Difference between revisions of "Kickapoo and English"

Latest revision as of 01:58, 17 May 2026

Resources for machine translation between Kickapoo and English

External Resources

1. Repository for KIC-ENG: https://github.swarthmore.edu/Ling073-sp26/ling073-kic-eng
2. Kickapoo Transducer: https://github.swarthmore.edu/Ling073-sp26/ling073-kic
3. English Transducer: https://github.swarthmore.edu/Ling073-sp26/apertium-eng

Developed Resources

1. Parallel Corpus: https://github.swarthmore.edu/Ling073-sp26/ling073-kic-eng-corpus

Additions

For the polished RBMT assignment, we have made the below expansions:

dix File Expansion

• Added 130 word stems in monolingual and bilingual dictionary; Specified Aorist tenses (2nd and 3rd) as well as added additional words (such as extra <vaux>).

Additional Disambiguation

• We've added two new rules to further distinguish ambiguity surrounding <prn> vs <det> that arises in manahi ('this') and iinini ('that'):

 SELECT Prn
 IF (-1 BOS) (0 Prn) (0 Det) (1 Vcnj) (2 EOS);

 SELECT Prn
 IF (-2 BOS) (-1 Verb) (0 Prn) (0 Det) (1 Noun) (2 EOS);

• For the word eecinoat that means both "be cheap" and "be easy", we've added a rule to distinguish their scenarios. Here's an example where it would be mapped to "cheap" before the word "coffee":

 <rule weight="1.2">
     <match lemma="eecinoat" tags="v.*">
       <select lemma="be cheap" tags="v.*"/>
     </match>
     <match lemma="kaahpiih" tags="n.*"/>
 </rule>

Polished rtx File

Transfer Rule Pattern 1: Negative Particle + Finite Verb

The first rule is dealing with the negative particle in the morphological transducer. We tested it with

Input:  aakwi noweewi
Before: not he said
After:  he did not say

This output is much closer to an acceptable English translation because English finite negation usually requires auxiliary do plus not, followed by the base form of the verb.

Transfer Rule Pattern 2: Locative Construction

We also attempted to handle locative constructions such as:

atoopooneki aiwi

meaning:

he was at the table

In this construction, the locative noun atoopooneki is first reduced into a DP with the tags sg and loc. The verb aiwi is reduced into a main-clause intransitive VP with past tense and subject agreement information. I then added a structural transfer rule that combines a locative DP and a main VP into a full English sentence.

Initial Coverage Analysis

1. monolingual: coverage: 94 / 166 (~0.56626506024096385542)
2. bilingual: coverage: 144 / 575 (~0.25043478260869565217)

KIC to ENG Evaluation

Ambiguity Resolution

Please see: Kickapoo and English/Ambiguity Resolution

Final Evaluation

Monolingual transducer

Machine translation: Kickapoo-English

Compared with the midterm evaluation, the final system has a much lower precision but a higher recall. At midterm, the system had high precision (92.31%) but low recall (29.45%), which means that the analyses it produced were usually correct, but it failed to analyze many forms in the gold standard.

In the final evaluation, recall increased to 39.39%, showing that the additional stems and expanded morphology helped the system recognize more correct forms. However, precision dropped to 31.90% because the system also produced many extra analyses that were not in the gold file. The final result has 52 true positives but 111 false positives, which suggests that the analyzer is now overgenerating.

This likely happened because we added many new stems and lexical categories, including some ambiguous entries where the same surface form can be analyzed as multiple parts of speech or as both animate and inanimate. We also used the same gold standard that we previously created. However, the gold standard should now be substantially different, especially since aorist forms have now been added to the transducer. Therefore, some analyses that are currently counted as false positives may actually reflect new analyses that were not represented in the original gold standard.

The WER is very high partly because the longer corpus is not made up of short isolated sentences, but instead comes from a complete tale about a motorcyclist. Translating this kind of narrative fluently would require many more structural transfer rules than our system currently has. Many Kickapoo sentences do not map directly onto English word order, and in several cases the English translation requires major reordering rather than word-by-word substitution. In addition, some English translations are idiomatic or context-dependent rather than literal. For example, a form based on the stem meaning “bother” may be translated as “harass” in the context of the story. Similarly, compounds or descriptive expressions may need to be translated into more natural English lexical items, such as “food-building” being translated as “restaurant.” Because our system mostly handles simpler lexical transfer and only a limited number of structural patterns, it often produces output that is partially analyzable but still far from the reference translation. As a result, the WER is high not only because of unknown words, but also because the reference translation reflects fluent English narrative structure and idiomatic meaning that our current transfer system does not yet fully capture.