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…


Laisser un commentaire

Votre adresse e-mail ne sera pas publiée. Les champs obligatoires sont indiqués avec *

Ce site utilise Akismet pour réduire les indésirables. En savoir plus sur comment les données de vos commentaires sont utilisées.