To share the practice from a project I work on - the Jisc Historical Texts platform[1] which provides searching across digitised texts from the 16th to 19th centuries. In this case we had the option to build the search application from scratch, rather than using a product such as ContentDM etc. I should say that all the technical work was done by K-Int [2] and Gooii [3], I was there to advise on metadata and user requirements, and so the following is based on my understanding of how the system works, and any errors are down to me :)

There are currently three major collections within the Historical Texts platform, with different data sources behind each one. In general the data we have for each collection consists of MARC metadata records, full text in XML documents (either from transcription or from OCR processes) and image files of the pages. 

The platform is build using the ElasticSearch [4] (ES) indexing software (as with Solr this is built on top of Lucene).

We structure the data we index in ES in two layers - the ‘publication’ record, which is essentially where all the MARC metadata lives (although not as MARC - we transform this to an internal scheme), and the ‘page’ records - one record per page in the item. The text content lives in the page record, along with links to the image files for the page. The ‘page’ records are all what ES calls ‘child’ records of the relevant publication record. We make this relationship through shared IDs in the MARC records and the XML fulltext documents.

We create a whole range of indexes from this data. Obviously field specific searchs like title or author only search the relevant metadata fields. But we also have a (default) ’search all’ option which searches through all the metadata and fulltext. If the user wants to search the text only, they check an option and we limit the search to only text from records of the ‘page’ type.

The results the user gets initially are always the publication level records - so essentially your results list is a list of books. For each result you can view ‘matches in text’ which shows snippets of where your search term appears in the fulltext. You can then either click to view the whole book, or click the relevant page from the list of snippets. When you view the book, the software retrieves all the ‘page’ records for the book, and from the page records can retrieve the image files. When the user goes to the book viewer, we also carry over the search terms from their search, so they can see the same text snippets of where the terms appear alongside the book viewer - so the user can navigate to the pages which contain the search terms easily.

For more on the ES indexing side of this, Rob Tice from Knowledge Integration did a talk about the use of ES in this context at the London Elasticsearch usergroup [5]. Unfortunately the interface itself requires a login, but if you want to get a feel for how this all works in the UI, there is also a screencast which gives an overview of the UI available [6].

Best wishes,

Owen

1. https://historicaltexts.jisc.ac.uk
2. http://www.k-int.com
3. http://www.gooii.com
4. https://www.elastic.co
5. http://www.k-int.com/Rob-Tice-Elastic-London-complex-modelling-of-rich-text-data-in-Elasticsearch
6. http://historicaltexts.jisc.ac.uk/support

Owen Stephens
Owen Stephens Consulting
Web: http://www.ostephens.com
Email: [log in to unmask]
Telephone: 0121 288 6936

> On 27 Jan 2016, at 00:30, Laura Buchholz <[log in to unmask]> wrote:
> 
> Hi all,
> 
> I'm trying to understand how digital library systems work when there is a
> need to search both metadata and item text content (plain text/full text),
> and when the item is made up of more than one file (so, think a digitized
> multi-page yearbook or newspaper). I'm not looking for answers to a
> specific problem, really, just looking to know what is the current state of
> community practice.
> 
> In our current system (ContentDM), the "full text" of something lives in
> the metadata record, so it is indexed and searched along with the metadata,
> and essentially treated as if it were metadata. (Correct?). This causes
> problems in advanced searching and muddies the relationship between what is
> typically a descriptive metadata record and the file that is associated
> with the record. It doesn't seem like a great model for the average digital
> library. True? I know the answer is "it depends", but humor me... :)
> 
> If it isn't great, and there are better models, what are they? I was taught
> METS in school, and based on that, I'd approach the metadata in a METS or
> METS-like fashion. But I'm unclear on the steps from having a bunch of METS
> records that include descriptive metadata and pointers to text files of the
> OCR (we don't, but if we did...) to indexing and providing results to
> users. I think another way of phrasing this question might be: how is the
> full text of a compound object (in the sense of a digitized yearbook or
> similar) typically indexed?
> 
> The user requirements for this situation are essentially:
> 1. User can search for something and get a list of results. If something
> (let's say a pamphlet) appears in results based on a hit in full text, the
> user selects the pamphlet which opens to the file (or page of the pamphlet)
> that contains the text that was matched. This is pretty normal and does
> work in our current system.
> 2. In an advanced search, a user might search for a name in the "author"
> field and a phrase in the "full text" field, and say they want both
> conditions to be fulfilled. In our current system, this won't provide
> results when it should, because the full text content is in one record and
> the author's name is in another record, so the AND condition can't be met.
> 3. Librarians can link description metadata records (DC in our case) to
> particular files, sometimes one to one, sometimes many to one, sometimes
> one to many.
> 
> If this is too unclear, let me know...
> Thanks!
> 
> -- 
> Laura Buchholz
> Digital Projects Librarian
> Reed College Library
> 503-517-7629
> [log in to unmask]