working out some thoughts here — this will change as I continue working through some of these issues.
If you follow the MarcEdit development, you’ll know that last week, I posted a question in a number of venues about the affects of Unicode Normalization and its potential impacts for our community. I’ve been doing a little bit of work in MarcEdit, having a number of discussions with vendors and folks that work with normalizations regularly – and have started to come up with a plan. But I think there is a teaching opportunity here as well, an opportunity to discuss how we find ourselves having to deal with this particular problem, where the issue is rooted, and the impacts that I see right now in ILS systems and for users of tools like MarcEdit. This isn’t going to be an exhaustive discussion, but hopefully it helps folks understand a little bit more what’s going on, and why this needs to be addressed.
So, let’s start at the beginning. What exactly are Unicode normalizations, and why is this something that we even need to care about….
Unicode Normalizations are, in my opinion, largely an artifact of our (the computing industry’s) transition from a non-Unicode world to Unicode, especially in the way that the extended Latin character sets ended up being supported.
So, let’s talk about character sets and code pages. Character sets define the language that is utilized to represent a specific set of data. Within the operating system and programming languages, these character sets are represented as code pages. For example, Windows provides support for the following code pages: https://msdn.microsoft.com/en-us/library/windows/desktop/dd317756(v=vs.85).aspx. Essentially, code pages are lists of numeric values that tell the computer how to map a representation of a letter to a specific byte. So, let’s use a simple example, “A”. In ASCII and UTF8 (and other) code pages, the A that we read, is actually represented as a byte of data. This byte is 0x41. When the browser (or word processor) sees this value, it checks the value against the defined code page, and then provides the appropriate value from the font being utilized. This is why, in some fonts, some characters will be represented as a “?” or a block. These represent bytes or byte sequences that may (or may not) be defined within the code page, but are not available in the font.
Prior to Unicode implementations, most languages had their own code pages. In Windows, the US. English code page would default to 1252. In Europe, if ISO-8859 was utilized, the code page would default to 28591. In China, the code page could be one of many. Maybe “Big-5”, or code page 950, or what is referred to as Simplified Chinese, or code page 936. The gist here, is that prior to the Unicode standard, languages were represented by different values, and the keyboards, fonts, systems – would take the information about a specific code page and interpret the data so that it could be read. Today, this is why catalogers may still encounter confusion if they get records from Asia, where the vendor or organization makes use of “Big-5” as the encoding. When they open the data in their catalog (or editor), the data will be jumbled. This is because MARC doesn’t include information about the record code page – rather, it defines values as Unicode, or something else. So, it is on catalogers and systems to know the character set being utilized, and utilized tools to convert the byte points from a character encoding that they might not be able to use, to one that is friendly for their systems.
So, let’s get back to this idea of Normalization Forms. My guess is that much of the Normalization mess that we find ourselves in is related to ISO-8859. This code page and standard has been widely utilized in European countries, and provides a standard method of representing extended Latinate characters [those between 129-255], though, Normalizations also affect other languages as well. Essentially, the unicode specification included ISO-8859 to ease the transition, but also provide new, composed code points for many of the characters. And Normalizations were born.
Unicode Normalizations, very basically, define how characters are represented. There are 4 primary normalization forms that I think we need to care about in libraries. These are (https://en.wikipedia.org/wiki/Unicode_equivalence):
- NFC – The canonical normalization, which will replace decomposed characters with composed code points.
- NFD – The canonical normalization, but in which data is fully decomposed
- NFKC – A normalization that utilizes a full compatibility decomposition, followed by the replacement of sequences with their primary composites, if possible.
- NFKD – A normalization that utilizes a full compatibility decomposition.
Practically, what does this mean. Well, it means that a value like: é can be represented in multiple ways. In fact, this is a good example of the problems that having differing Unicode Normalization Forms is having in the library community. In the NFC and NFKC notation, this value é is represented by a single code point, that represents the letter and its diacritic fully. In the NFD and NFKD notations, this character is represented by code points that correspond to the “e” and the diacritic separately. This has definite implications, as composed characters make indexing of data with diacritical marks easier, whereas, decomposed characters must be composed to index correctly.
And how does this affect the library community. Well, we have this made up character encoding known as MARC8 (https://en.wikipedia.org/wiki/MARC-8). MARC8 is a library specific character set (doesn’t have a code page value, so all rendering is done by applications the understand MARC8) that has no equivalent outside of the library. Like many character sets with a need to represent wide-characters (those with diacritics), MARC8 represented characters with diacritics by utilizing decomposed characters (though, this decomposition was MARC8 specific). For librarians, this matters because the U.S. Library of Congress, when providing instructions on support for Unicode in MARC records, provided for the ability to round-trip between MARC8 and UTF8 and back (http://www.loc.gov/marc/specifications/speccharucs.html). This round-tripability comes at a cost, and that cost is that data, to be in sync with the recommendations, should only be provided in the NFKD notation. This has implications however, as current generation operating systems are are generally implemented utilizing NFC as the internal representation for string data, and for programmers, who have to navigate challenges within their languages, as in most cases, language functions that deal with concepts like in-string searching or regular expressions, are done using settings that make them culturally aware (i.e., allows searching across data in different normalizations), but the replacement and manipulation of that data is almost always done using ordinal (binary) matching, which means that data using different normalization forms are not compatible. And quite honestly, this is confusing the hell out of metadata people. Using our “é” character as an example – a user may be able to open a program or work in a programming language, and find this value regardless of the underlying data normalization, but when it comes to making changes – the data will need to match the underlying normalization; otherwise, no changes are actually made. And if you are a user that is just looking at the data on the screen (without the ability to see the underlying binary data or without knowledge of what normalization is being used), you’d rightly start to wonder, why didn’t the changes complete. This is the legacy that round-trip support for MARC-8 has left us within the library community, and the implications of having data moving fluidly between different normalizations is having real consequences today.
Had We Listened to Gandalf
The ability to round-trip data from MARC-8 to UTF8 and back seemed like such a good idea at the time. And the specifications that the U.S. Library of Congress lays out were/are easy enough to understand and implement. But we should have known that it was going to be easy, and that in creating this kind of backward compatibility, we were just looking for trouble down the road.
Probably the first indication that this was going to be problematic was the use of Numeric Character Reference (NRC) form to represent characters that exist outside of the MARC-8 repertoire. Once UTF8 became allowed and a standard for representation of bibliographic data, the frequency in which MARC-8 records began to be littered by NRC representations (i.e., &#xXXXX; notation) increased exponentially, as too did the number of questions on the MarcEdit list for ways to find better substitutions for that data – primarily, because most ILS providers never fully adopted support for NRC encoded data. Looking back now, what is interesting, is that many of the questioned related to the substitution of NRC notations can be traced to the utilization of NFC normalized data and the rise in the presence of “smart” characters generated in our text editing systems. Looking at the MarcEdit archive, I can find multiple entries from users looking to replace NRC data elements that exist, simply because these elements represented the composed data points, and were thus, in compatible with MARC-8. So, we probably should have seen this coming…and quite honestly, should have made a break. Data created in UTF8 will almost always result in some level of data change when being converted back to MARC8…we should probably have just accepted that was a likely outcome, and not worried about the importance of round-tripability.
But….we have, and did, and now we have to find a way to make the data that we have, within the limitations of our systems, work. But what are the limitations or consequences when thinking about the normalization form of data? The data should render the same, right? The data should search the same, right? The data should export the same, right? The answer to those questions, is that this shouldn’t matter, if the local system was standardizing the normalization of data as it is added or exported from the system, but in practice, it appears that few (if any systems) do that, so the normalization form of the data can have significant impacts on what the user sees, can discovery, or can export.
What the user sees
Probably the most perplexing issues that arise related to the normalization form of data, arise in how the data is rendered to the user. While normalization forms have a binary difference, the system should be able to accommodate these differences aren’t visible to the user. Throughout this document, I’ve been using different normalized forms of the letter “é”, but if the browser and the operating system are working like they are suppose to – you – as the reader shouldn’t be aware of these differences. But we know that this isn’t always the case. Here’s one such example:
The top set of data represents the data seen in an ILS prior to export. The bottom shows the data once reimported, but the Normalization form had shifted from NFC to NFKD. The interface being presented to the user has chosen to represent the data as bytes to flag that the data is represented as a decomposed character. But this is jarring to the user as they probably should care.
The above example is actually not as uncommon as you might think. In experimenting with a variety of ILS systems, changes in Normalization form can often have unintended effects for the user…and since it is impossible to know which normalization form is utilized without looking at the data at the binary level – how would one know when changes to records will result in significant changes to the user experience.
The short answer, is you can’t. I started to wonder how OCLC treats Unicode data, and if internally, OCLC normalized the data coming into and out of its system. And the answer is no – as long as the data is valid, the characters, in whatever normalization, is accepted into the system. To test this, I made changes to the following record: http://osu.worldcat.org/title/record-builder-added-this-test-record-on-06262013-130714/oclc/850940559. First, I was interested if any normalization was happening when interacting with OCLC’s Metadata API, and secondly, I was wondering if data brought in with different normalizations would impact searching of the resource. And, the answers to these questions are interesting. First, I wanted to confirm that OCLC accepted data in any normalization provided (as was relayed to me by OCLC), and indeed that is the case. OCLC doesn’t do any normalization, as far as I can tell, of data going into the system. This means that a user could download a master record, and make no other change to the record but updating the normalized form, and replace that record. From the users perspective, the change wouldn’t be noticeable – but at the data level, the changes could be profound. Given the variety of differences in how different ILS system utilize data in the different Unicode normalization forms, this likely explains some of the “diacritic display issue” questions that periodically make their way on the MarcEdit listserv. Users are expecting that their data is compatible with their system because the OCLC data downloaded is in UTF8 and their system supports UTF8. However, unknown to the cataloger, the reliance of data existing in a specific normalized form may cause issues.
The second question I was interested in, as it related to OCLC, was indexing. Would a difference in normalization form cause indexing issues. We know that in some systems, it does. And for many European users, I have long recommended using MarcEdit’s normalization options to ensure that data converted to UTF8 utilizes the NFC normalization – as it enables local systems to index data correctly (i.e., index the letter + diacritic, rather than the letter then diacritic, then other data). I was wondering if OCLC would demonstrate this kind of indexing behavior, but curiously, I found OCLC had trouble indexing any data with diacritical values. Since I’m sure that isn’t the expected result, I’ve reached out to see exactly what is the expectation for the user.
As noted above, for years now, I’ve recommended that users who utilize Koha as their ILS system configure MarcEdit to utilize the NFC normalization as the standard data output when converting data between MARC-8 and UTF-8. The reason for this has been to ensure that data indexes correctly rather than flatly. But maybe this recommendation should have been more broadly. While I didn’t look at every system, one common aspect of many of the systems that I did look at show, is that data normalized as NFKD tends to not index a representation of data as diacritical value. They either normalize all diacritical data way, or they index the data as it appears in the binary – so for example, a record like this: évery would index as e_acute_very, i.e., the indexed value would be a plan “e”, but if the data appeared in NFC notation, the data would be indexed as an é (the combined character) allowing users to search for data using the letter + diacritic. How does your system index its data? It’s a question I’m asking today, and wondering how much of an impact normalization form has without the ILS, as well as outside the ILS (as we reuse data in a variety of contexts). Since each system may make different assumptions and indexing decisions based on UTF8 data presented – its an interesting question to consider.
The best case scenario is that a system would export data the same way that its represented in the system. This is what OCLC does – and while it likely helps to exasperate some of the problems I see upstream with systems that may look for specific normalizations, its regular and expected. Is this behavior the rule? Unfortunately it is not. I see many examples where data is altered on export, and often times, the diacritic related, the issue can be traced to the normalized form of the original data. Again, the system probably should care which form is provided (in a perfect work), but if the system is implementing the MARC specification as written (see LC guidance above), then developing operations around the expectation of NFKD formed would likely led to complications. But again, you’d likely never know until you tried to take the data out of the system.
Thinking about this in MarcEdit
So if you’ve stayed with me this long, you may be wondering if there is anything that we can do about the problems, short of getting everyone to agree that we all normalize our data a certain way (good luck). In MarcEdit, I’ve been looking at addressing this question in order to address the following problems that I get asked about regularly:
- When I try to replace x diacritic, I can find the instances, but when I try to replace, only some (or none) are replaced
- When I import my data back into my system, diacritics are decomposed
- How can I ensure my records can index diacritics correctly
The first two issues are ones that come up periodically, and are especially confusing to users because the differences in data is at a binary level – so hard to see. The last issue, MarcEdit has provided a 1/2 answer for. It has always provided a way to set normalization when converted data to UTF8, but once there, it assumes that the user will provide the data in the form that they require (I’m realizing, this is a bad assumption).
To address this problem, I’m providing a method in MarcEdit that will allow the user to force the normalization of UTF8 data into a specific normalization, and will enable the application to support search and replace of data, regardless of the normalized form of a character that a user might us. This will show up in the MarcEdit preferences. Under the MARCEngine settings, there are options related to data normalization. These show up as:
MarcEdit has included support for sometime to set normalization when compiling data. But this doesn’t solve the problem when trying to edit, search, etc. records in the MarcEditor or within the other areas of the program. So, a new option will be available – Enforce Defined Normalization. This will enable the application to save data in the preferred normalization and also force all user submitted data through a wrapper that will enable edit operations to be completed, regardless of the normalized form a user may use when searching for data or the underlying normalization form of the individual records. Internally, MarcEdit will make this process invisible, but the output created will be records that place all UTF8 characters into the specified normalization. This seems to be a good option, and its very unlikely that tomorrow, the systems that we use will suddenly all start to use UTF8 data the same way – and taking this approach, they don’t have to. MarcEdit will work as a bridge to take data in any UTF8 normalization, and will ensure that the data outputted all meets the criteria specified for the user.
Sounds good – I think so. But it makes me a little nervous as well. Why – because OCLC takes any data provided to it. In theory, a record could switch normalizations multiple times, if users pulled data down, edited them using this option, and uploaded the data back to the database. Does this matter? Will it cause unforeseen issues? I don’t know – I’m asking OCLC. I also worry that allowing users to specify normalization form could have cascading issues when it comes to record sharing. No everyone uses MarcEdit (nor should they) and its hard to know what impact this makes on other coding tools, etc. This is why this function won’t be enabled by default – but will need to be turned on by the user – as I continue to inquire and have conversations about the larger implications of this work. The short answer is that this is a pain point, and a problem that needs to be addressed somehow. I see too many questions and too many records where the normalization form of the data plays a role in providing confusing data to the user, confusing data to the cataloger, or difficulties in reusing or sharing the data with other systems/processes. At the same time, this feels like a band-aid fix until we reach a point in the evolution of our systems and metadata that we can free ourselves from MARC-8, and begin to think only about our data in UTF8.
So what should folks take away from all this? Let’s start with the obvious. Just because your data is in UTF8, doesn’t mean that its the same as my data in UTF8. Normalization forms of data, a tool that was initially used to ease the transition of data from non-Unicode to Unicode data, can have other implications as well. The information that I’ve provided, are just examples of challenges that make there way to me due to my work with MarcEdit. I’m sure other folks have had different experiences…and I’d love to hear these if you want to provide them below.