Who this list is for
This list is for doctoral researchers in digital humanities, literary studies, corpus linguistics and computational text analysis. Large digitised collections such as Eighteenth Century Collections Online (ECCO) have made computational reading of the period possible, but the text behind the page images was produced by OCR and carries errors. Text mining is now a major strand of digital humanities research (Joo et al., 2022), which is why the quality of the underlying text matters.

How to use this list
Treat each row as a starting point. Check what you can access (ECCO is licensed; other collections are open), whether ground-truth transcriptions exist for your sample, and what has been published since the sources below. If you are still choosing a broad area, start with how to choose a PhD topic and how to select a PhD topic.
Theme 1: Measuring OCR quality for the task in hand
Hill and Hengchen (2019) compared ECCO OCR output with hand-keyed versions of the same texts and showed that the effect of OCR errors depends on the analysis: some methods were fairly robust, others were distorted. Liimatta (2024) reached a similar conclusion for register analysis, finding that algorithmic identification of some linguistic features held up as OCR quality fell while others did not. Both point to the same gap: quality should be judged per task, and usually without full ground truth.
| Topic | Possible research question | Likely design and data | Why it matters now |
|---|---|---|---|
| Predicting OCR reliability for specific tasks without ground truth | Can features available without ground truth (dictionary match rates, language-model scores, confidence values) predict how far topic modelling or collocation results will diverge from clean text? | Paired OCR and hand-keyed texts as a training set; regression of task divergence on quality features; test on held-out texts | The impact of OCR errors differs by task (Hill & Hengchen, 2019; Liimatta, 2024) |
| OCR error and literary structure in eighteenth-century fiction | How do OCR errors affect the extraction of characters, dialogue and narrative structure from eighteenth-century novels? | Sample of novels with clean and OCR versions; named-entity and dialogue extraction; comparison of character networks | Existing studies focus on lexical and register features rather than literary structure (Liimatta, 2024) |
Theme 2: Corpus reliability beyond OCR
A corpus can be accurately transcribed and still unrepresentative. Beelen et al. (2023) compared a digitised newspaper collection with contemporary press directories and showed that some kinds of titles were over-represented and others under-represented. For eighteenth-century books, reprints, duplicate editions and metadata errors raise similar problems.
| Topic | Possible research question | Likely design and data | Why it matters now |
|---|---|---|---|
| A composite reliability index for historical text corpora | Can OCR quality, duplication, metadata accuracy and representativeness be combined into an index that predicts how stable a research finding is across corpus versions? | Index construction for one corpus; sensitivity analysis showing how findings change when low-scoring texts are removed or reweighted | Quality factors are usually assessed one at a time; bias in digitised collections is documented (Beelen et al., 2023) |
Theme 3: Post-correction and handwritten text recognition
Transformer models (Vaswani et al., 2017) now underpin both OCR post-correction with large language models and handwritten text recognition. They can reduce error rates, but they can also normalise historical spellings or insert plausible words that are not in the source, which is a serious problem for scholarship that depends on the exact text.
| Topic | Possible research question | Likely design and data | Why it matters now |
|---|---|---|---|
| Auditing LLM-based OCR post-correction | What proportion of LLM corrections in a sample of eighteenth-century texts introduce new errors (modernised spelling, inserted or deleted words), and can an audit protocol detect them? | Corrected and uncorrected OCR compared with hand-keyed text; error typology; precision and recall of automatic detection methods | LLM correction is spreading faster than methods for checking it |
| A cross-collection benchmark for eighteenth-century handwriting | How well do transformer-based HTR models trained on one archive transfer to letters and records from other archives of the same period and language? | Build a benchmark from several collections with transcriptions; train and test across collections; report character and word error rates | Models are often trained and tested within one collection, so transfer is poorly understood |
Keep the handwriting topic to one language and period.
How to test a topic before you commit
Topic choice in doctoral study depends on supervision, funding and access as well as interest (Olalere et al., 2014). For OCR topics, four checks matter most.
- Access: confirm licences for the collection and permission to process its text.
- Ground truth: identify hand-keyed or corrected texts you can use for evaluation.
- Task: name the downstream analysis you will test; quality is task-specific.
- Novelty: search recent digital humanities journals and conference proceedings.
Illustrative example (hypothetical). A candidate starts with "the impact of OCR on eighteenth-century literature". After the checks above, she narrows it to: "How far do character networks extracted from OCR versions of thirty novels published between 1740 and 1780 differ from those extracted from hand-keyed versions, and which quality measures predict the difference?" The question names a task, a sample, a period and a comparison, and it can be answered with texts that already have clean transcriptions.
See PhD topic selection and gap identification. For the digitisation side of archival work, see digitising colonial-era manuscripts.
Common mistakes
- Reporting one character error rate and assuming it predicts every analysis you plan to run.
- Evaluating on the same collection used for training and calling the result general.
- Ignoring duplicates and reprints when counting words or texts.
- Accepting LLM-corrected text without measuring what it changed.
- Planning a multilingual, multi-century benchmark in one doctorate.
- Calling printed books "manuscripts", which confuses OCR and HTR problems.
Key takeaways
- OCR errors affect text-mining tasks unequally, so quality must be judged per task (Hill & Hengchen, 2019; Liimatta, 2024).
- Corpus reliability includes representativeness and duplication, not only character accuracy (Beelen et al., 2023).
- Transformer-based correction and recognition need audit methods as well as accuracy scores (Vaswani et al., 2017).
- Narrow HTR benchmarks to one language and period for a feasible doctorate.
- Check access and ground truth before committing (Olalere et al., 2014).
