Difference between revisions of "Munsee/Final Project"

Latest revision as of 17:58, 14 May 2025

Project Focus

Despite ongoing efforts to revitalize the Munsee language, relatively few pedagogical resources are readily available to the public. Our tool aims to support a crucial step in the learning of this language: the translation of the grammatically complex verbs in Munsee to English, a language likely to be spoken by learners of Munsee.

Approach

1. We first generated a parallel corpus of verb phrases in Munsee and English.
2. We added verb stems to our bilingual dictionary (.dix) and monolingual transducer (.lexd), creating twol rules as necessary.
   1. New twol rules for VTIs that take nouns objects
   2. Many more twol rules for specific verbs conjugations (accounting for accent shifts, instances of metathesis, contractions, etc.)
3. We then created syntax trees to model the breakdown of various verb phrases.
   1. Structural transfer rules were developed by tracking tags backwards up the trees
   2. Attributes currently being tracked:
      1. Transitivity
         1. Subject
         2. Object
      2. Animacy
      3. Number
      4. Polarity
      5. Gender (for simplicity, we default to 'he' for third person)
4. We implemented structural transfer rules for translation from Munsee to English.
   1. Issues resolved after presentation on 5/9/25:
      1. Implemented "it" objects (applicable to transitive verbs)
      2. Implemented subject tracking for transitive verbs
5. Finally, we evaluated our output (see evaluation section below)

Future Directions

• Further investigate structural transfer rules regarding the "do" form in the negative
• Develop structural transfer rules so third person subjects (e.g. names) replace third person pronouns
• Develop structural transfer rules so that noun objects take the correct article/determiner (e.g. 'a rock' instead of 'rock')
• There is currently a bug in the structural transfer module; once this is resolved, the negative with 'mah' can be re-implemented
• Systematize differences between stable and unstable verb stems
  • This will require work on both .lexd and .twol files
• Implement future 'tense' with modal word 'uch'

Relevant Links

• Munsee Transducer Repository
• Repository for English Transducer
• Repository for Munsee-English Language Pair (includes test files)
• Munsee Corpus Repository

Evaluation

As of May 8, 2025

Line Evaluation Results

• Number of words in reference: 175
• Number of words in test: 82
• Number of unknown words in test: 50
• Percentage of unknown words: 60.98%
• Word Error Rate (WER): 97.14%
• Position-Independent Word Error Rate (PER): 97.14%

As of May 14, 2025 AM

Line Evaluation Results

• Number of words in reference: 307
• Number of words in test: 282
• Number of unknown words in test: 0
• Percentage of unknown words: 0.00%
• Edit distance: 52
• Word Error Rate (WER): 17.59%
• Position-Independent Word Error Rate (PER): 17.26%

As of May 14, 2025 PM (after fixing structural transfer bug with 'mah')

Line Evaluation Results

• Number of words in reference: 306
• Number of words in test: 297
• Number of unknown words in test: 0
• Percentage of unknown words: 0.00%
• Edit distance: 45
• Word Error Rate (WER): 14.71%
• Position-Independent Word Error Rate (PER): 13.07%