Computational Approaches to Digitised Historical Newspapers

The “digital sample” constitutes a subset of the totality or population of newspapers which existed at some point in time. This population is unknown, but can be approximated using contextual resources such as newspaper press directories or library catalogues [2]. Both are simultaneously useful and problematic. Catalogues record mainly what sources are preserved and are close to providing an overview of the complete newspaper record especially for countries with legal deposit. Newspaper press directories are a useful historical source, but also come with their own issues: classification of what “is” a newspaper changes over time and varies by directory. Ultimately, the population or “newspaper landscape” remains unknown, but we can describe those that have been recorded and/or preserved using metadata derived from catalogue or contextual resources [2].

A rich description of the newspaper landscape helps to contextualise and situate the “sample” of digitised newspapers. In other words, it enables us to at least approximate the “representativeness” of the collection, bearing in mind that the latter concept is more complex than simple proportionality and is always defined in relation to the research questions and ethical values or priorities of the researcher. When it comes to the composition of a corpus, we as researchers and content providers will never be able to get rid of biases and unwanted over- and under-representations. Our goal must rather be to identify, understand, acknowledge them, to infer how they may influence the research outcomes and to make them clearly visible within portals.

When using newspapers at scale and-to repeat the metaphor-as a “mirror” of the past, assessing diversity of collections emerges as a critical issue alongside the processes of media production which heavily influences how the presence was reflected. Researchers need to acknowledge whose perspectives or voices are absent in the data and which social categories dominate a collection.

The question of representativeness is closely intertwined with the issue of diversity and inclusion: which (social) perspectives are present in our data, which are missing? Coming back to the metaphor of newspapers as a “mirror” of the past, we can not simply trust the reflection but need to assess how it potentially distorts our image of the past. With rich descriptions of the sample (and population) we can situate data historically and socially. Of course these will always be rough and approximate descriptions, but nonetheless a crucial part of contextualising (the results derived from) big data. Finally, digitisation does not automatically mean accessibility which depends on the institutional policies (e.g. paywalls) and legal restrictions.

Keyword search may satisfy a large group of researchers (and laymen), but others may want to go beyond simply retrieving and reading newspaper content. Interfaces simultaneously provide and restrict access, i.e., the inbuilt functionalities set the limits to how users can navigate and analyse historical materials. They provide the tools and heuristics via which content becomes visible and users can create their research corpus or data set (see below).

However, while such type of access generally works for humanities’ scholars like historians, it does not necessarily meet the needs of those who follow more data-driven approaches, such as computational humanities researchers, computational historians, or NLP researchers. The latter often wish to process larger datasets for automatic enrichment and filtering, among other tasks. While most libraries or platforms provide access via search, accessing “newspaper collections as data” (i.e., at scale) is becoming more prevalent, but contemporary portals remain limited in their support for such interactions with the data.

As newspapers are embedded in specific spatial and temporal contexts, their content also contains traces of historical biases, both in text and image. In the most extreme cases, historical newspapers contain “toxic” content, to use a term common in today’s research on language models and ethical AI. The textual (or visual depiction) of people, especially the more marginalised and underprivileged, articulate attitudes which are considered offensive within contemporary norms.

But not all biased language is “toxic”. A more neutral term would be “overrepresentation” of specific textual patterns among certain subsets of the data, for example conservative newspapers may mention words such as “agriculture” more frequently than newspapers of other political leaning.

Providers of digitised historical newspapers, such as cultural heritage institutions and specific newspaper portals (e.g., ANNO, BelgicaPress, Chronicling America, Delpher, impresso, NewsEye, etc.) can provide detailed information regarding the collection. For example: list of newspaper titles, dates of publication, how much of a newspaper title has been digitised. It could also be useful to provide whatever contextual information about a newspaper and the entire collection is available, e.g., concerning selection criteria, geographical scope, number of editions, print runs, publishers and editors. Information such as political orientation of the newspaper titles, even when imperfect and tied to specific time periods, will be useful here. Ideally, such contextual information is accompanied by sources such as bibliographic references. The question of who is responsible for providing this information was raised, e.g., the cultural heritage institution or the researcher undertaking the research. Perhaps a partnership between these two actors would be most valuable.

Figures 18 illustrates how contextual information could be displayed: firstly, the Newspaper Timelines⁶⁰⁶⁰60https://impresso-project.ch/app/newspapers/ from the impresso project, and secondly, the Press Picker⁶¹⁶¹61https://livingwithmachines.ac.uk/press-picker-visualising-formats-and-title-name-
changes-in-the-british-librarys-newspaper-holdings/ from the Living with Machines project.

The provision of explicit information regarding digitisation quality (e.g., Optical Character Recognition, OCR) would also be useful, ideally provided at a number of levels: for the whole newspaper title, for an issue, or per article.

Digitised newspaper providers should provide explicit guidelines regarding terms of use, particularly in terms of legal consideration. For example, the impresso platform requires users to sign a Non-Disclosure Agreement (NDA)⁶²⁶²62https://impresso-project.ch/assets/documents/impresso_NDA.pdf before access to full collection is granted. Furthermore, it would be useful for cultural heritage providers to provide information about what percentage of the total collection has been digitised. This helps to provide transparency on the “missingness” in a collection, ideally at the level of each of the newspaper titles.

When considering measures to ensure transparency and fairness of research corpus creation, selection and sharing, it is important to consider diversity, equality, equity. To assist both researchers and cultural heritage institutions with this, an equity monitor could be developed. A number of aspects could be considered; for example, the identification of contested terms in a corpus (see [11, 6], as well as [3] and Conconcor⁶³⁶³63https://www.cultural-ai.nl/conconcor). If collection holders are aware that their corpora include contested terms, a disclaimer alerting users to this could be added to the website. This could be particularly relevant when contentious terms are used for query expansion or are used in a visualisation. For researchers whose corpus includes contested terms, it is advisable to explicitly acknowledge this in their publications.

Regarding “low” level processing tasks (OCR/OLR), bias is mainly related to the quality of the tools. Do they work equally well on each part of the collection? How does OCR/OLR quality impact the retrievability and accessibility of each (type of) document? The quality of the tools on each part of the collection will depend on the data that they were originally trained on or developed for. They will likely work best on data that resembles the training/development set.

Similarly, for higher level, automated enrichment tasks (part-of-speech tagging, NER, linking to knowledge graphs, topic detection, sentiment analysis), we can ask: how well do they work for each part of the collection? What data were the tools trained on or developed for, and to what extent is this different from the data currently under investigation? NER tools may show a higher performance on some entities than others. Knowledge graphs may not cover all relevant entities.

In some cases, bias is ingrained in the collection [3]. As a product of their times, historical newspapers will also reflect the norms, values and language of e.g. colonising nations and their perspectives on their colonies. The use of automated enrichment tools may lead to unwanted side effects with respect to colonial or otherwise outdated terminology. Words may be taken out of context. Consider, for example, that a topic detection algorithm may define a geographically-focused topic as a list of terms including racist references to people.

Manual annotation is prone to bias that relates to the viewpoints, background and knowledge of the annotator. In some cases, these highly personal characteristics will be unknown, such as when making use of crowdsourcing. In some cases, we might not want to expose anonymous crowd workers to bias that is ingrained in the collection, such as when offensive, colonial terminology is present.

Search engines typically rank documents based on a combination of the following factors: a matching score between a query and the content of the document, usage data in the form of previous queries and clicks, and an importance score of the document, e.g., using a PageRank-like algorithm. Whether a document will appear high in the ranking will therefore be influenced by its popularity (if usage data is taken into account), connectedness (if PageRank is used), and document properties such as length, which impact the matching score [20]. This may introduce distorted perceptions of search results. Therefore ranking-algorithms should be made transparent to users who rely on them.

Detailed information about OCR and segmentation quality helps a user to decide not only whether the quality is good enough, but also whether bias towards certain parts of the research corpus is to be expected. This requires fine-grained performance metrics, for example at the level or articles, newspaper titles or time periods.

Another solution to mitigate or at least understand bias due to OCR errors is to provide access to the original “raw” scans, i.e., the images of pages.

For automated enrichment tools, documentation of how, for what purpose, and on which training set they were created, helps a user to assess whether bias is to be expected when these tools are applied to their data. Several documentation approaches have been proposed in recent years, the most notable being Datasheets for Datasets [6] for documentation of training sets, and Model Cards [10] for documentation of trained models. Also, the research on provenance is relevant here, i.e. a formal representation of the consecutive processes involved in the creation of the enrichments. PROV⁶⁴⁶⁴64https://www.w3.org/2001/sw/wiki/PROV is an approach to formally capture provenance information on the semantic web.

The content of historic newspaper collections will often be “biased” in the sense that the articles display the perspectives of the time in which they were created. Removing this type of bias will mostly not be feasible or desirable in the context of historical research. We recommend to include an “equity monitor” as part of a research design, where a user critically assesses whether contentious terminology is present in the corpus, and whether this is problematic. As noted above, contentious terminology could be problematic when used as input to automated enrichment tools, or when presented to crowd workers. In these cases, a user could decide to not include certain articles in their research. Note that detection of contentious terminology is not trivial and automation of this task is still in its infancy [3].

A user may include a disclaimer as part of the dataset, to warn (other) users and/or annotators that there may be offensive content. This is especially recommended when sharing the corpus for reproducibility or future research.

We consider human annotators to be always biased. A diverse or representative group of annotators helps to avoid annotations that are skewed towards one background or viewpoint.

Explicit information about who created the annotations helps users to assess whether an (unwanted) bias in the annotations is to be expected. This could consist of age, gender, country of citizenship, (native) language, level of expertise, and way of recruitment of the annotator(s). Note that this information is often not available when using crowdsourcing.

Bias introduced by the ranking algorithm of a search engine may be explicated and mitigated by providing multiple rankings. Many search engines already include additional rankings next to relevance ranking, such as a chronological order. However, specifically the inclusion of multiple relevance rankings would allow a user to understand to what extent the ranking algorithm impacts their goals.

In order to make the research traceable, researchers have to be explicit about the methods and tools they use including, for the latter, the version, the used settings, and why they were appropriate for the task in question.

Often, the analysis of a newspaper corpus involves a set of tools organised in a pipeline. In order to obtain transparency, users should have insight into the composition and performance of the various components of the pipeline. In the case of machine learning tools, it should be clear which versions are used and on which dataset they have been trained.

In order to be fully transparent, ideally all these things are documented, and the tools or queries stored alongside the data. This raises the question what level of documentation is required to make the research traceable or repeatable for others. Some researchers provide tools to document the settings, such as the Gephi Fieldnotes plugin developed by [22]. Others have proposed strategies for tracing all the data handling steps [7]. We see, however, the risk that the effort to produce such documentation may take a disproportionate amount of time and effort. Tool standardisation may make this need less urgent (e.g., the role of SPSS⁶⁵⁶⁵65https://www.ibm.com/de-de/analytics/spss-statistics-software in Social Science research) and therefore reduce this burden.

To facilitate qualitative research, where users explore the corpus at object level, an interface should present research results in the form of scans next to the OCR and metadata (which most portals currently afford). For quantitative analysis, tools will be used both inside and outside the portals. In order to improve the traceability of the research, researchers should have the ability to store tools and their settings alongside the datasets, perhaps on publicly accessible platforms such as GitHub and Zenodo, or using tools specifically designed for this purpose, such as the Gephi Fieldnotes plugin. “How to cite” text blocks with detailed and multimodal information on the tools and their settings could be helpful here.

The transparency of analysis pipelines can be supported by interfaces that allow researchers to compare the performances of different algorithms on a specific task. An example is the NEWSGAC platform⁶⁶⁶⁶66https://github.com/newsgac/platform that allows users to compare different algorithms for automatic genre detection in newspapers. In order to increase the transparency of machine learning tools, references to publications on models used and documentation on training datasets is provided (e.g., in the form of “datasheets for datasets”[6]).

Ideally, users should have the possibility to save, export and reuse their own analysis pipeline and the results. This output should ideally connect to the changing publication and presentation modes for the research results, that allow researchers to include data, code and narratives alongside each other (e.g., the Journal of Digital History⁶⁷⁶⁷67https://journalofdigitalhistory.org/en/about that publishes the data alongside the narrative and a description of the methodological issues, or the ESWC conference⁶⁸⁶⁸68https://2022.eswc-conferences.org/call-for-papers-research-track/ where linking to data, code and other resources is a review requirement); this is further discussed in the Acknowledgements section below.

Visual interfaces are in many contexts suitable, necessary means to make patterns inherent in the data set in question salient to the observer. However, visualisations are abstract representations of (typically) numerical data, and the visual mapping of the underlying information always imposes a level of distortion because numbers are rather easier to compare when they are served in textual form than when our brain has to approximate them when they appear in the form of visuals such as bars in a bar chart or dots in a scatter plot. The complexity of comprehending data in textual form increases with the size of the data set, but visualisations help to arrange the data in a way that users get a quick, understandable overview even for vast data sets. In order to limit the level of distortion, visual representations of data have to be carefully designed.

. Following Edward Tufte’s guidelines for graphical excellence, first of all, visualisations should “show the data”, make it coherent and avoid distorting what the data has to say [19]. An appropriate indicator for good visual design is when viewers are induced to think about the substance rather than about methodology, graphic design or the technology of graphic production. Moreover, visualisations should not “lie,” i.e., the size of effect shown in the visual display needs to correspond to the size of effect in the data.

Of particular importance for visual design is the selection of appropriate colour maps. Qualitative colour maps (a set of different hues) should be used to display categorical data, and continuous colour maps (sequential or diverging) to communicate quantitative data (colour gradients). Although powerful, a general advice is not to encode the most important feature with colour (“get it right in black and white”). Visualisations should also be colorblind-safe, i.e., one should not mix diverse shades of green and red. Several online tools support defining accurate colour maps, e.g. ColorBrewer⁶⁹⁶⁹69https://colorbrewer2.org/.

Next to choosing inappropriate colour maps, visualisation designers should avoid visual clutter that reduces the readability of the displayed data and conceals occurring patterns. Clutter can also occur when choosing 3D over 2D representations, which are the means of choice when visualising data that does not inhere 3D structures. Textures that cause visual stress (moiré vibrations) should furthermore be avoided.

In order to support Information seeking, visualisation tools should implement Shneiderman’s mantra “Overview first, zoom and filter, then details on demand” [17]. Whereas the overview corresponds in digital humanities terminology to distant reading, details on demand refers to close reading. Thus, visual interfaces should support gradual zooming and filtering of the data to be analysed.

Especially, data in the context of humanities applications often embody uncertainty of different kinds (imprecision, inhomogeneity, incompleteness). Visualisations are suitable means to communicate these uncertainties, for example through transparency or grey glyphs, indispensable for increased reliability of visual display of information.

To ensure a transparent visual interface with a minimised level of data distortion, we suggest conducting a participatory visual design process that involves visualisation experts on the one hand, and domain experts that ensure the suitability of the visual design for its intended purpose on the other. An exhaustive overview of visual design principles can be found in [12].

There are some good examples on how interfaces can offer transparency on the collections. The CLARIAH-NL Media Suite⁷⁰⁷⁰70https://mediasuitedata.clariah.nl/dataset/nisv-catalogue , and the Sound and Vision Archive⁷¹⁷¹71https://mediasuitedata.clariah.nl/dataset/audiovisual-collection-daan, offer explanations and graphs to contextualise the audiovisual collection of the Netherlands Institute for Sound and Vision (NISV), explaining both the role of the archive in the archival field in the Netherlands as well as the time the collection spans, what the digital archives cover, what kind of media is included, when updates happen, what part or the collection is digitised, how the collection is enriched, where additional information and help can be found and how it can be searched (as exemplified in Figure 19). It also indicates what the differences between the metadata of two media management systems are and offers links as well as downloads to the description of the metadata fields.

Digital newspaper collections also contextualise their datasets to a certain extent although additional information like the one offered for the Sound and Vision archive might be added. The impresso interface indicates the provenance of its collection in the detail view of individual titles. A simple graph visualises the overall temporal distribution of the newspaper collection and a bar chart enables users to get a quick overview of the time span and amount of data each newspaper contributes to the collection (see Figure 17). Due to the efforts of the National Digital Newspaper Program⁷²⁷²72https://www.loc.gov/ndnp/ which “is a long-term effort to develop an Internet-based, searchable database of U.S. newspapers with descriptive information and select digitization of historic pages” Chronicling America is currently also adding information on its dataset. Figure 20 shows that contextualisation regarding the collection is made in the “Maps and Visualisation” section of the portal and it can be seen that most of this information is of very recent date (mostly updated February and June 2022) and that the “Map and Timeline” of the collection are updated on a weekly basis using the ArcGIS Instant App (see example in Figure 21). The interactive map visualisation has the added value that it supports a scalable reading of the collection, which is often required by humanities researchers.

These examples show that awareness for the necessity of transparency and biases is growing constantly. In this context visualisations can be helpful to support the communication of complex issues at a glance. The visualisation in Figure 16 frames the issue at hand very clearly. It is a good example of how well visualisations are able to communicate difficult and complex information. This is also true for analyses that are being done by researchers (see Section Data analysis).

There is a considerable amount of hidden labour in the process of generating digital newspaper collections. It would be ideal if this hidden labour could be made visible and all the actors in the process could be acknowledged for their work. This would help increase the level of transparency and fairness of the whole digitised historical newspaper ecosystem.

In order for historical newspaper collections to exist, they have to have been acquired by libraries, archives or museums, often in physical form. They then need to be digitised, processed (e.g. metadata generation, creation of lower resolution images for public display), and then enriched (e.g., OCR and article segmentation). It is not feasible to undertake the whole digitisation workflow at once, and therefore selection or prioritisation is needed. All these steps involve the expertise of cultural heritage professionals, computer and data scientists and software engineers, as well as (digital) humanities researchers who would like to use the digitised newspapers as historical sources for their research.

In addition, the creation of research data sets or corpora, which requires sustained intellectual effort, is often not recognised or acknowledged as formal research output. This both discourages researchers to spend time and properly document this crucial step in the research workflow, but also devalues this work as a necessary evil before the “real” research work can begin. Both cultural heritage professionals and computer scientists contribute significantly to this phase, e.g., by providing historical context information about the historical newspaper collections or by working on information extraction methods to computationally facilitate the corpus building process. It is therefore important that this valuable work is uncovered and made visible.

Finally, such work is often made possible by the financial contribution of (public) funding agencies, the acknowledgement of which is also important.

We recommend the formal and visible acknowledgement of all contributing partners at each step of the digitisation process. This includes the contributions by cultural heritage professionals, software engineers, researchers, and (public) funding agencies. For example, when building a platform for the exploration and analysis of digitised historical newspapers, both the funding agency can be acknowledged, such as is the case with the impresso project (“impresso. Media Monitoring of the Past. Supported by the Swiss National Science Foundation under grant CR- SII5 173719, 2019.” ⁷³⁷³73https://impresso-project.ch) or the NewsEye project (“This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 770299.” ⁷⁴⁷⁴74https://www.newseye.eu/about/). Additionally, when building a research dataset or corpus, it is important to acknowledge all contributors, such as “Biltereyst, Daniël, Philippe Meers, Dries Moreels, Julia Noordegraaf and Christophe Verbruggen. Cinema Belgica: Database for Belgian Film History.” ⁷⁵⁷⁵75https://www.cinemabelgica.be, all consulted on July 27, 2022. Not only does this publicly acknowledge the work of people involved in the development of the platform or dataset in question, but it also enables it to be cited in articles or other research outputs that have made use of it. Being able to demonstrate the impact of such platforms becomes increasingly important for their sustainability.

Acknowledgement in publications is another key method to make all parties who contributed to the development of data and the design of platforms visible. For instance, the domains in which each actor has contributed to such a dataset or platform could be stated explicitly, as is already required with regard to authorship by some journals. The Journal of Open Humanities Data Author Guidelines⁷⁶⁷⁶76https://openhumanitiesdata.metajnl.com/about/submissions/ for example provides a number of recommendations based on the ICMJE (Internal Committee of Medical Journal Editors)⁷⁷⁷⁷77https://www.icmje.org/recommendations/browse/roles-and-responsibilities/defining-the-role-of-authors-and-contributors.html, outlining criteria for authorship. An area where there is still room for further development is the (academic) recognition of more innovative digital research outputs. Innovative Journals such as the Journal of Open Humanities Data (JOHD)⁷⁸⁷⁸78https://openhumanitiesdata.metajnl.com, which focuses on the publication of “peer reviewed publications describing humanities data or techniques with high potential for reuse”⁷⁹⁷⁹79https://openhumanitiesdata.metajnl.com/about/; the Journal of Digital History which intends “serve as a forum for critical debate and discussion in the field of digital history by offering an innovative publication platform and promoting a new form of data-driven scholarship and of transmedia storytelling in the historical sciences”⁸⁰⁸⁰80https://www.degruyter.com/journal/key/jdh/html and the Journal of Data Mining & Digital Humanities (JDMDH)⁸¹⁸¹81https://jdmdh.episciences.org which is situated at “the intersection of computing and the disciplines of the humanities, with tools provided by computing such as data visualisation, information retrieval, statistics, text mining by publishing scholarly work beyond the traditional humanities.”⁸²⁸²82https://jdmdh.episciences.org/page/editorial-policies