How a learner of a polysynthetic language like Arapaho would use an electronic or printed dictionary?

Question

Along the lines of this question about agglutinative dictionaries, I just finished reading through the paper "Creating lexical resources for polysynthetic languages — the case of Arapaho, by Kazeminejad et al." from the University of Colorado.

In that paper they talk somewhat deeply (albeit only briefly, with minimal examples) about how they built a Finite State Transducer (FST) to what sounds like generate all possible combinations of how you might use a verb base/stem in Arapaho. They said there could easily be up to 100 variations of how you might piece together the prefixes and suffixes around a verb base/stem, in addition to the complicated morphophonological changes that sometimes occur across short or longer distances within words.

But basically, they built some sort of generator that can take a verb base (or stem is it called?) and generate all the possibilities. I guess it also served as a parser too, and could parse an incoming "complete sentence-word" (verb stem + prefixes and suffixes) into the several parts of speech and whatnot. This seems extremely difficult to pull off (they just casually gloss over it like it was nothing haha), even for a skilled programmer. It seems you have to completely master the grammar as well before you can then begin to start creating this parser/generator, and those grammars are often hefty tomes.

My question though is, what goes into the print dictionary, and what goes into this "electronic" dictionary? (which they didn't really describe from the customer perspective very much) Say you built this parser/generator after mastering the Arapaho grammar (or some other polysynthetic language grammar), what then? How does a "user" (customer) use this parser / electronic dictionary? Is it that they would type in a "full sentence-word" (as I'm calling it, with verb stem and several prefixes/suffixes), and it would figure out the verb stem/base, then generate a bunch of derived variations from it on the fly? What is the user experience you would need for such a tool? And what "data" would back the dictionary (i.e. what stuff would they hardcode into the dictionary, like verb stems)?

Finally, they only briefly mention what they might do if they created a print dictionary. I think they just wanted to list the stem, even though you might never use the stem directly in speech, and it requires "advanced skill" to be able to figure out what the verb stem/base is. Can you clarify this part on how a "static" (or "print") dictionary would work in their paradigm for Arapaho (or a polysynthetic language)?

To summarize, I would like to know how a polysynthetic dictionary should work, both dynamically (electronically) and statically (printed), according to what they are going for in the paper. I am not asking "what is the best way to create a polysynthetic dictionary", but instead I just want to get a better sense what these authors are hinting at (they didn't go into enough detail, spent too much time talking about the details of the parser/generator). What does it look like for a learner of Arapaho, how they would use these two tools? Can you provide a richer example?

Next up I am going to take a look at Modeling the Noun Morphology of Plains Cree by Snoek et al., which was linked to by the other paper and which seems like it might offer a little more insight.

Answer 1

You can ask that question with a number of different auxiliaries – would, should, could, did, does... each generating a different answer. My answer is not about the Arapaho project (which seems to be 404), but the outline of my answer may be applicable to Arapaho. This is based on Saami and the product of Giellatekno.

North Saami is taught as a language course at the University of Tromsø, so that is one source of actual experience in language teaching issues. One can also inquire at the Giellagas Institute in Finland. I took intro Saami at Uit, at a time when the computational product was not quite mature (the teaching product Oahpa was just an idea), so the product was not used in instruction at the time. My first recommendation would be to inquire of the teaching staff in universities, especially Uit, to see how the current computational tools from Giellatekno are actually used in language instruction. They developed analogous tools for other Saami languages as well as other languages.

There exists a (modern) print Sámi-Norwegian and Norwegian-Sámi dictionary. However, at least at the elementary level of language-learning, I don't think it is generally used by elementary learners, instead typical learners use pedagogically fine-tuned materials from class (local supplemental materials plus the Davvin textbooks). For non-typical or advanced learners, what you can get from the print dictionary is a given North Saami word in typical citation form, POS coding, a somewhat cryptic system of paradigm-indication, and a Norwegian gloss. You have to know the morphology of the language to figure this work out.

Compare that to the various GT tools. You can plug in a nominative singular like bárdi and get a medium-sized paradigm or a full paradigm (which includes all of the possessive suffixes). IMO this is a stunningly useful language-learning tool, for a language with a massively complex inflectional system. You can go to class and as the instructor what the illative singular of bárdni is, or you can ask the computer and get immediate satisfaction. (I am willing to say that the responses it gives are correct, having tried to trip it up). This does not tell you what a bárdni is or what the syntactic function of the illative singular is, nor how to pronounce the word. You can elsewhere get a Norwegian gloss by entering the word in a text box (bárdni translates as gutt, and now you are in the simple area of Norwegian-to-English Google translation). You can plug in any text: Fonologalaš transkripšuvnnas fas dát erohus ii merkejuvvo, dasgo dat ii leat sámegiela sániid earuheami dáfus deaŧalaš iige dárbbašlaš translates as På en fonologisk transkripsjon igjen denne forskjellen merkes ikke, fordi den ikke er samiske ords skillet med hensyn til viktig eller nødvendig.

The question of how exactly this happens is irrelevant from the language-learner's perspective – what matters is that it does happen. The learner needs some basic knowledge of the language to move forward (bárdni is pronounced [baareʔni]; you have to know something about case forms and the fact that there is inflectional consonant gradation and various vowel changes). You can get much of that information from written books such as the Daavin course, though it is all in orthography and the spelling-to-speech mapping is tricky.

So to summarize, the one disadvantage of the existing GT tools is that the dictionary tools are not as tightly integrated into the product as I might like, and this is not a complaint. I would ordinarily complain that pronunciation is given short shrift, but that is necessitated by the huge linguistic diversity of the spoken language (and the substantial uniformity of the written language). There is a text-to-speech product, but it is (ahem) restricted technology using a USB dongle, and I have not experienced that speech. The earlier free online version was stunningly accurate.

As for "how do they do it?", things are pretty well documented and open, at least on the Giellatekno site.