EXTRACT STATEMENTS.
MAP RELATIONSHIPS.

# SKILL: Statement Extraction with corp-extractor

Use the `corp-extractor` Python library to extract structured subject-predicate-object statements from text. Returns Pydantic models with confidence scores.

## Installation

```bash
pip install corp-extractor[embeddings]  # Recommended: includes semantic deduplication
```

For GPU support, install PyTorch with CUDA first:
```bash
pip install torch --index-url https://download.pytorch.org/whl/cu121
pip install corp-extractor[embeddings]
```

## Quick Usage

```python
from statement_extractor import extract_statements

result = extract_statements("""
    Apple Inc. announced the iPhone 15 at their September event.
    Tim Cook presented the new features to customers worldwide.
""")

for stmt in result:
    print(f"{stmt.subject.text} ({stmt.subject.type})")
    print(f"  --[{stmt.predicate}]--> {stmt.object.text}")
    print(f"  Confidence: {stmt.confidence_score:.2f}")
```

## Output Formats

```python
from statement_extractor import (
    extract_statements,        # Returns ExtractionResult with Statement objects
    extract_statements_as_json,  # Returns JSON string
    extract_statements_as_xml,   # Returns XML string
    extract_statements_as_dict,  # Returns dict
)
```

## Statement Object Structure

Each `Statement` has:
- `subject.text` - Subject entity text
- `subject.type` - Entity type (ORG, PERSON, GPE, etc.)
- `predicate` - The relationship/action
- `object.text` - Object entity text
- `object.type` - Object entity type
- `source_text` - Original sentence
- `confidence_score` - Groundedness score (0-1)
- `canonical_predicate` - Normalized predicate (if taxonomy used)

## Entity Types

ORG, PERSON, GPE (countries/cities), LOC, PRODUCT, EVENT, WORK_OF_ART, LAW, DATE, MONEY, PERCENT, QUANTITY, UNKNOWN

## Precision Mode (Filter Low-Confidence)

```python
from statement_extractor import ExtractionOptions, ScoringConfig

options = ExtractionOptions(
    scoring_config=ScoringConfig(min_confidence=0.7)
)
result = extract_statements(text, options)
```

## Predicate Taxonomy (Normalize Predicates)

```python
from statement_extractor import PredicateTaxonomy, ExtractionOptions

taxonomy = PredicateTaxonomy(predicates=[
    "acquired", "founded", "works_for", "headquartered_in"
])
options = ExtractionOptions(predicate_taxonomy=taxonomy)
result = extract_statements(text, options)

# "bought" -> "acquired" via semantic similarity
for stmt in result:
    if stmt.canonical_predicate:
        print(f"Normalized: {stmt.predicate} -> {stmt.canonical_predicate}")
```

## Batch Processing

```python
from statement_extractor import StatementExtractor

extractor = StatementExtractor(device="cuda")  # or "cpu"
for text in texts:
    result = extractor.extract(text)
```

## Best Practices

1. Use `[embeddings]` extra for semantic deduplication
2. Filter by `confidence_score >= 0.7` for high precision
3. Use predicate taxonomies for consistent knowledge graphs
4. Process large documents in chunks (by paragraph/section)
5. GPU recommended for production (~2GB VRAM needed)

## Links

- PyPI: https://pypi.org/project/corp-extractor/
- Docs: https://statement-extractor.corp-o-rate.com/docs
- Model: https://huggingface.co/Corp-o-Rate-Community/statement-extractor