Kikuyu/Transducer

The Kikuyu Transducer, in its work-in-progress state, can be found online at Swarthmore's Github.

Generator

Upon changes to the generator, the coverage has gone down to 37.58%, but there appear to be far fewer false positives (although some words which were formerly analysed correctly are no longer due to complications; this will hopefully be corrected in the future).

Of the new tests, 118 of the 147 pass, and 29 fail. (These tests work differently than the analyser's tests, in that they count unexpected results as failures; that is, they do not only require a correct analysis, but also a lack of incorrect analyses). The reason for some of the failures is unknown, but others occur due to known issues which have not been resolved, such as issues with the negation system regarding which negation marker is allowed in different situations. Other failures may not be failures at all; although the forms they produce are not anywhere in the corpus available, there is no clear reason yet why they are not correct forms (other than perhaps idiosyncratic variation). There is likely still much to understand with regard to the tense and aspect system that may eliminate the number of false forms generated, although the generator's handling of tense and aspect is far more sophisticated than that of the analyser.

A total of 27 rules were added to twol for this step, as well as several to twoc.

Analyser

Evaluation

The transducer is currently able to analyze 49.72% of the corpus.

It is very important to note that this number may be misleading, since it is possible the transducer is analyzing forms in the corpus incorrectly; given the massive overgeneration of the transducer in its current state, it is likely it generates an enormous number of forms that it should not, and it is possible some of those are coincidentally actually words. However, the transducer produces accurate analyses for the words in the test files.

Furthermore, it is important to note that the transducer was built from words in this corpus, so it is unsurprising it is capable of analyzing a large chunk of it. (In particular, many of the roots which are prevalent in the corpus can be analyzed by the transducer.)

At the time of this writing, 4 words in commonwords.yaml can be analyzed correctly, and 123 words in kik.yaml can be analyzed correctly. The top unanalyzed words in the corpus are now:

• 29 nĩ
• 16 thiomi
• 10 kana
• 8 nako
• 6 no
• 6 Abĩrika
• 5 mũtego
• 5 ũndũ
• 4 harĩa
• 4 Mũtiiri
• 4 ũyũ
• 4 ta
• 4 mũtiiri
• 4 rũũi
• 3 kuona
• 3 ciothe
• 3 njamba
• 3 maingĩ
• 3 ng
• 3 thĩ

Of these, nĩ is known to be a form of the copula, a very complicated piece of morphology. Abĩrika means Africa, and Mũtiiri and mũtiiri are both the name of the website from which some of the corpus was taken.

Currently, the transducer can analyze 82 noun stems, 54 verb stems, and 15 adjective stems, for a total of 151 stems.

Notes

At the moment, there are seven test failures in kik.yaml:

• ndigũthiĩ: It is unclear why this fails; in general, there seems to be a problem with negatives beginning with nd-, although it is not obvious why.
• wagui: This is an irregular plural, and has not been implemented.
• guoko: This is a plural word with a slightly abnormal plural morpheme, gu- instead of gũ; the .lexc file is not yet able to deal with it.
• ndĩagũire: Again, this is a negative word beginning with nd-, and it is failing for an unknown reason.
• ndũthiaga: See above.
• ndaroka: See above.
• rũigi: The singular prefix rũ- does not appear when the root already begins with rũ-, as in rũigi. The .lexc file is not yet able to account for this.

At first, the transducer was able to analyze 37.04% of the corpus. After adding the ability to analyze several common words (na "with", kagui "boy", maĩ "water", and o "all"), the transducer was able to analyze 49.72% of the corpus.