in vector space.

Pedrazzini, Nilo and Barbara McGillivray. 2022. Machines in the media: semantic change in the lexicon of mechanization in 19th-century British newspapers. In Proceedings of the 2nd International Workshop on Natural Language Processing for Digital Humanities, pages 85–95, Taipei, Taiwan. Association for Computational Linguistics.

Řehůřek, Radim and Petr Sojka. 2010. Software Framework for Topic Modelling with Large Corpora. In Proceedings of the LREC 2010 Workshop on New Challenges for NLP Frameworks, pages 45– 50, Valletta, Malta. ELRA. http://is.muni.cz/publication/884893/en

Ridge, Mia. 2013. ‘From Tagging to Theorizing: Deepening Engagement with Cultural Heritage through Crowdsourcing’. Curator: The Museum Journal 56 (4).

Ridge, Mia, Samantha Blickhan, Meghan Ferriter, Austin Mast, Ben Brumfield, Brendon Wilkins, Daria Cybulska, et al. 2021. The Collective Wisdom Handbook: Perspectives on Crowdsourcing in Cultural Heritage. Community Review edition.https://doi.org/10.21428/a5d7554f.1b80974b.

Ridge, Mia, and Nilo Pedrazzini. ‘Public Domain Newspaper Titles in Living with Machines’. Living with Machines (blog), 7 May 2024. https://livingwithmachines.ac.uk/public-domain-newspaper-titles-in-living-with-machines/.

Sakoe, H. and S. Chiba, "Dynamic programming algorithm optimization for spoken word recognition," in IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 26, no. 1, pp. 43-49, February 1978, doi: 10.1109/TASSP.1978.1163055.

Spufford, Francis and Uglow, Jenny (1996). Cultural Babbage: Technology, Time and Invention. London, Boston: Faber and Faber.

Sussman, Herbert (1969). Victorians and the Machine : The Literary Response to Technology. Harvard University Press.

Sussman, Herbert (2009). Victorian Technology: Invention, Innovation, and the Rise of the Machine. Praeger.

Tahmasebi, Nina, Lars Borin, Adam Jatowt, Yang Xu, and Simon Hengchen (Eds.) (2021). Computational approaches to semantic change. Berlin: Language Science Press.

Tolfo, Giorgia, Olivia Vane, Kaspar Beelen, Kasra Hosseini, Jon Lawrence, David Beavan and Katherine McDonough. 2023. Hunting for Treasure: Living with Machines and the British Library Newspaper Collection. In Estelle Bunout, Maud Ehrmann and Frédéric Clavert (Eds.), Digitised Newspapers – A New Eldorado for Historians?: Reflections on Tools, Methods and Epistemology, 23-46. Berlin, Boston: De Gruyter. DOI:10.1515/9783110729214-002.

van Strien, Daniel, Kaspar Beelen, Mariona Coll Ardanuy, Kasra Hosseini, Barbara McGillivray, and Giovanni Colavizza (2022). “Assessing the Impact of OCR Quality on Downstream NLP Tasks.” In Proceedings of the 12th International Conference on Agents Artificial Intelligence (ICAART2020)-Volume, pages 484-496.

Appendix 1 – Crowdsourcing design and implementation

528 individual volunteers contributed to the 'language of mechanisation' Zooniverse workflows to complete a series of tasks related to the analysis of historical newspaper articles (subjects in the data model below, Figure 5.31). The tasks focused on the analysis of the words ‘trolley’, ‘car’, ‘coach’, and ‘bicycle’/‘motorcycle’ and involved classifying each article according to the meaning that those words conveyed in context. Data about the newspaper articles came from the digitised newspapers, combined with the information available in Mitchell’s Press Directories (https://en.wikipedia.org/wiki/Mitchell%27s_Press_Directories) such as price and political leaning, processed by the Living with Machines project. The ‘Explore the data sources’ section of the Observable Notebook around the language of mechanisation presents some of this data as showcased below.

Whenever using human annotated data for research, measuring the inter-annotator agreement, i.e. the extent to which the annotators produced the same annotation for the same initial data, helps the researchers assess the reliability and challenges of the annotation.

We calculated agreement rate and frequency as follows. We grouped all annotations available per article. For each article, we calculated the total number of annotations for that article and identified the annotation that was used the most times for the article, we refer to this as “max”. We calculated the agreement rate by dividing the max by the total of annotations. Finally, we counted the number of times each agreement rate occurs. The Notebook allows the users to set a value for the minimum agreement required. We set this value to 65 per cent to ensure that at least two thirds of the annotators (typically two out of three) agreed. The analysis presented in section 4.2 is based on the annotations for which this minimum value was reached.

While the number of articles for which meanings were annotated vary (371 for ‘bike’, 1,243 for ‘coach’, 436 for ‘car’), proportionally the higher agreement is found for the word ‘coach’, where 1,243 articles reach an agreement rate of annotation greater or equal to 65% (i.e. at least 2 out of 3 annotators agree on the meaning associated to the occurrence of ‘coach’).

Appendix 2 - Design and implementation of Visualisation and Analysis Notebooks

Two of the notebooks designed and developed by KDL in 2023, are relevant for the analysis of the language of mechanisation discussed in section 4, namely:

1. A GitHub repository (https://github.com/kingsdigitallab/lwm-davizct/) with code to produce Jupyter Notebooks to analyse and visualise various crowdsourcing tasks on the Zooniverse platform to provide insight into the overall annotation activity across workflows. One of the main goals of this work was to make the code used for this project reproducible so that others can use or adapt it to visualise data from their own crowdsourcing projects;
2. An Observable interactive Notebook (https://observablehq.com/@jmiguelv/language-of-mechanisation) to analyse words that potentially changed meaning during the nineteenth century, examining their connection to the process of mechanisation in British society, specifically focused on terms used to describe vehicles.

The design and development phase of these Notebooks was concentrated within approximately 10 development iterations, grouped into increments and scheduled into a set of 2-week periods of development called timeboxes or sprints. In the following section we will describe our design process that produced these two products and associated visualisations, and walk the readers through a series of visualisations to support the analysis.

Technologies and processes

The high-level Zooniverse data model that underpins the crowdsourced datasets fed to the Language of Mechanisation Notebook is structured around subjects (articles/images) and classifications (annotations on the articles/images as described in Section 3 above).

The components of the data model can be summarised as follows:

• workflow: a series of tasks that accomplish a specific purpose on selected subject sets;
• user: a Zooniverse volunteer contributor;
• subject: a media file shown to volunteers during a task. In this case an image of a newspaper article;
• subject_set: groups or 'folders' of subjects;
• annotation: a value assigned to a subject by a volunteer when completing a task in a workflow, for example, a transcription, classification or count;
• comment: all posts on a Zooniverse project Talk board, which can be about specific subjects or general discussion threads;
• tag: hashtags used in comments.

Once data are collected from Zooniverse the following steps are applied to data workflows using tools developed by the Living with Machines team:

• Any duplicates in the subjects are merged;
• The relevant data are selected;
• High-level metadata are aggregated;
• Subjects, annotations and data sources are selected.

Jupyter Notebooks were chosen by the project team as a technical platform as they are a widely adopted interactive environment that any Zooniverse project manager could use as a starting point to visualise information about their crowdsourcing workflows. Our initial Notebooks were later extended to prepare and export data for further data visualisation and analysis in other platforms.

Observable was selected as the platform to deliver the Language of Mechanisation visualisations because it offered a powerful platform for data visualisation and exploration that enabled collaborative and rapid prototyping. For example, team members could work together iteratively developing visualisations and sharing insights in real-time. Its interactive nature allowed for experimentation and quick iteration, facilitating the exploration of Living with Machines’s complex datasets. Additionally, Observable's integration of code and visuals, along with its Notebook-style interface, aligned with KDL and project values in terms of transparency, reproducibility, and documentation, as well as project objectives, making the material suitable for data analysis, storytelling, and knowledge sharing.The design process to deliver the products for this project derives from the double diamond process.^[40] It consists of four phases with relatively quick iterations within each phase.

Phase 1 – Discover

Part of the discovery phase included some user research - supported by open-ended questionnaires - to reach an agreed understanding of the project scope, objectives and audiences as well as for King’s Digital Lab team to gain familiarisation with the data points of the crowdsourced datasets inclusive of newspapers metadata. The main results of this phase were: (1) the definition of a high-level data model that would underpin interactions with the datasets by potential users; (2) analysis work that informed the following phases.

Phase 2 – Define

The outcomes of the discovery phase were integrated with the project requirements (defined at high-level prior to the project starting) and tasks to guide development were created or integrated as needed in an iterative fashion following review of each part of the technical solution (also called an increment). A Miro (https://miro.com/) board was used to support this phase and guide future design iterations.

Phase 3 - Develop

A range of different technical solutions - from off the shelf plotting libraries to custom built self-hosted sites - were initially discussed and tested; however, given the time and budget at our disposal, the options quickly converged to adopt existing Notebook platforms (Jupyter and Observable respectively). With the platforms default design system in place, there were obvious user interface constraints and customisation of these platforms would have not been practical; the focus of our work shifted to improving the user experience. Considerable effort went into organising the narrative around the visualisations, improving visual elements, accessibility, and adding explanatory text. We explored and tested different chart types, interactions, and aimed for a simplified terminology and language for what was possible, to encourage engagement and facilitate understanding.Digital accessibility testing and statement – The design and development of the interactive Notebook was a very small part of the overall project; the choice of the Observable platform for this imposed limitations which meant that a full accessibility assessment could not be performed on the page. KDL and theLwM project both prioritised digital accessibility in their work to enable as many people as possible to access the information, and we set aside resources to improve the issues we found. Some limitations we could not overcome are as follows:

• Visual limitation: most sections of the Notebook were labelled but some charts include dimensions or data elements that are defined only by colour, causing potential difficulty for users with visual colour impairment.
• Device limitation: The Notebook is responsive (i.e. adapts well to mobile as well as desktop devices) but it is better viewed on larger screens as some of the charts might not display or could be hard to view due to the lack of space.
• Browser issues: feedback suggested that some of the charts do not load on certain browsers; the table of contents links do not work in Safari browser.

These issues were noted in the accessibility statement of the Observable Notebook product with a call for user feedback to raise any other issues that would be identified.

Phase 4 - Deploy

Adjustments were iteratively made in this phase based on a pragmatic selection of the feedback given by the users interacting with the Notebook.Near the end of the Develop phase, KDL and the research team set up a quick usability test to stay within budget constraints. KDL prepared a set of questions for test participants to answer - partners gathered participants and conducted the usability test. There were two separate tests for the two outputs. Both had about 8-10 participants recruited from the British Library and wider LwM team. There were no set tasks and the questions were open-ended with the aim to get quick feedback to identify top issues. KDL collated the responses and gave suggestions to the research team on what we found were the top issues and worked together to implement those improvements.

1. M. Gorlach, English in Nineteenth-Century England (Cambridge: Cambridge University Press, 1999); C. Kay and K.L. Allan, English Historical Semantics (Edinburgh: Edinburgh University Press, 2015). ↑

2. Francis Spufford and Jenny Uglow, Cultural Babbage: Technology, Time and Invention (London, Boston: Faber and Faber, 1996); Herbert Sussman, Victorian Technology: Invention, Innovation, and the Rise of the Machine (Westport CT: Praeger, 2009). ↑

3. BMcG conceived the scope of the work reported on in this chapter and led the design and implementation of the research and the writing process; she also drafted and edited the first and last sections of this chapter. As typical for any project King’s Digital Lab (KDL) undertakes, the design of the visualisations discussed under section 4 was conducted collaboratively with members of the Living with Machines project (Barbara McGillivray, Nilo Pedrazzini, Mia Ridge, Kalle Westerling) and coordinated by KDL sub-team of Research Software (RS) Analyst (Arianna Ciula), RS Engineer (Miguel Vieira), and UI/UX Designer (Tiffany Ong). Other KDL team members contributed to management and administration (Lab Manager Pam Mellen). NP drafted section 2 and co-led (with BMcG) the experiments described therein; MR led the design and implementation of the crowdsourced annotation tasks; JL acted as historical adviser and co-wrote historical interpretations in Section 4. ↑

4. Herbert Sussman, Victorians and the Machine: The Literary Response to Technology (Cambridge MA: Harvard University Press, 1969); Colin Manlove, ‘Charles Kingsley, H. G. Wells, and the Machine in Victorian Fiction,’ Nineteenth-Century Literature 48, 2 (1993): 212–239; Tamara Ketabgian, The Lives of Machines: The Industrial Imaginary in Victorian Literature and Culture (Ann Arbor, MI: University of Michigan Press, 2011). ↑

5. Alexander Bergs and Laurel J. Brinton (eds), English Historical Linguistics (Berlin, Boston: De Gruyter Mouton, 2012), 1063 ff. ↑

6. Gorlach, English in Nineteenth-Century England, 10-25; David F. Mitch, The Rise of Popular Literacy in Victorian England: The Influence of Private Choice and Public Policy (Philadelphia, PA: University of Pennsylvania Press, 1991); David Vincent, Literacy and Popular Culture: England, 1750-1914 (Cambridge: Cambridge University Press, 1989). ↑

7. Martin Hewitt, The Dawn of the Cheap Press in Victorian Britain: the End of the 'Taxes on Knowledge', 1849-1869 (London: Bloomsbury Academic, 2014); Emma Griffin, ‘The Making of the Chartists: Popular Politics and Working-class Autobiography in Early Victorian Britain’, English Historical Review, 129.538 (2014), 578-605, Doi: 10.1093/ehr/ceu078; Simon Eliot, ‘Circulating Libraries in the Victorian Age and After’, in The Cambridge History of Libraries in Britain and Ireland, vol. 3: 1850-2000, ed. by Alistair Black and Peter Hoare (Cambridge: Cambridge University Press, 2006), 125-46. ↑

8. Gorlach, English in Nineteenth-Century England, 92 ff.; Bergs and Brinton, English Historical Linguistics; Kay and Allan, English Historical Semantics, 28ff. ↑

9. Gorlach, English in Nineteenth-Century England, 125 ff. ↑

10. Florentina Armaselu et al., LL(O)D and NLP perspectives on semantic change for humanities research (Semantic Web 13(6), 2022). ↑

11. See Barbara McGillivray, How to Use Word Embeddings for Natural Language Processing (SAGE Research Methods: Doing Research Online, 2022), https://doi.org/10.4135/9781529609578; for a non-technical introduction to this technique, and Nina Tahmasebi, Lars Borin, Adam Jatowt, Yang Xu, and Simon Hengchen (eds.), (2021). Computational approaches to semantic change. (Berlin: Language Science Press, 2021) for an overview of recent NLP research on semantic change detection. ↑

12. Discussed further in Ridge, Mia, Nilo Pedrazzini, Miguel Vieira, Arianna Ciula, and Barbara McGillivray. ‘Language of Mechanisation Crowdsourcing Datasets from the Living with Machines Project’. Journal of Open Humanities Data 10, no. 1 (29 April 2024). https://doi.org/10.5334/johd.195. ↑

13. Raphael Samuel, ‘Workshop of the world: steam power and hand technology in mid-Victorian Britain,’ History Workshop Journal, 3 (1977), 6-72; Maxine Berg and Pat Hudson, ‘Rehabilitating the industrial revolution’, Economic History Review, 2nd ser., 45 (1992), 24-50. ↑

14. Gorlach, English in Nineteenth-Century England. ↑

15. See McKernan, Luke. 2019. ‘Heritage Made Digital - the newspapers’ https://blogs.bl.uk/thenewsroom/2019/01/heritage-made-digital-the-newspapers.html. ↑

16. See Giorgio Tolfo, Olivia Vane, Kaspar Beelen, Kasra Hosseini, Jon Lawrence, David Beavan and Katherine McDonough, ‘Hunting for Treasure: Living with Machines and the British Library Newspaper Collection’, in Estelle Bunout, Maud Ehrmann and Frédéric Clavert (eds.), Digitised Newspapers: A New Eldorado for Historians?: Reflections on Tools, Methods and Epistemology, (Berlin, Boston: De Gruyter), 23-46, DOI:10.1515/9783110729214-002. For a breakdown of the newspaper titles included in this corpus, their temporal coverage and the number of tokens, see Ridge, Mia, and Nilo Pedrazzini. 2024. ‘Public Domain Newspaper Titles in Living with Machines’. Living with Machines (blog). 7 May 2024. https://livingwithmachines.ac.uk/public-domain-newspaper-titles-in-living-with-machines/. ↑

17. Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean, ‘Efficient estimation of word representations in vector space’, arXiv (2013), https://doi.org/10.48550/arXiv.1301.3781. We used the implementation of Word2Vec by Gensim (Řehůřek and Sojka 2010). ↑

18. William L. Hamilton, Jure Leskovec, and Dan Jurafsky, ‘Diachronic word embeddings reveal statistical laws of semantic change,’ in Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, vol. 1: Long Papers, (Berlin: Association for Computational Linguistics, 2016), pp. 1489–1501. ↑

19. Because the models are trained separately, they are not initially comparable with one another, i.e. they only make sense ‘internally’. For example, we can compare how similar the vectors for king and queen are in the 1860s, but we cannot compare king in the 1860s and queen in the 1890s, or king in the 1840s and king in the 1870s. The vector assigned to each word at the end of a model training is exclusively computed based on the text from the relevant subcorpus (e.g. 1860s, 1870s, etc.), namely on the words occurring there, their frequency, and their co-occurrence. The alignment of models is typically achieved by applying an algorithm (Orthogonal Procrustes) which transforms the vectors of a model so that they are comparable to those of other models, while also ensuring that the relationships between words within the original model are preserved. For a more technical explanation on the alignment of diachronic word embedding models, see Nilo Pedrazzini and Barbara McGillivray, ‘Machines in the media: semantic change in the lexicon of mechanization in 19th-century British newspapers,’ in Proceedings of the 2nd International Workshop on Natural Language Processing for Digital Humanities (Taipei, Taiwan. Association for Computational Linguistics, 2022), pp. 85–95. ↑

20. We did this with the help of off-the-shelf Python packages based on Levenshtein distance (pyspellchecker). ↑

21. Daniel van Strien, Kaspar Beelen, Mariona Coll Ardanuy, Kasra Hosseini, Barbara McGillivray, and Giovanni Colavizza, ‘Assessing the Impact of OCR Quality on Downstream NLP Tasks,’ in Proceedings of the 12th International Conference on Agents and Artificial Intelligence, Volume 1, ARTIDIGH, (2020), 484-496, DOI: 10.5220/0009169004840496. ↑

22. Pedrazzini and McGillivray, ‘Machines in the media’. ↑

23. Roberta Killick, Paul Fearnhead, and Idris A. Eckley, ‘Optimal detection of changepoints with a linear computational cost,’ Journal of the American Statistical Association, 107, 500 (2012): 1590–1598. ↑

24. H. Sakoe and S. Chiba, ‘Dynamic programming algorithm optimization for spoken word recognition,’ IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 26, no. 1 (1978), pp. 43-49, doi: 10.1109/TASSP.1978.1163055. ↑

25. Kaspar Beelen, Jon Lawrence, Daniel C.S. Wilson, David Beavan, ‘Bias and representativeness in digitized newspaper collections: introducing the environmental scan,’ Digital Scholarship in the Humanities, 38, 1 (2023), 1–22, https://doi.org/10.1093/llc/fqac037. ↑

26. Sarah J. Leslie, Sangeet Khemlan, Sandeep Prasada and Sam Glucksberg. ‘Conceptual and linguistic distinctions between singular and plural generics,’ Proceedings of the 31st Annual Cognitive Science Society, (2009), 479-484; Alda Mari, Claire Beyssade and Fabio Del Prete, Genericity (Oxford: Oxford University Press, 2012). ↑

27. Pedrazzini and McGillivray, ‘Machines in the media’. ↑

28. Zooniverse https://www.zooniverse.org is the world's largest platform for digital volunteering on citizen science and citizen humanities projects, with nearly 3 million registered volunteers. ↑

29. Mia Ridge, ‘From Tagging to Theorizing: Deepening Engagement with Cultural Heritage through Crowdsourcing,’ Curator: The Museum Journal 56, 4 (2013): 435-50. ↑

30. Ridge, Mia. ‘Outreach and Marketing for Crowdsourcing Tasks’. Living with Machines (blog), 27 June 2024. https://livingwithmachines.ac.uk/outreach-and-marketing-for-crowdsourcing-tasks/. ↑

31. Mia Ridge, Nilo Pedrazzini, Miguel Vieira, Arianna Ciula, and Barbara McGillivray, ‘Language of Mechanisation Crowdsourcing Datasets from the Living with Machines Project,’ Journal of Open Humanities Data 10, 1 (2024), https://doi.org/10.5334/johd.195.

    The datasets are: Language of Mechanisation: annotated historical newspaper articles https://doi.org/10.23636/5t9m-0g59, and

    OCR and crowdsourced annotations, Language of Mechanisation, JSON files https://doi.org/10.23636/z634-km37↑

32. Oxford English Dictionary, s.v. “trolley (n.),” March 2024, https://doi.org/10.1093/OED/2298568645. ↑

33. Oxford English Dictionary, s.v. “bicycle (n.),” December 2023, https://doi.org/10.1093/OED/4063130776. ↑

34. Oxford English Dictionary, s.v. “bike (n.2),” March 2024, https://doi.org/10.1093/OED/1131314560, Glen Norcliffe, Ride to Modernity: the Bicycle in Canada, 1869-1900 (Toronto: University of Toronto Press), 47-58. ↑

35. Oxford English Dictionary, s.v. “coach (n. & adv.),” March 2024, https://doi.org/10.1093/OED/2505898911. ↑

36. Oxford English Dictionary, s.v. “car (n.1),” February 2024, https://doi.org/10.1093/OED/4164178241. ↑

37. Oxford English Dictionary, s.v. “coach (n. & adv.),” March 2024, https://doi.org/10.1093/OED/2505898911. ↑

38. Oxford English Dictionary, s.v. “car (n.1),” February 2024, https://doi.org/10.1093/OED/4164178241. ↑

39. Chris A. Williams, ‘Risk on the Roads: Police, Motor Traffic and the Management of Space, c. 1900-1950’, in Governing risks in modern Britain: danger, safety and accidents, c. 1800-2000, ed. by Tom Crook and Mike Esbester (Basingstoke: Palgrave Macmillan, 2016), pp. 195-219. ↑

40. See https://www.designcouncil.org.uk/our-work/skills-learning/the-double-diamond/ ↑