Skip to content

Latest commit

 

History

History
253 lines (196 loc) · 11.4 KB

File metadata and controls

253 lines (196 loc) · 11.4 KB

0.12 — Cross-Linguistic Corpus Specification

Version: 0.12 Date: 2026-05-22 Status: Complete — Sprint 4.2 deliverable

1. Purpose

This specification defines the format, sourcing strategy, and expansion plan for the Nested Semantic Graph (NSG) cross-linguistic corpus. The corpus provides empirical ground truth for validating the central claim of "Few Become One" (DOI: 10.5281/zenodo.20328374): that diverse surface languages linearize the same Nested Semantic Tree — an ultrametric structure invariant under translation.

2. Current Corpus (Baseline)

The project corpus (0.9.py, build_expanded_corpus()) contains 12 documents across 5 languages:

Language Type Count Documents
English Isolating 6 EN-1 through EN-6
Turkish Agglutinative 2 TR-1, TR-2
Mohawk Polysynthetic 1 MH-1
Finnish Agglutinative 2 FI-1, FI-2
Inuktitut Polysynthetic 1 IK-1

All 12 share 1-2 core propositions (DOG-BITE-MAN), built from hand-crafted semantic trees. All pass ultrametric verification (0 violations).

Limitations:

  • Small (12 documents) — insufficient for quantitative claims
  • Narrow propositional coverage (predominantly DOG-BITE-MAN)
  • Hand-crafted trees, not parsed from real text
  • No pre-existing parallel corpus used

3. Target Expansion

The expansion target by project close-out (S4.4) is 25+ documents:

Requirement Current Target
Total documents 12 25+
Languages 5 8+
Language families 3 (IE, Turkic, Eskimo-Aleut) 5+
Proposition clusters 2 5+
Ultrametric verification 100% pass 100% pass

4. Corpus Format Specification

Each corpus entry is a (doc_id, sentence, gloss, english, language, source, tree_spec) tuple defined as follows:

4.1 Entry Fields

Field Type Required Description
doc_id string Yes Unique identifier: {LANG}-{N} (e.g., "JA-1")
sentence string Yes Surface sentence in source language (UTF-8, native script)
gloss string Yes Morpheme-by-morpheme gloss in English (Leipzig Glossing Rules)
english string Yes Natural English translation
language string Yes Language name (English)
source enum Yes Origin: HAND-CRAFTED, PARALLEL-TEXT, UD-TREEBANK, FIELDWORK, GRAMMAR
tree_spec dict Yes NST construction specification (see §4.2)
iso_639 string ISO 639-3 language code (e.g., "tur", "moh")
family string Language family (e.g., "Turkic", "Iroquoian")
typology enum Morphological type: ISOLATING, AGGLUTINATIVE, POLYSYNTHETIC, FUSIONAL
proposition_cluster string Semantic grouping (e.g., "TRANSFER-OBJECT", "MOTION-PATH")
source_ref string Citation or URL for source text
notes string Linguist's notes on interesting features

4.2 Tree Specification (tree_spec)

The tree_spec dict encodes how the surface sentence maps to an NST:

tree_spec = {
    "root": "S",                    # Root node category
    "verb": {                       # Verb phrase
        "stem": "bit",              # Verb root lemma
        "category": "BITE",         # Semantic category
        "tense": "PAST",            # TAM marking
        "voice": "ACTIVE",          # Voice
    },
    "participants": [               # Ordered argument list
        {"role": "AGENT", "stem": "dog", "category": "ANIMAL"},
        {"role": "PATIENT", "stem": "man", "category": "HUMAN"},
    ],
    "adjuncts": [                   # Optional modifiers
        {"type": "TEMPORAL", "stem": "yesterday"},
        {"type": "LOCATIVE", "stem": "park"},
        {"type": "INSTRUMENTAL", "stem": "stick"},
    ],
    "morphology": {                 # Language-specific encoding
        "case_marking": "NOM-ACC",  # or "ERG-ABS", "DIRECT-INVERSE"
        "agreement": "SUBJ",        # Agreement target
        "incorporation": False,     # Noun incorporation?
    }
}

4.3 Python Data Format

# Corpus entries as list of tuples
CORPUS_ENTRY = (
    doc_id: str,        # "JA-1"
    sentence: str,       # "Inu ga kino otoko o kanda"
    gloss: str,          # "dog NOM yesterday man ACC bit"
    english: str,        # "The dog bit the man yesterday"
    language: str,       # "Japanese"
    source: str,         # "PARALLEL-TEXT"
    tree_spec: dict,     # See §4.2
    _meta: dict          # Optional: {iso_639, family, typology, cluster, ref, notes}
)

5. Parallel Text Resources

5.1 Universal Dependencies (UD)

URL: https://universaldependencies.org/ Access: Free, open-source (CC BY-SA 4.0) Size: 200+ treebanks, 100+ languages Relevance: UD provides dependency parses with universal POS tags and dependency relations. The dependency tree can be transformed into an NST by mapping universal relations to semantic roles.

Mapping strategy:

UD Relation NST Role
nsubj AGENT (if active) / PATIENT (if passive)
obj / iobj PATIENT / RECIPIENT
obl ADJUNCT (LOCATIVE, TEMPORAL, INSTRUMENTAL)
root (VERB) VERB_PHRASE root

Recommended treebanks for NSG:

  • English (EWT, GUM) — isolating baseline
  • Turkish (IMST, BOUN) — agglutinative
  • Finnish (TDT) — agglutinative
  • Japanese (GSD) — SOV, postpositional
  • Hindi (HDTB) — split-ergative
  • Arabic (PADT) — non-concatenative morphology
  • Basque (BDT) — ergative-absolutive
  • Warlpiri (if available) — non-configurational

Limitation: UD trees capture syntactic dependencies, NOT semantic roles directly. NST requires semantic role mapping, which adds a manual step.

5.2 The Parallel Bible Corpus

URL: https://github.com/christos-c/bible-corpus Access: Free, public domain Size: 1,600+ languages, ~7,000 verse-aligned translations Relevance: The most extensive parallel corpus available. Every verse has translations in hundreds of languages, providing natural parallel data for NST isomorphism verification.

Recommended verses for NSG:

  • Genesis 1:1-3 (creation — causal chain)
  • Exodus 3:14 (being — existential)
  • Psalm 23:1-3 (shepherd — benefactive)
  • John 1:1 (logos — equative)
  • Matthew 5:3-5 (blessed — attributive)

Advantage: Guaranteed semantic parallelism — all translations encode the same proposition. Limitation: Biblical language is formal/archaic. May not reflect natural spoken language structures.

5.3 WALS (World Atlas of Language Structures)

URL: https://wals.info/ Access: Free, open-access Size: 192 structural features, 2,600+ languages Relevance: WALS provides typological features for language selection, ensuring the corpus spans diverse morphological and syntactic types. NOT a text corpus — used for language sampling, not document sourcing.

Features relevant to NSG sampling:

WALS Feature Relevance
20A: Fusion of inflectional formatives Morphological type (isolating → polysynthetic)
49A: Number of cases Case marking richness
81A: Order of Subject, Object, Verb Word order — affects linearization
98A: Alignment of case marking Nominative-accusative vs ergative-absolutive
101A: Expression of pronominal subjects Pro-drop parameter
102A: Verbal person marking Agreement morphology
104A: Order of possessor and possessum Genitive construction

Sampling strategy: Select 2 languages from each of the 5 major morphological types (isolating, agglutinative, fusional, polysynthetic, introflexive), ensuring at least one non-Indo-European per type.

5.4 Other Resources

Resource Type Relevance
ODIN (Online Database of Interlinear Text) Glossed text Pre-glassed examples from linguistic field manuals
Tatoeba Parallel sentences Community-contributed translations in 400+ languages
CMU Pronouncing Dictionary Phonological Not directly relevant to NST
Ethnologue Language catalog ISO codes, family classification
Glottolog Language catalog Comprehensive classification with references

6. Corpus Expansion Plan (Target: 25+ documents)

Phase 1: Propositional Diversity (6 new, target 18)

Add documents covering new proposition clusters:

Cluster Example Proposition Target Languages
TRANSFER-OBJECT "The woman gave the book to the child" English, Turkish, Finnish, Japanese, Hindi
MOTION-PATH "The man went from the house to the store" English, Turkish, Mohawk
PERCEPTION "The child saw the dog in the garden" English, Finnish, Inuktitut
CAUSATIVE "The woman made the child eat" Turkish, Japanese (causative morphology)
BENEFACTIVE "The man baked bread for the woman" English, Mohawk (applicative)

Phase 2: Language Diversity (6 new, target 24)

Add 3 new languages with 2 documents each:

Language Family Type Source
Japanese Japonic Agglutinative (SOV) UD-GSD treebank
Basque Isolate Ergative-absolutive UD-BDT treebank
Swahili Bantu Agglutinative (noun class) Grammar-based

Phase 3: Parallel Text (3 new, target 27)

Add 3 documents from the Parallel Bible:

Verse Languages Proposition
Genesis 1:1 English, Turkish, Mohawk "God created the heavens and the earth"
Psalm 23:1 English, Finnish, Inuktitut "The Lord is my shepherd"
John 1:1 English, Japanese "In the beginning was the Word"

7. Verification Protocol

Each new corpus entry MUST pass:

  1. Lexical coverage: Every morpheme in tree_spec has a corresponding entry in the language's LEXICON
  2. Tree construction: build_nst_from_morphology() produces a valid tree with ≥2 leaves
  3. Ultrametric property: All triples of leaf nodes satisfy the strong triangle inequality (0 violations)
  4. Isomorphism check: Documents in the same proposition cluster share isomorphic tree structure (same node count, same category hierarchy, same leaf ordering up to linearization)

8. Deliverables

File Purpose
0.12.md This specification
0.12.py Corpus loader and format validator (optional, for S4.3 integration)
0.9.py Expanded corpus implementation (existing, target expansion)

9. Open Questions

  1. Real Zemberek? The Turkish pipeline (0.10.py) simulates Zemberek output. Should S4.3-S4.4 attempt actual Zemberek integration? Decision: NOT in this sprint. Simulation is sufficient for prototype.
  2. UD dependency → NST mapping: The mapping is non-trivial (syntax ≠ semantics). Should we implement a rule-based mapper? Decision: Manual tree specification for key examples, automated mapper as future spinoff.
  3. Publication scope: Should the corpus spec be included in the final publication (0.7.md)? Decision: Yes — as §3 (Methodology) appendix.

10. References