Technically Speaking, Adapt It is not CARLA
H. Andrew Black
CARLA Consultant, Mexico Branch; International CARLA Communications Coordinator; and International Linguistics Consultant with respect to CARLA

16 October 2001

1 Introduction

Primarily because of a number of recent emails with Bruce Waters, Bill Martin, John Hatton, and Allen Johnson, I've recently come to a better understanding of what Adapt It is. I had mistakenly thought that it was on the lower end of the CARLA scale of tools, much like CC and WrdOChg are (after all, this is where the Appropriate CARLA Tools PowerPoint presentation had it). So I was taken by surprise when some of these folks started making some amazing claims about Adapt It that did not fit with my understanding of what these tools are capable of. That led to a number of discussions.

Thanks to these folks, I now see that Adapt It is actually not even on the CARLA scale of tools, given the standard, technical definition of CARLA (see the introduction to CARLA that's on the CARLA CD-ROM for a fuller explanation). Rather, Adapt It is in the class of tools that are sometimes called "Machine Assisted Human Translation" (or MAHT) tools.

2 Explanation

CARLA tools all seek to model the systematic linguistic changes that have occurred over time. In producing the draft, the computer tool uses the linguistic model the user has encoded to produce the results. The human assists by adjusting the rules and manually editing the output.

Adapt It, on the other hand, does not seek to model any of these systematic linguistic changes (and even the startup splash screen implies this). Instead it provides a very nice user interface for doing manual adaptation with a memory device. That is, the user is doing all of the changes for each word or phrase as they are encountered and the tool prompts the user with what they changed before whenever it sees something they've adapted before. Since Adapt It can also be used to produce things like backtranslations to an obviously unrelated language, it is clearly a different kind of tool than those we've usually associated with CARLA. In Adapt It's case, the human is doing the translation while the computer just makes the human's job easier (and probably more consistent).

Since Adapt It is fundamentally Machine Assisted Human Translation, several things follow:

  1. The user(s) must be adequately bilingual in the source and target languages. Ideally, of the people involved in the adaptation, at least one person knows all the innuendos of meaning, idioms, etc. of the source and at least one person knows the same for the target. One person could conceivably fill both roles.
  2. The users do not need to be linguists or to even be aware of linguistic concepts. Thus non-linguists can successfully use the tool (and this is very encouraging for those situations where there are truly bilingual national colleagues eager to produce translations, but they are not linguists).
  3. The output is indeed potentially better than the pre-revision output of traditional CARLA approaches (I include CC, ShoeBox for Windows, and Ample/SenTrans/Stamp here). The reason is that the human is doing all the adaptation work.
  4. The output of any manual adaptation (whether done 100% by hand or via a MAHT tool like Adapt It) is likely to be expected to be of similar quality to any translation done the traditional way (by this I mean the way SIL has done translations for years). There's no kind of "computer magic" going on, like some people might think happens with CARLA. Therefore, all the traditional quality checks will apply and will apply just as well as they do for any "traditional" translation. Since the output of a MAHT approach will be on a par with a "traditional" translation, it's easy to see why such a tool is easily accepted as a way to produce translations. (Of course, Adapt It is also being accepted because it is so easy to use and one can be productive with it from the very beginning.)
  5. Neither the source nor the target texts have any additional morphological and grammatical checking applied to them (like happens with the Ample/SenTrans/Stamp approach, especially for the source). Thus Adapt It does not offer any morphology-based spelling checking or interlinearization. Since the output is on a par with the traditional method of producing translations, however, this additional checking may not be missed or considered crucial. (Whether this is truly wise is another question.)
  6. Since the user is supplying all the semantics every time they make a change, Adapt It can give the impression that it is more "semantic" than the morphological approaches. Of course, the reality is that the tool itself does absolutely nothing with semantics or linguistics of any kind at all, but since one fully expects the output to faithfully represent the meaning of the source, one can see why it would give this impression.
  7. I now see why Bruce Waters (the author of Adapt It) has no qualms about saying that it will work with all kinds of languages, including highly agglutinative languages. The human does all the real work, so the only possible way it could fail would be if the users did not know what they were doing (which could happen if they were not adequately bilingual in the source and target languages).
  8. Since there is no formal linguistic modeling whatsoever, the tool will never allow the user to capture any kind of linguistic generalizations. This has special implications for any linguistically-aware user working with languages on the agglutinative end of the spectrum. They need to understand up front that the tool will not capture any linguistic generalizations at all. They will have to make potentially hundreds, if not thousands, of changes that could be handled by rules or mappings of abstract morphemes.
  9. Since there are no ordered rules of any kind, the user is never faced with incorrect results due to improper rule orderings. On the other hand, because there is no linguistic modeling, the user will not be learning near as much about the syntax, morphology and phonology of the languages involved as they would if they used some other approach to CARLA.
  10. When is it appropriate to use Adapt It? Whenever manual adaptation would be appropriate. In fact, in my opinion, if you are going to do a manual adaptation, use Adapt It.[1]

Endnotes

[1]

Suppose you have a situation with a cluster of three or more languages where it would be appropriate to do some kind of adaptation between them. Further suppose that the available personnel would allow you to either use the Adapt It approach or, say, the CarlaStudio approach. In the long run, which approach would most likely be the most efficient?

In my opinion, the CarlaStudio approach would be the most efficient in the long run, especially as the number of languages in the cluster increases. This is because once you set up a language pair in CarlaStudio, adding a new language takes considerably less time than the first one or two did. This is simply because they are related and share so much that you've already modeled linguistically. With Adapt It, you would effectively have to start over from scratch for each new pair.

In addition, you would obtain the benefits of adding morphology-based spelling checking and morphological grammar checking for all of the languages involved. You can also produce interlinear text and use this to further check on the quality of the translation.