Ressources numériques en sciences humaines et sociales OpenEdition Nos plateformes OpenEdition Books OpenEdition Journals Hypothèses Calenda Bibliothèques OpenEdition Freemium Suivez-nous

A tale of precision and recall

Back in the day when “text retrieval” was a thing, I remember learning the difference between precision and recall, and the need for a philosophical attitude to the fact that an optimal search has to maximize both these fairly incompatible factors. I now realise how much this whole ATCL exercise has been about that fact. My earlier efforts to identify ATCL titles in the catalogues of existing digital archives involved comparison on the basis of a manufactured key, algorithmically derived from each resource by the same process, which seemed a good compromise. This method also seemed necessary because of the limited facilities some of those resources offered for querying and manipulating the results of queries. With the availability of the wonderful “opentexts.world” service neither of these constraints applies – but the difficulties of balancing precision and recall have not gone away.

Here are the steps I am jumping through:

1. Generate a list of queries, one for each title in ATCL which doesn’t yet have any digital copy

2. Using CURL, send the queries off to the opentexts.world server and get back an XML representation of the results, including catalogue information and a link to the digital version

3. Process the results to check that this is actually the title we are looking for, and then extract the link to add to my atcl-links database

As of today, my query list has 7791 items. The NLS server doesn’t seem to mind dealing with several thousand CURL requests in rapid succession: it takes about ten minutes to run and dutifully sends me back a fat file containing a fairly straightfoward XML representation of the data.

This is fortunate since I am finding it difficult to decide how exactly to construct my query with maximal precision (to avoid false positives) and maximal recall (to avoid missing any) Most titles contain lots of words, most of which are preserved in most catalogues, so an exact word match for the full title is a good start. There are however still a few problems: punctuation and articles sometimes disappear; some titles appear more than once; some titles are very short, and thus generate many false positives. Quite a few titles have the BTAO problem – that tendency of Victorian publishers to improve the title of a new work by adding to it the formula “By The Author Of [insert previously successful titles by this author]” which results in multiple titles containing the same (irrelevant) string. What’s a good filter to cut down the noise from such things? My first thought was to require that the author’s name should be included; my second was to use the date of publication.

The problem with using the author’s name of course is that that it isn’t necessarily present on the title page, and therefore not necessarily present in the title field of the catalogue record. Many novels are anonymous; many authors published under a pseudonym. The ATCL has done a great job of rounding up and normalising authors, grouping under a single entry all variations of an author’s names. Using this it would be possible to find all the works of “Isabella Harwood” whether published under her name or the more usual pseudonym of “Ross Neil”, by increasing the recall of my “creator” search to allow for either name, but I haven’t yet done that. Instead, for my first experiment, I just use the main ATCL surname of the author, and resign myself to less recall, but more precision.

Running my 7791 queries like this

curl “https://design.opentexts.world/search/export?advanced=true&format=xml&title=Abbot%27s%20Cleve%3A%20or%20Can%20It%20be%20Proved%3F%20A%20Novel&creator=Harwood

gets me a total of 6503 results saying “nothing doing”, and 1288 for which there is one or more matching record. I anticipate multiple hits for each title, since there are multiple editions, and of course most of these catalogues list works by volume rather than by work. A very large number of hits usually indicates a problem: for example, there is a novel with the title “Arthur” by Christiana Jane Douglas. Searching just for “title: arthur AND creator:douglas” gets many many titles containing the word “arthur”, some of them editions of the Morte D’Arthur, edited by James Douglas, and others being numerous editions of Crimean War memoirs by one Douglas Arthur Reid. But 1288 hits is not too big a list to refine further.

My second experiment searches for the full title as above, but filters by date of publication. This produces slightly different numbers: there are now 6143 “nothing doing” responses, and 1648 with at least one hit. More interesting perhaps is that I can now compare the two result sets and see which titles are not found by either query – by hypothesis these are genuinely not available, because they don’t exist in the OpenTexts database – and which are found by one but not the other. There are 659 records not found by the search-with-author queries but found by the search-with-date option, whereas there are only 299 records not picked up by the search-with-date query but found with the search-with-author option. Looking down that list very quickly, I see that in most cases the disparity in dates is because the digitized copy is of a later edition of the same work, and this starts me wondering how much later an edition has to be before I decide it’s not satisfactory. The ideal might be to include only digitizations of the first edition, but an edition produced a year or two later is probably fine. Some of these texts have a long and complicated publishing history in which distinguishing the edition is quite critical; others were reprinted once or twice and then disappeared forever.

I am now leaning to the view that the way forward is to maximize recall, simply by combining the 299 records missed by the search-with-date strategy with the rest, and then to pass those results through another filter to improve its precision. This filter would check, for example, whether the publication details for each candidate match, or are within an acceptable range. But it’s very pleasing to note that I have now identified at least one digital version for 13,769 of the 19,912 titles in ATCL, i.e. 69%. Now, if I could only persuade the British Library to be a bit less secretive…

Counting the books: yes, there’s more

My efforts to find links to digitized versions of all the titles in ATCL made one huge methodogological leap forward last week, and is now poised on the brink of another.

Going through the titles I had managed to extract from a rather uncooperative Google Books interface last week, I noticed that rather a lot of them were marked as “not available” for some reason: more precisely although my 11,104 searches, each corresponding to an entry in ATCL for which I had not yet found a digitized version, had succeeded in identifying 2186 previously unseen titles, they had also thrown up 3885 titles which Google considered inaccessible, presumably for copyright reasons, and 5033 of which it flatly denied any knowledge. Yet when I looked up a few of these same titles (whether allegedly “inaccessible” or “non-existent”) in SOLO – the Bodleian’s wizard student-friendly query interface to its catalogue– there they were, page images downloadable in PDF, no sweat.

Now, amongst other delights, SOLO allows quite rich facetted searching, so it is easy to formulate a query like “find me all titles classed as fiction published in London or Scotland between 1830 and 1900, which have also been digitized by Google”, which made me think for a few moments that my work was now done. But as with many other classy library interfaces, SOLO stops short of allowing a mere automaton to carry out any searching: you have to sit at a keyboard and type, though it will grudgingly allow you to save and download the results of your query … provided it contains no more than 50 (FIFTY!) hits. Which (as I politely pointed out to the harassed librarian on online-chat duty last week), is almost entirely useless for my purposes.

Then I remembered that Real Librarians Do It With Z39.50 and dusted off my YAZ skills. The Bodleian, like all real libraries, has a perfectly good Z39.50 interface, which is not only entirely unbothered by a succession of several hundred queries but also happy to send back directly as many full catalogue entries for the hits as you can (err) handle. The only catch is that the queries have to be expressed in some antique syntax called PQN (Prefix Query Notation) and the results come back in MARC 11. I cut my programming teeth on Fortran IV, so these ancient tongues scare me a lot less than, say, JSON. I turned my list of queries unsatisfied by Google Books into PQN, fired them at library.ox,ac.uk:210/ALEPH and put the kettle on for a nice cup of tea. PQN is not very discriminating, or not in my hands at any rate, and my queries massively overgenerated. But once my 11297 results had passed through through a couple of utilities (yaz-marcdump to produce marcxml, and my very own `marctotei` to identify and fillet the relevant records) I had a set of 780 CPF format records to add to the ATCL database list, and the tea wasn’t even cold (774 once I’d weeded out some duplicates and mismatches).

A natural question is: can we do the same trick with the British Library? Or any other library offering a Z39.50 interface? In principle, yes. But of course the Bodleian’s use of MARC fields may not be entirely the same as everyone else’s, and so the script I wrote to fillet the results of a query may need fine tuning. For example, the BL does not seem to use Marc code 856 (which I rely on) at all: its digital texts are stored in something called the Digital Store, and their identifiers there don’t seem to map directly to anything like a URL. And while I was thinking about that, something unexpected happened.

A tweet arrived, alerting me to the existence online of the “OpenTexts.world” search engine: a search interface to a much more ambitious and much more comprehensive view of the world’s digital resources, namely the Global Digitised Dataset Network (GDD Network), originally a research project into the feasibility of creating a global catalogue of digitised texts. At the end of this project’s first funding year it has made available not only a nice search interface but also (applause) the underlying complete dataset. The latter looks a bit like the HT snapshot dumps I have processed before, though it is missing quite a few useful fields, such as type of text, place of publication, etc. And the nice search interface so far has only limited functionality: nice if you are exploring the data, and really quite annoying if you know exactly what you want to find. On the bright side, it allows you to download the results of the query as a CSV file and even has a sort of API, apparently supporting Lucene-style queries to be passed in via a URL to a SOLR-indexed version of the data. This could well be the answer…

Counting the Books contd.

A couple of days ago I reported here on some imbalance in the representation of male and female novelists in current digital archives, written while I was still trying to persuade the Google Books server to do tricks for me. I can now report further progress. After 21 iterations, I did finally manage to confect a complete list of all the ATCL titles freely available from Google Books. Putting this together with data already stored in ATCL for Google, I have now identified 2823 ATCL titles in Google Books, which brings the total number of known digitizations up to 11,510: 58% of all available titles. This seemed good enough pretext to revisit the summary table I produced last time, so here it is again in a new and hopefully slightly more comprehensible form:

Revised counts

As might have been predicted, with more data the situation becomes more nuanced. Note first that the percentages of available titles which get digitized (column “%dig”) apparently decreases as the actual number of texts available for digitization (column “All”) increases, suggesting that the more titles you have available the less likely you are to deal with any one of them. Only tentative conclusions are warranted, since we are lacking so much data for the later part of the century. That said, comparing the columns M-dig and F-dig suggests that throughout the century, digitizers are consistently and disproportionately more likely to go for a male-authored text. Even in titles from the 1850s, where there are substantially more female authored titles available than male (778 as opposed to 595), the proportion of them which get digitized is still lower than the proportion of male authored titles (79% as opposed to 84%). In the 1880s, 55% of titles are explicitly female-authored, as opposed to 41% male (the remainder being unspecified); yet the male authors are still sampled for digitization at a far higher rate (59% as opposed to 37%).

My previous accusations of sexism amongst the digitizers en masse thus vindicated I next considered the practice of individual archives. The following table shows the numbers of ATCL titles I found in each of five major archives, and the proportions attributed to male and female authors in each.

A-digM-digF-digU-dig%Male%Fem
All115106207505025354%44%
Hathi Trust5655356820226563%36%
InternetArchive16658897482853%45%
Google Books28231138158010540%56%
British Lib51042742225211054%44%
Gutenberg22751682590374%26%
Digitization choices by archive

Overall, the balance is comparable with that shown in the previous table: a small preference for male as opposed to female authored titles (54% to 44%). But this is perhaps concealing a marked variation in practice amongst the archives. At one extreme, Project Gutenberg has nearly three times as many male authored titles as female, while at the other Google Books actually has significantly more female authors than male (56% as opposed to 40%). In between is the British Library Microsoft collection, which matches exactly the proportions for all the archives combined.

Irrespective of gender, how much variation is there in the holdings of these archives? Here’s a frequency distribution showing how many ATCL titles are available from 1, 2, or more archives (note that I cannot distinguish how many of these are actually copies of the same digital version).

1724763%
2278024%
3120110%
42672.3%
5160.7%
Archive overlap: how many titles are available from how many archives?

Encouragingly, this suggests that there is little overlap amongst the holdings of the main digital archives: 63% of all 11,511 digitized titles listed occur in only one archive, 87% in one or two.

Which titles get digitized most frequently? This is hard to tell, for several reasons. Some archives list multi-volume titles as multiple copies; some archives list items simply copied from other archives. For my Google Books listing I excluded titles which were already listed by another archive. But for what it’s worth, here, in no particular order, are the fifteen titles listed as available from all five archives I looked at:

  • Caine, Hall (1853-1931). The Deemster: A Romance (1887)
  • Caine, Hall (1853-1931). A Son of Hagar: A Romance of Our Time (1887)
  • Hamerton, Philip Gilbert (1834-1894). Wenderholme: A Story of Lancashire and YorkshireEdinburgh: Blackwood 1869
  • Collins, Wilkie (1824-1889). Antonina: or, The Fall of Rome. A Romance of the Fifth Century London: Bentley 1850
  • Collins, Wilkie (1824-1889). The Woman in White London: Sampson Low1860
  • Eliot, George (pseud.) (1819-1880). The Mill on the Floss. Edinburgh: Blackwood 1860
  • Dickens, Charles (1812-1870). Oliver Twist: or, The Parish Boy’s Progress. London: Bentley 1838
  • Eliot, George (pseud.) (1819-1880). Middlemarch: A Study of Provincial Life. Edinburgh: Blackwood 1872
  • Gaskell, Elizabeth Cleghorn (1810-1865). Mary Barton: A Tale of Manchester Life. London: Chapman and Hall 1848
  • Grant, James (1822-1887). The Romance of War: or, The Highlanders in Spain London: Henry Colburn 1847
  • Dickens, Charles (1812-1870). Barnaby Rudge: A Tale of the Riots of ‘EightyLondon: Chapman and Hall 1841
  • Dickens, Charles (1812-1870). Bleak HouseLondon: Bradbury and Evans 1853
  • Wood, Mrs. (-). It May be True: A Novel. London: T. C. Newby 1865
  • Oliphant, Margaret (1828-1897). Harry Jocelyn. London: Hurst and Blackett 1881
  • Ouida, (pseud.) (1839-1908). Folle-Farine. London: Chapman and Hall 1871

No, it makes no sense to me either. I expected to see Charles Dickens and George Eliot and Mrs Gaskell on the list, but Hall Caine and Philip Hamerton? Clearly one needs to be very careful in interpreting this data.

See previous bloggage for details of how the numbers were obtained. Supporting data and scripts have been updated in my github repo.

An experiment in counting the books

A couple of years ago I spent some time trying to determine which of the titles in the wonderful “At the Circulating Library” (ATCL) database were freely available online in digital form. This was for largely pragmatic reasons to do with building the ELTeC English language collection: other blog entries describe the method I used and some preliminary results. It’s not as easy as you might suppose to download reliable catalogue information from most digital libraries, nor is it always readily tractable when you do. After some experimentation, I hit on the idea of creating a magic key, a kind of fingerprint, derived from the title and author name as specified, which could then be matched against keys in the same format derived from ATCL entries.

More recently, it occurred to me that this data might also provide some interesting numbers to contribute to current debates about digitization priorities. Exactly why some titles make it to Project Gutenberg, or the HathiTrust, or the Internet Archive and others don’t is not a question to which simple un-nuanced answers are likely or even (maybe) possible, but we should still ask them. Those responsible for the digitization efforts of major libraries are a little coy about the principles on which books are chosen for digitization, or even whether they actually have explicit selection policies, for some reason. I assume that there is a difficult tightrope walk between on the one hand practical but purely adventitious matters such as the relative locations of volume and scanner, the size and state of the volume, the time of day, the temperament of the scanner operator etc.) and on the other principled criteria aiming to ensure a balance of say titles by female and male authors, or high and low brow, date of production, longevity of readership, and so on. It would be surprising if the choices were completely unrelated to characteristics of the population being sampled, or totally failed to reflect the cultural priorities of the scanning operation; the same uncertainties apply, of course, to the collection being sampled for digitization itself.

Anyway, I recently read an interesting article by Allen Riddell and Troy Bassett (“What library digitization leaves out”; preprint available from https://arxiv.org/abs/2009.00513)  which reports that in the data they looked at – the comparatively small sample of surviving English novels published in 1836 and 1838 – shorter books, and books with male authors are disproportionately more likely to be digitized. I naturally wondered whether this applies equally well across the whole of the 19th century.  Which is what led me to revisit my efforts of two years ago. But first, here are the results.

There are 19,912 titles in the current ATCL database. Of these, 9152 (46%) have authors identified in the database as male, 9809 (49%) are identified as female, and 951 (4%) are identified as unknown. These relative proportions are rather different if we look at titles with at least one digital surrogate, of which there are in total 9099 (45%). Of these 9099 digitized texts, we find 5221 (57%) are of male authorship, 3718 (41%) of female authorship, and 160 (2%) are unsexed.

Look at that again. Although there are actually more titles available for digitization from female authors than for male, the number that actually gets digitized is significantly smaller (if, like me, you think a gap of 16 percentage points is pretty significant). Hmmm. These counts of course derive from the whole period covered by ATCL, from 1800 to 1900, so I also calculated them for each decade, only to find that the proportions and their imbalance remain fairly consistent across the century. And this despite huge changes in the numbers: for the last decade of the century ATCL lists nearly 6000 titles, a six-fold increase on (for example) the fourth decade. What percentage of those titles were digitized? In both decades, over 51%. And what proportion of those digitized titles were male-authored ? In both decades, 62%. There is some variability across the decades, but the basic picture remains the same

One possible explanation might be that titles with unknown or unsexable authorship (e.g. the ubiquitous “Anonymous”) are more likely to have been female, and that hence we are not seeing all the truly female authors. But even were this the case (after all, why should we not equally well hypothesize that male authors might be bashful or crave secrecy?), the proportions for books ostensibly male-authored with respect to books ostensibly not male-authored (i.e. those classed as either F or U by ATCL) remain stubbornly higher than the proportions for books definitely not male-authored. And indeed, the same mutatis mutandis is true for the ostensibly-female to ostensibly-not-female ratio.

Here’s a table showing the raw counts:

Decade All “Male” “Female” “U” A-dig M-dig F-dig U-dig
19912 9152 9809 951 9099 5221 3718 160
1830s 482 256 174 52 250 164 85 1
1840s 1037 543 422 72 538 334 202 2
1850s 1483 595 778 110 718 347 358 13
1860s 2341 1019 1093 229 1015 540 456 19
1870s 2866 1189 1514 163 1300 642 633 25
1880s 4126 1693 2287 146 1765 945 782 38
1890s 5979 2995 2863 121 3092 1929 1103 60

 

And here’s another showing the percentages:

               
Decade Ad% M% Md% F% Fd% U% Ud%
45.70% 45.96% 57.38% 49.26% 40.86% 4.78% 1.76%
1830s 51.87% 53.11% 65.60% 36.10% 34.00% 10.79% 0.40%
1840s 51.88% 52.36% 62.08% 40.69% 37.55% 6.94% 0.37%
1850s 48.42% 40.12% 48.33% 52.46% 49.86% 7.42% 1.81%
1860s 43.36% 43.53% 53.20% 46.69% 44.93% 9.78% 1.87%
1870s 45.36% 41.49% 49.38% 52.83% 48.69% 5.69% 1.92%
1880s 42.78% 41.03% 53.54% 55.43% 44.31% 3.54% 2.15%
1890s 51.71% 50.09% 62.39% 47.88% 35.67% 2.02% 1.94%

 

In an ideal world, you’d expect the percentages for titles with male authors (M%)  and for digitized titles with male authors (Md%)  to be roughly the same, right?  Think on… And feel free to download the csv file behind these tables for your own experimentation.

One should always suspect the data, so I make no excuse for the following detailed blow by blow account of how I got these numbers. Full gruesome details, including the scripts mentioned below, are available from https://github.com/lb42/bookLists

The basic method was to download a complete catalogue of relevant titles available from each target digital library, and then try to match them with records in the ATCL. For Google Books, which does not seemingly provide a complete catalogue online, I tried a different method, discussed further below.

I started by downloading the latest (June 2020) dump of the ATCL database, and converting it to a basic TEI XML format. I then did much the same for the holdings of five digital libraries with good holdings of 19th century novels: the Hathi Trust, the British Library, the Internet Archive, Project Gutenberg, and Google Books. As a control, and for testing purposes, I also looked at a few smaller collections, notably the Victorian Women Writers Project at Indiana University and the (now defunct) University of Adelaide “ebooks” repository. I wanted to provide something similar to John Mark Ockerbloom’s lovely Online Books Pages at https://onlinebooks.library.upenn.edu/ but more precisely tied in to ATCL.

Hathi Trust makes available a monthly dump of their entire collection as a huge tab-delimited file. Working with the most recent dump, dated September 1 2020, I used a simple minded perl script `hathiProcess.prl` to parse this file and select from it only freely-available English language books published in Great Britain between 1800 and 1920; an  XSLT stylesheet `htConv.xsl` then converted the results to the common project format (CPF).

The British Library website makes available an Excel spreadsheet providing metadata for the titles from their collection which were digitized some time ago by the Microsoft Books project I downloaded this, converted it to TEI with `csvtotei` and converted the result to CPF, (selecting just the 19th century titles) with `blConv.xsl`.

Project Gutenberg makes available several versions of its catalogue data. I worked with the most recently updated one, which is a vast archive of unbelievably verbose RDF files. Despite its complexity, this data doesn’t include any publication data for the source texts concerned (unsurprising really), though it does provide birth and death dates for the authors. To cut down the numbers a little, I rejected titles whose authors were not born during the 19th century, and also those which specified a MARC relator field “edt” (to cut out non-original editions). Once I had remembered how on earth to handle a gazillion tiny files of RDF (I did this back in 2018 ), I used the `gutConvRDF.xsl` script to process them all to CPF, and concatenated the results into a single file.

The Internet Archive, so far as I can see, doesn’t have any generally available or downloadable catalogue, though it does have a really good query interface. The method I used for attacking Google Books would presumably work equally well (or equally badly; see below) in this case, but I haven’t tried it. Instead I just used a predefined collection called `19thcennov` which someone at UIC Urbana Champaigne thoughtfully created back in December 2008. This gave me 7828 XML records which were easily converted to CPF using `iaConv.xsl`.

The common project format files all consist of TEI <bibl> elements with either an @xml:id attribute or an <idno> specifying the identifying code for this item in the relevant repository, e.g. ‘ia:foreignersnovel03pric` identifies the Internet Archive’s digitization of volume three of Eleanor Price’s novel “The Foreigners” . Each <bibl> also has an @n attribute supplying the magic key for the title, which is confected as follows:

  • remove the full stop following Mr or Mrs in any title containing one
  • take the substring of the title up to the first occurrence of one of the punctuation marks . , : ; or /
  • concatenate this with the author’s last name
  • convert to lower case and remove all punctuation characters and spaces

So, for example ATCL lists a work with the title “The Foreigners: A Novel” attributed to author “Eleanor C. Price”. The same work appears in the Internet Archive list, but with the author “Price, Eleanor C. (Eleanor Catherine)” and the title “The foreigners : a novel”. Despite the differing strings, both will get the same magic key “theforeigners|price”. This method is far from bullet proof, but it’s serviceable.

For Google Books, as noted above, there is no readily downloadable catalogue. But there is an API, which in a moment of madness I thought it might be cool to learn how to use. A day of poking around led me to a neat python script some helpful person had written to look up ISBN numbers (hat tip to AO8’s treasury , which I mercilessly hacked to my own purposes. My version reads a file of URL-encoded search requests like this “inTitle:the+inTitle:foreigners+inAuthor:Price”, fires them at the Google API, and processes the returns into a rudimentary bibl or a comment lamenting the absence or unavailability of the item in question. The file of search requests is rather long (one for each title in ATCL for which I have not yet found any digital version – a total of 11,203 ) so I make the program sleep for a while after firing off about 40 consecutive requests, to help the Google server catch up. Despite this considerate behaviour on my part, it did not take Mr Google long to decide that my program (or my IP address) was a threat, and then to start returning unco-operative HTTP messages like 503 (“Service Unavailable”) and 429 (“Too many requests”). The API Help pages confirm that Google considers “using an app, program or script to perform a large number of searches in a short time” prima facie justification for temporarily blocking the IP address in question; though it’s not clear what exactly is meant by “large” (more than 100?) and “short” (less than a minute?) in that phrase. Furthermore, when I search using my specially-minted API key, there seems to be a hard limit of 1000 queries per day in any case: so this job is not going to be finished very quickly. Still, I do now have an extra 1517 records to show for two day’s work.

Once I’ve created all these lists, I run the merger.xsl script to add <ref> elements to the ATCL-TEI file I created in the first step. This makes for some redundancy, for two reasons: firstly, for most of the archives a three volume novel is likely to get a separate entry for each volume; secondly, for many titles, there exist multiple digitizations – which may (or may not) derive from the same source. The following table shows for each archive the number of records selected for processing, the number of references to ATCL titles found, and the number of titles affected. Note that I haven’t yet done any de-duplication to remove overlaps.

British Library 62015 9920 5104  
Hathi Trust 460070 18891 5655  
Internet Archive 7829 4691 1655  
Project Gutenberg 38338 2880 2275  
Google Books ? 1517 1517  

I haven’t made available the CPF files for each archive, nor the final merged TEI version of the ATCL dump, since this is not really my data to share. But I have made available a file called atcl-links.csv, which is a spreadsheet with a row for each ATCL title digitized in one or more publicly available digital collection, mapping its ATCL identifier to its identifier in each repo. I’ll  update these as and when the data improves.

Can we trust the ATCL database?

I’ve been enthusiastic about the database behind the At the Circulating Library website ever since I discovered it nigh on three years ago. Troy Bassett, its creator, deserves much respect and lots of credit both for the work put into creating it and maintaining it, and for his generous open minded policy of making the data itself freely available in snapshot form for nerds like me to play with. I think of the ATCL as the nearest thing we are ever likely to get to an authoritative catalogue of the 19th century novel, and use it as such in other work which I’ll report on later. But (you guessed there was going to be a but) just how reliable is its coverage? Does it cover each decade of the 19th century equally well, or are there gaps?

When asked, Troy assures me that he’s fully aware that there are gaps. He estimates that the true size of the database should be well over 25,000 titles, rather than its current 18,000. He and his team are slowly and meticulously filling the gaps, by hoovering up and checking data from resources such as published catalogues and bibliographies, or online collections like Proquest’s NSTC. It’s tedious and fiddly work, ideally suited to a small farm of graduate students, should you have such a resource at your disposal.

My interest in ATCL is to use it as a reference point when assessing the coverage of other resources, in particular the catalogues of digital libraries. Take, for example, the current Hathi Trust catalogue. This lists 17,433,331 titled volumes in all; 235,035 of them being volumes published in the UK between 1800 and 1900. Removing obvious duplicates (Hathi catalogues each volume of a three-decker separately, usually though not always) brings this down to an only slightly more manageable 129,817 titles. How good is its coverage of 19th century novels? The difficulty of course is to winnow out just the entries which are novels. As I suggested on my blog back in June 2018, the word “novel” in a title turns out to be a very good indicator (other good words include “or” and “tale”, “history” etc), so I extracted from the HT records just those titles containing the word “novel” for further investigation: there are 953 of them.

In my earlier work, I’d just assumed that if a title in HT didn’t appear in ATCL then it wasn’t a novel. But how complete is ATCL? A good way of finding out might be to work the other way round: first get a list of things which are definitely (probably) novels — and then check to see whether  they also appear in ATCL.

On my first pass, 284 of my 953 “novels” (i.e. titles containing that word) did not appear in ATCL, which surprised me. But this was mostly a result of my matching procedure being insufficiently robust to cope with the amount of variation in cataloguing practice within the HT records and their occasional divergence from ATCL practice. I spent a happy day or two going through the list of delinquents by hand, taking the opportunity to discard 19th c. reprints of earlier works, translations from other languages (mostly scandalous Zola), and a handful which were clearly not novels at all (e.g. “Photographic amusements : including a description of a number of novel effects obtainable with the camera / by Walter E. Woodbury.”) – all of which brought the number down to 191. I then fixed the “magic keys” used to identify each title in ATCL as necessary, all of which brought me to a manageable 40 titles apparently missing from ATCL, with which to torment Troy Bassett and his team. 40 titles missing from a sample of 953 is pretty good; even 191 missing is not so bad in this always very approximate work. I remain persuaded that the ATCL is a reliable approximation to a representative sampling of the 19th century novel.