Zazaki/Transducer
For code, check here: https://github.swarthmore.edu/Ling073-sp26/ling073-zza.git
Analyzer Evaluation
Notes
Tests that don't match gold and why:
- mar<n> <m> <f> <sg> <dir> - Uses both <m> and <f> tags simultaneously, which is non-standard for Apertium. Would need a separate gender class for dual-gender nouns.
- panj roj-i - Multiword expression with a hyphenated plural. Lexd cannot handle multiword units as single analyses.
- merdımêndo and çayênda (indefinite + ezafe) - The ezafe suffix interacts with the indefinite suffix (-ên + ezafe), requiring additional morphology beyond simple indefinite -ê. Not yet implemented.
- Number/Case ambiguity - Masculine singular oblique and plural direct both surface as -i (e.g., roji, bergiri, laji). The analyser correctly produces both analyses; this is genuine ambiguity in the language.
Test Breakdown
Here's the test and match gold result for each corpus:
Corpus 1 of 16: Gender-morph: 4/4 (100.0%) tests pass (3/4 (75.0%) match gold)
Corpus 2 of 16: Number-morph: 3/3 (100.0%) tests pass (0/3 (0.0%) match gold)
Corpus 3 of 16: Case and Specificity-morph: 5/5 (100.0%) tests pass (3/5 (60.0%) match gold)
Corpus 4 of 16: Indefinite Suffix-morph: 4/4 (100.0%) tests pass (2/4 (50.0%) match gold)
Corpus 5 of 16: Personal Pronouns-morph: 6/6 (100.0%) tests pass (6/6 (100.0%) match gold)
Corpus 6 of 16: Interrogative Pronouns-morph: 5/5 (100.0%) tests pass (5/5 (100.0%) match gold)
Corpus 7 of 16: Descriptive Adjectives-morph: 5/5 (100.0%) tests pass (5/5 (100.0%) match gold)
Corpus 8 of 16: (Demonstrative) Determiners-morph: 4/4 (100.0%) tests pass (4/4 (100.0%) match gold)
Corpus 9 of 16: Present Tense-morph: 4/4 (100.0%) tests pass (4/4 (100.0%) match gold)
Corpus 10 of 16: Preterite Tense-morph: 4/4 (100.0%) tests pass (4/4 (100.0%) match gold)
Corpus 11 of 16: Subjunctive Mood-morph: 3/3 (100.0%) tests pass (2/3 (66.67%) match gold)
Corpus 12 of 16: Imperative-morph: 6/6 (100.0%) tests pass (6/6 (100.0%) match gold)
Corpus 13 of 16: Adverbs-morph: 5/5 (100.0%) tests pass (5/5 (100.0%) match gold)
Corpus 14 of 16: Postpositions-morph: 5/5 (100.0%) tests pass (5/5 (100.0%) match gold)
Corpus 15 of 16: Conjunctions-morph: 3/3 (100.0%) tests pass (3/3 (100.0%) match gold)
Corpus 16 of 16: commonwords-morph: 14/14 (100.0%) tests pass (0/14 (0.0%) match gold)
Not Yet Implemented
Verb negation — Present tense negation prefix nê- and prohibitive prefix me- are not implemented.
Copula and verb of existence — These irregular verb paradigms (beno, esto, etc.) are not yet in the transducer.
Periphrastic verb forms — Perfect, past perfect, imperfect, and progressive tenses are not implemented.
Coverage note — By adding ra<post>, u<cnjcoo>, zî<post>, and rê<post> to the transducer, coverage went from 23.94% to 29.42%.
Evaluation
Stem counts (excluding inflection, punctuation, and anonymous lexicons)
Nouns: 11 (8 masc + 3 fem)
Adjectives: 2 (1 consonant-final + 1 vowel-final)
Determiners: 4
Personal pronouns: 6
Interrogative pronouns: 5
Verbs: 6 (1 present + 1 preterite + 1 subjunctive + 3 imperative)
Adverbs: 5
Postpositions: 8
Conjunctions: 4
Total: 51 stems
Coverage
The initial coverage is 23.94%. By adding ra<post>, u<cnjcoo>, zî<post>, and rê<post>, the improved coverage is 29.42%.
Test Analysis
What tests still don't work
All tests are passed. 76/76 pass (all corpora)
For commonwords-morph-*.txt files, 14/14 commonwords passed.
List of Top Unknown Words
| Count | Unknown Word |
|---|---|
| 48 | ke |
| 45 | bi |
| 33 | de |
| 29 | xo |
| 25 | Îsayî |
| 19 | va |
| 16 | cı |
| 16 | Fatose |
| 13 | xwu |
| 12 | ki |
| 10 | kerd |
| 9 | ser |
| 9 | da |
| 9 | Ali |
| 8 | vêrdi |
| 7 | mi |
| 7 | eno |
| 7 | biye |
| 7 | beno |
| 7 | Homayî |
Generator Evaluation
Initial evaluation of morphological generation
At the start of generator work (tagged as "analyser"), the generation test results were:
| Corpus | Pass | Gold Match |
|---|---|---|
| Gender-gen | 4/4 | 3/4 |
| Number-gen | 3/3 | 2/3 |
| Case and Specificity-gen | 5/5 | 5/5 |
| Indefinite Suffix-gen | 4/4 | 2/4 |
| Personal Pronouns-gen | 6/6 | 6/6 |
| Interrogative Pronouns-gen | 5/5 | 5/5 |
| Descriptive Adjectives-gen | 5/5 | 5/5 |
| (Demonstrative) Determiners-gen | 4/4 | 4/4 |
| Present Tense-gen | 4/4 | 4/4 |
| Preterite Tense-gen | 4/4 | 4/4 |
| Subjunctive Mood-gen | 4/4 | 4/4 |
| Imperative-gen | 6/6 | 6/6 |
| Adverbs-gen | 5/5 | 5/5 |
| Postpositions-gen | 5/5 | 5/5 |
| Conjunctions-gen | 3/3 | 3/3 |
| Total | 72/72 | 68/72 (94.4%) |
Initial coverage on the basic corpus (zza.corpus.basic.txt): 29.42%
Final evaluation of morphological generation
To improve coverage, we tested against a larger corpus (zza.bible_luke1.txt, 25,537 tokens) and added high-frequency unknown stems. The added stems include common postpositions (ra, do, ci), conjunctions (ke, de, bi, u), adverbs (wina, her, ey), pronouns (xwu, înan, mi), and nouns (ser, yew).
Twol rules
0 twol rules were added. The morpheme boundary deletion (>:0) was already handled in the Alphabet section of the twol file. No morphophonological alternations required additional rules, as Zazaki morphology in our implementation uses clean suffixation.
Updated stem counts
Nouns: 13
Adjectives: 2
Determiners: 4
Personal pronouns: 9
Interrogative pronouns: 5
Verbs: 6
Adverbs: 8
Postpositions: 9
Conjunctions: 7
Total: 63 stems
Final generation test results
72/72 tests pass, 68/72 match gold (94.4%). Results are unchanged from initial evaluation since no twol rules were modified.
Coverage
| Corpus | Coverage |
|---|---|
| zza.corpus.basic.txt (initial) | 23.94% |
| zza.corpus.basic.txt (after adding common words) | 29.42% |
| zza.bible_luke1.txt (initial) | 27.24% |
| zza.bible_luke1.txt (after adding common words) | 40.92% |
Top unknown words (zza.bible_luke1.txt)
| Count | Unknown Word |
|---|---|
| 152 | Îsayî |
| 121 | va |
| 99 | şima |
| 74 | kerd |
| 72 | ci |
| 70 | Homayî |
| 61 | Îsa |
| 55 | da |
| 51 | yê |
| 50 | enê |
| 45 | bikerê |
| 40 | eno |
| 39 | bî |
| 34 | yo |
| 31 | verê |
| 29 | seba |
| 29 | kesî |
| 28 | ê |
| 27 | kerdêne |
| 27 | dî |
