PropMatch Encoder Bugs

[abhishekreddyallu]

Property encoder workflows have dispatch, prompt-context, and inverse-label inconsistencies

1. Keyword dispatch breaks property LLM encoder overrides

Location: ontoaligner/encoder/llm.py:39, ontoaligner/encoder/llm.py:44

Current code used keyword dispatch:

self.get_owl_items(owl=source)
self.get_owl_items(owl=target)

But PropertyLLMEncoder overrides use a different parameter name, such as:

def get_owl_items(self, prop):

This caused property encoding to fail.

Observed before fix:

TypeError: ... got an unexpected keyword argument 'owl'

Fix:

encoded_source = self.get_owl_items(source)
encoded_target = self.get_owl_items(target)

Verified after fix:

PropertyLLMEncoder completed successfully.
PropertyFullTextLLMEncoder completed successfully.

---

2. Full-text property context was missing from LLM prompts

Location: ontoaligner/encoder/property.py
Related: ontoaligner/aligner/llm/dataset.py

The PropertyFullTextLLMEncoder and the PropertyFullTextLLMDataset were not using a consistent encoded schema.

The encoder prepared domain/range/inverse context, but the prompt dataset was reading fields in a way that could leave these sections empty or process already-encoded text incorrectly.

Expected prompt fields as PropMatchEncoder:

domain_text
range_text
inverse_text

Fix in PropertyFullTextLLMEncoder:

return {
    **prop,
    "domain_text": domain_text,
    "range_text": range_text,
    "inverse_text": inverse_text,
    "text": combined_text,
}

Fix in PropertyFullTextLLMDataset:

source_domain = self.preprocess(input_data["source"].get("domain_text", ""))
target_domain = self.preprocess(input_data["target"].get("domain_text", ""))

source_range = self.preprocess(input_data["source"].get("range_text", ""))
target_range = self.preprocess(input_data["target"].get("range_text", ""))

source_inverse = self.preprocess(input_data["source"].get("inverse_text", ""))
target_inverse = self.preprocess(input_data["target"].get("inverse_text", ""))

Verified after fix:

PropertyFullTextLLMDataset created successfully

Domain present: True
Range present: True
Inverse present: True

PASS: Full-text property context is present in the generated prompt.

---

3. Inverse labels were split character-by-character

Location: ontoaligner/encoder/property.py:136
Location: ontoaligner/encoder/property.py:264

OntologyProperty.extract_data() stores inverse_label as a string.

Example:

"inverse_label": "transforms into"

But the encoders used " ".join(...) on inverse_label, which could split a string into characters.

Problem pattern:

inverse_text = ""
        if property['inverse_of']:
            inverse_text = " ".join(property['inverse_label']) if len(property['inverse_label']) > 0 else ""

This can produce incorrect text such as:

t r a n s f o r m s   i n t o

Fix:

inverse_text = ""
if property.get("inverse_of"):
    inverse_text = property.get("inverse_label") or ""

And for PropertyFullTextLLMEncoder:

inverse_text = ""
if prop.get("inverse_of"):
    inverse_text = prop.get("inverse_label") or ""

Verified after fix:

Raw inverse label: 'transforms into'
Encoded inverse: 'transforms into'
PASS: inverse label remains a complete string.