> -----Original Message-----
> From: Code for Libraries [mailto:[log in to unmask]] On Behalf Of Jakub Skoczen
> Sent: Tuesday, April 03, 2012 1:55 AM
> To: [log in to unmask]
> Subject: Re: [CODE4LIB] presenting merged records?
>
> On Mon, Apr 2, 2012 at 6:38 PM, graham <[log in to unmask]> wrote:
> > Hi Peter
> >
> > I'm actually using pazpar2 from Indexdata to merge my records; this
> > limits you in some ways (in particular it merges on-the-fly, which
> > blocks the use of your option 2 - the initial record on which the
> > merger is based is not guaranteed to be the most fully populated), but
> > has a
>
> Yes -- it's not guaranteed and that's why pazpar2 has the merge option 'longest' which will select the
> richest metadata from all records on a field by field basis. It usually produces better results. Not
> to mention that in a federated search setting the 'initial' record cannot be considered stable -- it
> may and will change from a search to search.
>
These two comments exemplify one of the biggest trade-off areas. (Perceived) Response Time and Quality of Results. We have generally opted to go with the "speed" options as that is what users wanted, with functionality to create better results left to a user initiated post-search option.
Fed search has a reputation for being slow, often because of the slowness of sources to respond, and waiting for all results to arrive before displaying any. No modern fed search systems do this and most adopt some form of "first in, first displayed", with the results list filling out as the slower results arrive. This has problems when processing is applied to the whole record set (such as de-duping or, more interestingly, sorting). Results shuffling around before your eyes is fun, and exactly right for some circumstances (such as Kayak's list of flights ordered by price), but also confusing if results disappear from view while you blink. Usually this is solved with a "update now" type button for users.
So we chose the first record as the "head", and de-duped against it, because it was the simplest (hence fastest) thing to do in the real time melee that occurs when records start pouring in from a fed search. Certainly running the same search against the same sources may produce a variation in the first record of a de-dupe set, but only if the sources have overlapping response times. And it does not seem to be a practical limitation to usefulness - especially if the record is the head of a list of duplicates. They are all there every time.
Merging to get a virtual record provides the best quantity of data, which is generally beneficial. It is not necessarily the best quality, even adopting a merge rule of "longest is best" (which we have as an option), but it is the most fulsome.
Peter
> Since you are using pazpar2 you know that, but I will just mention that especially for bibliographic
> records, it really comes down to looking at the metadata fields and selecting the best-fitting merging
> algorithm: e.g for publication dates you may want to generate date ranges that will be representative
> for the whole merged result or you may want to pick-up all unique author fields from among all records.
>
The general range of algorithms for creating a virtual record from two which are deemed "the same" (a different process to decide that) that we use is based on the characteristics of the data fields. So we create a "merge map" which list the "best" thing to do for each field. And, as in everything, this is a "generally best, and it will make some horrible mistakes from time-to-time" situation. Note that this is to create a synthetic record with an attempt at being the "best", with no intention of allowing drill down to the original source of a data item. If the user wants to do that they open the list and look for the original records. Simplest (A) is to "add field". Just create a combined list of everything. Useful for location, holdings and instance information, even for classifications as long as the type is retained (more than one field/subfield involved).
(B) is to select one values as the "best". This means an on the fly rule (such as "longest " as mentioned above) or a much more complex data mining process, which is very slow - so we don't use it. This is good for titles - but see below as well.
(C) is to merge, but de-dupe at the (sub)field level. This would be used for subject headings, keywords, even authors to reduce the "clutter". This becomes real fun in for example the case of a title, where the 245$a is identical, but the various notes subfields are different, and so end up merged. Often they are contradictory, or orthogonal to each other. Language codes are a good example of chaos.
In fact language codes are a type (D) operation to choose one of the many data element variants (usually the most frequent) and make it the head, with the others relegated in some typographic (or hidden list) way.
There are also specialist treatments such as ranging mentioned above. Useful for dates (how to deal with "c"?) and not much else. Even so dates can have their own idiosyncrasies if dealing with articles, not monographs; how to handle some with dd's and some with mm"s and some with only years, or some combination. A "vote" for the most frequent would seem to give the best chance of a full publication date, but can go seriously wrong - especially when confronted with dd/mm/yy and mm/dd/yy from different aggregators.
Other specialist processing deals with individual data elements such as journal titles and, further afield, phone numbers. Many of these processes can be used for virtual records for display, but generally are reserved for more thorough processing to normalize and sanitize the data in a later stage, which can involve recourse to outside authority tables, and sources, and a lot of other stuff.
And really it is for reasons of doing a lot of work, slowing down general processing and display, and producing sometimes "iffy" virtual records, that we have adopted the simplest display of choosing just one as an exemplar, hiding the rest in a sub-list to keep the display clean, and allowing the user to see for him/herself and make decisions. Having said that, we do display a count, and often a list of the sources of the records as part of the head, and leave it at that. Pragmatism beats perfection.
Peter
> > lot of configuration options, which makes it quite flexible.
> >
> > So, you can choose:
> >
> > 1. Which fields need to be identical for a record to be merged at all.
> > I was using author, title, edition when I first mailed the list, but
> > have found allowing records with different publication dates to be
> > merged just caused too many unpredictable problems, and have now added
> > publication date to the list of required fields. The test for
> > identical authors, dates etc is just a string comparison, so a
> > proportion of records which ought to be merged by these criteria never
> > are, due to typos, variant names, date formats, etc.
> >
> > 2. What to do with fields which differ between records which are being
> > merged. You can choose either 'unique', which appends all unique field
> > values (this is what I use for subject headings, so exactly repeated
> > subject headings are dropped, but variants are kept), and 'longest',
> > which picks the longest field value from all the candidates (this is
> > what I use for abstracts).
> >
> > At the end of the process you have a merged record which has a 'head'
> > with the merged record itself, but which contains each of the original
> > records, so you could potentially do as you suggest and let users see
> > any of the input records if they wanted. However, by default this
> > isn't Marc but an internal format (a processed subset of the Marc
> > input) so it may not be much use to most users.
> >
> > I'm finding the 'head' section is mostly quite usable but does often
> > have individual fields with strange or repeated values (eg values
> > identical apart from punctuation). So I'm doing some post-processing
> > of my own on this, but it's very arbitrary at the moment.
> >
> > Graham
> >
> > On 03/30/12 01:09, Peter Noerr wrote:
> >> Hi Graham,
> >>
> >> What we do in our federated search system, and have been doing for some few years, is basically
> give the "designer" a choice of what options the user gets for "de-duped" records.
> >>
> >> Firstly de-duping can be of a number of levels of sophistication, and a many of them lead to the
> situation you have - records which are "similar" rather than identical. On the web search side of
> things there are a surprising number of real duplicates (well maybe not surprising if you study more
> than one page of web search engine results), and on Twitter the duplicates well outnumber the original
> posts (many thanks 're-tweet').
> >>
> >> Where we get duplicate records the usual options are: 1) keep the first and just drop all the rest.
> 2) keep the largest (assumed to have the most information) and drop the rest. These work well for WSE
> results where they are all almost identical (the differences often are just in the advertising
> attached to the pages and the results), but not for bibliographic records.
> >>
> >> Less draconian is 3) Mark all the duplicates and keep them in the list (so you get 1, 2, 3, 4, 5,
> 5.1, 5.2, 5.3, 6, ...). This groups all the similar records together under the sort key of the first
> one, and does enable the user to easily skip them.
> >>
> >> More user friendly is 4) Mark all duplicates and hide them in a sub-list attached to the "head"
> record. This gets them out of the main display, but allows the user who is interested in that "record"
> to expand the list and see the variants. This could be of use to you.
> >>
> >> After that we planned to do what you are proposing and actually merge record content into a single
> virtual record, and worked on algorithms to do it. But nobody was interested. All our partners (who
> provide systems to lots of libraries, both public, academic, and special) decided that it would
> confuse their users more than it would help. I have my doubts, but they spoke and we put the
> development on ice.
> >>
> >> I'm not sure this will help, but it has stood the test of time, and is well used in its various
> guises. Since no-one else seems interested in this topic, you could email me off list and we could
> discuss what we worked through in the way of algorithms, etc.
> >>
> >> Peter
> >>
> >>
> >>> -----Original Message-----
> >>> From: Code for Libraries [mailto:[log in to unmask]] On Behalf
> >>> Of graham
> >>> Sent: Wednesday, March 28, 2012 8:05 AM
> >>> To: [log in to unmask]
> >>> Subject: Re: [CODE4LIB] presenting merged records?
> >>>
> >>> Hi Michael
> >>>
> >>> On 03/27/12 11:50, Michael Hopwood wrote:
> >>>> Hi Graham, do I know you from RHUL?
> >>>>
> >>> Yes indeed :-)
> >>>
> >>>> My thoughts on "merged records" would be:
> >>>>
> >>>> 1. don't do it - use separate IDs and just present links between
> >>>> related manifestations; thus
> >>> avoiding potential confusions.
> >>>
> >>> In my case, I can't avoid it as it's a specific requirement: I'm
> >>> doing a federated search across a large number of libraries, and if
> >>> closely similar items aren't merged, the results become excessively
> >>> large and repetitive. I'm merging all the similar items, displaying
> >>> a summary of the merged bibliographic data, and providing links to
> >>> each of the libraries with a copy. So it's not really FRBRization
> >>> in the normal sense, I just thought that FRBRization would lead to similar problems, so that there
> might be some well-known discussion of the issues around... The merger of the records does have
> advantages, especially if some libraries have very underpopulated records (eg subject fields).
> >>>
> >>> Cheers
> >>> Graham
> >>>
> >>>>
> >>>> http://www.bic.org.uk/files/pdfs/identification-digibook.pdf
> >>>>
> >>>> possible relationships - see
> >>>> http://www.editeur.org/ONIX/book/codelists/current.html - lists 51
> >>> (manifestation)and 164 (work).
> >>>>
> >>>> 2. c.f. the way Amazon displays rough and ready categories
> >>>> (paperback, hardback, audiobooks, *ahem* ebooks of some sort...)
> >>>>
> >>>> On dissection and reconstitution of records - there is a lot of
> >>>> talk going on about RDFizing MaRC
> >>> records and re-using in various ways, e.g.:
> >>>>
> >>>> http://www.slideshare.net/JenniferBowen/moving-library-metadata-tow
> >>>> ard -linked-data-opportunities-provided-by-the-extensible-catalog
> >>>>
> >>>> Cheers,
> >>>>
> >>>> Michael
> >>>>
> >>>> -----Original Message-----
> >>>> From: Code for Libraries [mailto:[log in to unmask]] On
> >>>> Behalf Of graham
> >>>> Sent: 27 March 2012 11:06
> >>>> To: [log in to unmask]
> >>>> Subject: [CODE4LIB] presenting merged records?
> >>>>
> >>>> Hi
> >>>>
> >>>> There seems to be a general trend to presenting merged records to
> >>>> users, as part of the move towards
> >>> FRBRization. If records need merging this generally means they
> >>> weren't totally identical to start with, so you can end up with conflicting bibliographic data to
> display.
> >>>>
> >>>> Two examples I've come across with this: Summon can merge
> >>>> print/electronic versions of texts, so uses a new 'merged' material
> >>>> type of 'book/ebook' (it doesn't yet seem to have all the other
> >>>> possible permutations, eg book/audiobook). Pazpar2 (which I'm
> >>>> working with at the
> >>>> moment) has a merge option for publication dates which presents dates as a period eg 1997-2002.
> >>>>
> >>>> The problem is not with the underlying data (the original unmerged
> >>>> values can still be there in the
> >>> background) but how to present them to the user in an intuitive way.
> >>> With the date example, presenting dates in this format sometimes
> >>> throws people as it looks too much like the author birth/death dates you might see with a record.
> >>>>
> >>>> I guess people must generally be starting to run into this kind of
> >>>> display problem, so it has maybe
> >>> been discussed to death on ... wherever it is people talk about
> >>> FRBRIzation. Any suggestions? Any mailing lists, blogs etc any can recommend for me to look at?
> >>>>
> >>>> Thanks for any ideas
> >>>> Graham
>
>
>
> --
>
> Cheers,
> Jakub
|