I took Monday and Friday off this week and on Wednesday and Thursday I attended the Digital Humanities Congress in Sheffield, meaning I only had one day to spend on regular work this week.  I spent most of this writing a specification document for the database and CMS for the Scottish Life Writing project, which I sent off to PI Craig and RA Emily for feedback.  I also had a further chat about the new search and results filtering options for the DSL.

I also spent a bit of time fixing a few bugs that I’d spotted in my recent updates to the HiMuJe Malabar map.  Category labels in the legend were getting duplicated when switching between menus in the map (e.g. ‘Area’ would become ‘AreaArea’), and I managed to fix this.  I’d also spotted that the categorisation type was getting removed from the URL when switching between menus, which meant that bookmarking or citing the page didn’t give the correct categorisation.  I managed to fix this too.  Finally, I’d spotted that loading a travel route stage from a URL into a fresh browser or browser tab meant that the map didn’t zoom to the stage’s location but instead displayed the full extent of the route.  This was because the code to position the map to the full extent was still running when the code to zoom to a specific location was triggered, meaning the latter did not take effect.  I fixed this by adding in a delay.  There is now a brief pause between the route loading in and then the map zooming to the chosen stage, which I think works quite well.

It was great to attend the conference in Sheffield and I attended a lot of interesting sessions (see https://www.dhi.ac.uk/dhc/2026/), a lot of which involved discussions of AI, which is having as huge impact on digital humanities research at the moment.  The opening plenary talk discussed ‘Literary Criticism after Large Language Models’.  The speaker noted that LLMs can produce a close reading of texts in seconds, but discussed whether this can actually ever be considered literary criticism or is it just imitation, and is there actually a difference?  Can it be considered criticism if there’s no understanding behind it?

The speaker gave an interesting overview of the stages of the use of computers in the humanities, starting with concordances and indexes in the 50s to the 90s, the emergence of distant reading in the 2000s, the shift to algorithmic criticism and interpretative visualisation in the 2010s and in the 2020s a major shift to using LLMs to interpret text – no longer just counting words and visualising but interpreting texts too.

The speaker pointed out the criticism involves interpretation (extracting meaning), surprise (new insights), context (historical and cultural), revision (deciding what is relevant) and reproducibility.  He discussed a study that compared how GPT-4 and humans can interpret metaphors, and how the AI was pretty effective, but whether this was from generalising or interpreting from existing knowledge.  Another study looked at authorship and genre in short passages and LLMs performed very well, with Llama 38b proving to be the best in this study.  The speaker then discussed a test he carried out on a poetry magazine corpus that he created and asked GPT-5 to analyse.  This was a new corpus so the LLM had no prior knowledge and consisted of around 1300 poems.  So this was ‘blind corpus analysis’ with no existing or contextual knowledge.  He asked the LLM to identify formal, stylistic, thematic, rhetorical and linguistic patterns to formulate a critical account of the collection.  The results were good, with genuine insights and the results could plausibly appear in literary scholarship.  He pointed out that LLMs can perform corpus level criticism but tend towards overcoherence, but can revise when challenged.  He also pointed out that everything the LLM does is heavily prompt dependent – the AI only ever does what is specifically asked – a human critic is still absolutely needed.

The speaker concluded by stating that criticism is a text about a text – it is an accountable act of reading and LLMs can do this.  He stated that there is ‘genuine’ (human) versus ‘false’ (AI) interpretation, but that the dividing line between the two is very thing now and wondered what forms of criticism will become more viable when plausible readings are cheap.

For the first parallel session I chose session 1, which involved three talks about AI.  The first looked at creating handwritten text recognition (HTR) for complex manuscripts using generative AI in the context of the Newton Project (https://newtonproject.ox.ac.uk/).  This has been around since 1998 and currently incudes around 12 million words of transcribed text.  They have a new project launching in 2027 which will digitise three books such as the Philosophae Naturalis Principia Mathematica.  These works have a lot of mathematics in them and are not uniform documents, including lots of hands, tables, diagrams, revisions, calculations and both Latin and English text.  The manuscripts are already digitised and are available in IIIF hosted at Cambridge Digital Library.  The speaker sampled 5% of the manuscript (around 80 pages) to evaluate the document using a 5-tier system.  19% were evaluated as being extremely difficult to transcribe with only 6% being easy.

The speaker had experimented with Transkribus and wanted to compare this to more general AI tools (Leo AI and Gemini).  She picked out four pages at different tiers to test, performing a line-by-line comparison of the tools.  For a tier-2 page (pretty easy) Leo performed best, followed by Gemini then Transkribus.  For a tier-4 page (difficult) Transkribus failed, Leo quit mid-way through and Gemini started to give summaries.  The speaker updated the prompt, which improved things but the results were still not highly accurate.  The speaker noted that as Transkribus works line by line it does not work so well whereas the AI tools take the whole page into consideration.

The speaker focussed on Gemini and created a detailed prompt with lots of instructions and developed a tool using Claude Sonnet that can be used to evaluated and correct HTR output (see https://johedesan.github.io/newt/).

The second speaker discussed marking up historical texts using AI.  He used data from the Proceedings of the Old Bailey project (https://www.oldbaileyonline.org/).  This was manually annotated using XML tags to note who did what, sentencing and other things.  The speaker wanted to investigate how AI could do this now, given prompts and examples.  He used the a locally installed instance of Qwen AI to perform the tests.  It did well, but there were some issues.  The output specified was not a conversational response, so there weren’t any hallucinations and he set it up with a ‘head’ using the spaCy NLP tool.  This was able to pick out defendants, victims and verdicts from trial documents.  He stated that the tools found assigning the correct verdict to be hard as this often appears at the end of the text and is tricky when there are multiple people.  Long trials are also tricky.

The third speaker discussed person identification in historical correspondence using an LLM.  The data was the Hammer-Purgstall correspondence (https://gams.uni-graz.at/hpe) which features 3545 transcribed letters.  This started with a print edition that was converted to TEI, although the semantic annotation is incomplete.  The project used the CorrespSearch webservice (https://correspsearch.net).

The project had to deal with many issues such as deciding on the boundaries for annotation (what should be included in the annotation) , dealing with tables, working out who pronouns refer to, and issues with families and groups.  Also people with mononyms, issues with different people with the same name (e.g. fathers and sons).  The project used an LLM as it could make use of context and works with multiple languages.  As the project didn’t want the LLM to mess with the XML the output was not updated XML but a series of positions where the relevant tokens start and end.

The results of the tests were compared to human annotation.  AI tests were carried out using GPT-5, Claude 5 and Qwen 3.5 and the tools were able to extract 68% of the names.  It took humans 13 hours to process the data but the AI tools only took 20-35 minutes.

For the next parallel session I chose Session 5, which was mostly focussed on GIS and mapping.  The first speaker discussed visualising the historical evolution of urban form using GIS.  The speaker used ArcGIS to analyse maps of colonial Calcutta over 100 years.  She looked at street sinuosity and street width to compare colonial centres to indigenous areas , and also ‘non-plot’ areas (areas in transition without formal buildings).  She created colour coded clusters based on these criteria (e.g. wide, straight, very low non-plot) to show the character of districts, comparing this to historical data and also showing how the situation changed through various different maps over 100 years.

The second speaker looked at land reclamation in 19^th century Denmark based on historical documents that were available as PDFs.  The speaker used an LLM as it could work well with Danish.  This was able to extract coordinates and create georeferenced data that could then be searched.

The third speaker discussed geoparsing data found in archaeological records such as field reports – ‘grey literature’ that is otherwise difficult to extract data from.  The project developed a tool to work out which places are mentioned in the texts.  This is a two-part process involving recognition (finding places in the texts) and resolution (working out what the place actually is).  The latter needs the most work.  NER tools can struggle, e.g. when multiple places share the same name such as Memphis in the US and Memphis in Egypt.  The speaker discussed how the project’s tool creates a ‘bounding box’ to work out which places are likely correct based on the proximity of other named places in the text.  Further information can be found here: https://cloud.gate.ac.uk/.

Day 2 of the conference began with more parallel sessions, and I chose to attend Session 7.  The first speaker talked about the digitisation and study of marginalia.  He pointed out the marginalia were once frowned upon and attempts were sometimes made to remove them, such as through chemical erasure, but they are now of great interest.  He presented the work of a project called Magic (https://www.magic.unina.it/en) that is attempting to read these marginalia using modern technology.  The project uses FITS images that have embedded metadata and runs processes to reduce things like bleed-through to enhance the visibility of the notes – processes that also help with HTR.

The second speaker discussed the Law Versus Practice project (https://www.lawvpractice.ie/), which extracts data about women and property ownership in early modern Ireland.  It is creating biographical profiles of women from many primary sources, such as the list of claims, which contains 3,100 claims.  The data is fairly standardised, tabular data that is mostly printed, although there are inconsistencies, omissions and errors.  The speaker pointed out that disambiguation was an issue and needed research using titles, parentage, dates and geography.  This was tricky for women due to marriage and name changes.  The project has extracted over 5000 unique people that are linked to places, although the latter is often incomplete.

The third speaker discussed slavery in the Dutch East India Archive (see https://voices.iisg.nl/).  The project alanysed 7.5 million pages of transcriptions from the archive, mainly focussing on early modern Asia and mapped slave voyages.  HTR was carried out by a tool called Loghi (https://github.com/knaw-huc/loghi-htr) with linguistic analysis using AntConc.  The speaker noted that the transcriptions were not perfect but were good enough for searching and finding the co-occurrence of slave-related terms using AntConc.  Around 800 terms were analysed, including spelling variants.

The next session I attended as Session 10, and the first speaker was Jeremy Smith, who gave a fascinating talk about religious vocabulary in the 17^th century, using the ‘quads’ concept from the Linguistic DNA project and data from EEBO-TCP.  He talked specifically about evidence from the (mostly religious) works of Lucy Hutchinson.  He discussed how the phrase ‘infinite love’ was not scriptural but became commonplace during this period.  He also discussed ‘saving faith’, a term which also appeared during this period.

The second speaker discussed how ideas travel in published works, mapping the influence of John Locke in the 18^th century through direct quotation, paraphrase, concepts, critique, adaptation and appropriation.  The vocabulary can change so the speaker investigated connections using semantic similarity across 32 million pages in ECCO using 100-token chunks to investigate twenty high-use Locke quotes.  The third speaker discussed identifying essay-length republication in books and newspapers in the 18^th century, looking at the work of David Hume.

For the final parallel session I attended Session 13.  The first speaker discussed news consumption, social media and youth engagement in Portugal.  The speaker pointed out how social media is now the gateway for news, either intentionally or incidentally, and how young people are more disengaged with other media.  There is no stable relationship with a specific new brand and engagement is formed through algorithmically curated feeds that shape an image of the world, including exposure to misinformation and propaganda.  The speaker looked at people aged 15-29 in Portugal using focus groups and questionnaires.  The data consisted of around 1500 responses, with more female than male participants, with most from the older age groups having higher education.  53% used social media for news daily, with TV news at 25% and newspapers 17%.  The respondents viewed traditional media as being more trustworthy, but social media also includes traditional news accounts as well as influencers and amateur news accounts.

The second speaker discussed propaganda and linguistic divergence in Russian online discourse during the war in Ukraine.  She discussed linguistic variation across state and social media, with social media more focussed on mobilisation and disinformation while state media focussed on pacifying the population and normalisation.

The third speaker gave an overview of preparations for a study of zeitgeist in UK parliamentary debates, looking at the period from Callaghan, Thatcher and Major (1976-1997).  The project will use the Hansard corpus.

The final session of the conference was a roundtable discussion about Digital Humanities in a changing higher education landscape and discussions covered the position of DH at a time when funding is increasingly hard to come by, and the role of AI in DH research.

The conference was hugely interesting and featured some great speakers, lots of interesting research and some interesting challenges.  It was really great to be able to attend.