The Stevenson Letters

All three units use the same running example: a collection of 462 letters by the Scottish writer Robert Louis Stevenson. This page describes the collection, shows what one of the letters looks like, and explains why it makes a good test case for AI tools.

Robert Louis Stevenson

Robert Louis Stevenson (1850–1894) was born in Edinburgh, and is best known for Treasure Island, Kidnapped, and Strange Case of Dr Jekyll and Mr Hyde. He was often ill, and spent much of his life travelling in search of a healthier climate: to France, to California, where he married Fanny Osbourne, and finally to the South Pacific. He spent his last years in Samoa, at a house called Vailima, and died there in 1894. His family and friends called him Louis, which is how he signs many of his letters.

He was a prolific letter writer, and his letters follow all of this: his work, his health, his travels, his friendships, and the business of writing for magazines.

The Collection

The letters come from The Letters of Robert Louis Stevenson to His Family and Friends, selected and edited by his friend Sidney Colvin and published in 1899 (Stevenson 1899), as transcribed on Wikisource. They are in the public domain. The collection has 462 letters, written between 1863 and 1894, each saved as a plain-text file, with a table recording who each letter was written to and when.

The people he wrote to most often in the collection:

Correspondent Who they were Letters
Sidney Colvin Critic and museum curator, a close friend, and the editor of these letters 54
Charles Baxter A lawyer, and a friend from his student days in Edinburgh 38
W. E. Henley A poet and editor 35
Edmund Gosse A critic and poet 31
Mrs Sitwell (Frances Sitwell) A friend and confidante 29
Mrs Thomas Stevenson His mother 25
E. L. Burlingame The editor of Scribner’s Magazine 23
Henry James The novelist 18

Keep in mind that this is a selection. Colvin chose which letters to publish, and what to leave out. The complete edition of Stevenson’s letters, published by Yale University Press in 1994–95, has nearly 2,800 letters, only 1,100 of which had been published before. Whatever you find in these 462 letters, with or without AI, reflects Colvin’s choices as well as Stevenson’s life: a reminder that every dataset was shaped by someone (Model Training comes back to this).

A Sample Letter

Here is one of the letters, exactly as it appears in its file. It’s the letter you’ll work with throughout Unit 2:

Letter: TO SIDNEY COLVIN

[EDINBURGH, AUGUST 1877.]
MY DEAR COLVIN, - I'm to be whipped away to-morrow to Penzance, where at the post-office a letter will find me glad and grateful. I am well, but somewhat tired out with overwork. I have only been home a fortnight this morning, and I have already written to the tune of forty-five CORNHILL pages and upwards. The most of it was only very laborious re-casting and re-modelling, it is true; but it took it out of me famously, all the same.
TEMPLE BAR appears to like my 'Villon,' so I may count on another market there in the future, I hope. At least, I am going to put it to the proof at once, and send another story, 'The Sire de Maletroit's Mousetrap': a true novel, in the old sense; all unities preserved moreover, if that's anything, and I believe with some little merits; not so CLEVER perhaps as the last, but sounder and more natural.
My 'Villon' is out this month; I should so much like to know what you think of it. Stephen has written to me apropos of 'Idlers,' that something more in that vein would be agreeable to his views. From Stephen I count that a devil of a lot.
I am honestly so tired this morning that I hope you will take this for what it's worth and give me an answer in peace. - Ever yours,
LOUIS STEVENSON.

Even this short letter has most of the features that make the collection hard for software, and for AI models in particular:

  • Capitals. The heading, the salutation (MY DEAR COLVIN), and the signature are in capitals, as they are in the printed edition. So are the place and date (EDINBURGH, AUGUST), the magazines (CORNHILL, TEMPLE BAR), and a word for emphasis (CLEVER), but those were italics in print: the transcription turned them into capitals. Models trained on modern text rely heavily on capital letters to spot names, and are easily misled here (Mayhew et al. 2019).
  • The editor’s additions. The heading (TO SIDNEY COLVIN) and the place and date in square brackets ([EDINBURGH, AUGUST 1877.]) come from the edition, not from Stevenson’s letter.
  • Old spellings and phrasing: to-morrow, re-casting, a devil of a lot.
  • Names that need outside knowledge. Stephen is almost certainly Leslie Stephen, who edited the Cornhill Magazine, but the letter never says so. Villon isn’t the poet François Villon, but Stevenson’s own writing about him, and the letter mentions two different pieces: the story that TEMPLE BAR liked (“A Lodging for the Night”), and an essay in the Cornhill that is out this month. Nothing in the letter tells them apart.
  • The same thing in different forms. Stevenson’s friend is SIDNEY COLVIN in the heading and COLVIN in the salutation. In other letters, he is my dear Colvin, or just S. C..

The Choice of Letters

The letters are real, messy, and unfamiliar to most models, which makes them a fair test. They are also typical of the sources humanities and social science researchers actually work with: historical, edited, inconsistent, and full of references that a reader with some background understands immediately, and a model often doesn’t. What’s easy for you, such as knowing that CORNHILL is a magazine here, and not a street in London, is often exactly what an AI tool gets wrong.

Uses in These Materials

Unit Page How the letters are used
1 Basic Concepts Finding the people mentioned in the letters, with rules or with examples
1 Model Training
Evaluation Metrics
Measuring how well a model finds the names in the letters
1 Language Models How a language model splits names and old spellings into tokens
1 LLM Behaviour A small model invents three different Sidney Colvins
2 All of Unit 2 A six-step exercise with the letter above
2 Activity: Gemini Notebook Exploring a sample of 20 letters with an AI research tool
3 Case Study: NER Finding the people, places, and organizations in all 462 letters with a pre-trained model, and mapping the results
3 Activity: Annotation
Activity: Evaluation
Annotating some of the letters yourself, and measuring the model’s accuracy against your annotations

The Files

The letters, the table of correspondents and dates, the 20-letter sample, and the annotations are all in the repository’s data/ folder. Its README describes every file, where it came from, and its known issues.

References

Mayhew, Stephen, Tatiana Tsygankova, and Dan Roth. 2019. “ner and pos when nothing is capitalized.” In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), edited by Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan. Association for Computational Linguistics.
Stevenson, Robert Louis. 1899. The Letters of Robert Louis Stevenson to His Family and Friends. Edited by Sidney Colvin. 2 vols. Methuen.