text2rel

TestTestTest

pip install text2rel

This is the package implementation of the code for the kinship ties and relational information from large text corpora, such as genealogies, biographies, and historical dictionaries.

Important information:

• For example package usage, see test_workflow.ipynb
• For information about individual functions before the package implementation and release, see examples folder.

Expected input files

Each document should normally have these three files together in the same folder under documents_root:

0_GenelogiesAndBiographies/
├── Example genealogy/
│   ├── Example genealogy_mod.htm       # source HTML used for the inventory
│   ├── Example genealogy_mod.pdf       # PDF with a readable OCR text layer
│   └── Example genealogy_Original.pdf  # original PDF without OCR
└── Biographies/
    └── Example biography/
        ├── Example biography_mod.htm
        ├── Example biography_mod.pdf
        └── Example biography_Original.pdf

The filenames must share the same base name. The package uses the HTML for inventory creation, prefers *_mod.pdf for page matching, and falls back to *_Original.pdf, which can be OCRed with Tesseract when necessary.

Current release:

Current release includes inventory creation, setting text bounds, assigning font usage and selecting appropriate text chunks from the files (including restriction to "MainText" fonts and JumPJumP insertion).

Files structure:

1. cleaner.py

Contains the HTMLCleaner class and thus the main logic.

2. utils.py

Contains helper functions (like JumPJumP insertion)

3. io.py

Contains the file processing logic, such as loading and saving the JSON inventory.

4. page_matching.py

Adds one-based PDF start_page and end_page values to chunk metadata using exact and fuzzy text matching. Missing PDFs are skipped with null page values and explicit filename guidance.

Add PDF page numbers

If you already created an HTMLCleaner, use cleaner.add_pdf_pages(). If you only have an inventory JSON file, use add_pages_to_inventory_file() directly.

summary = cleaner.add_pdf_pages(
    documents_root="../0_GenelogiesAndBiographies",
    output_path="exact_html_inventory_new_ids_cleaned_pages.json",
)

The default fuzzy threshold is 0.80. The matcher prefers *_mod.pdf, falls back to _Original.pdf when needed, and can use Tesseract for image-only originals when Tesseract is installed.

Human Labeler Allocation

Automatic project creation attaches the existing Kinship ontology to Relations and Events ontology to Events, then starts new projects after upload and member setup. Existing projects are still skipped. Other workspaces can pass ontology_ids={"Relns": "kinship_id", "Events": "events_id"} to upload_labelers_data_to_labelbox().

When text chunks are allocated to labelers, each generated TXT filename indicates its task type:

• Task_0 contains chunks shared across all labelers.
• Task_1 through Task_9 contain labeler-specific, unshared chunks.

New section 4 assessment workflow

The reusable assessment follows sections 4.2.2.1 and 4.2.2.2 of 240426 Information extraction system 260801.ipynb. Install the optional plot dependencies, load the processed human and AI tables, and then create the assessment tables before drawing anything:

pip install -e ".[plots]"

from text2rel import (
    ReliabilityAssessment,
    build_sweep_report,
    plot_sweep_report,
    plot_threshold_curves,
)

assessment = ReliabilityAssessment("inventory.json")
assessment.load_relations_data(relations_csv="human_relations.csv")
assessment.load_chatgpt_relations("machine_relations.csv")

# one_to_one is the package's stricter default for headline metrics.
relation_results = assessment.assess_llm_against_humans(data="relations")
plot_threshold_curves(relation_results["threshold_metrics"], direction="one_to_one")

# directional_best reproduces the notebook's two reusable-candidate views.
diagnostics = assessment.assess_llm_against_humans(
    data="relations", matching_mode="directional_best"
)
fn_report = build_sweep_report(
    diagnostics["human_to_llm_best_pairs"], data="relations"
)
plot_sweep_report(fn_report)

assess_llm_against_humans() returns the selected pairs, threshold metrics, false-negative and false-positive review tables, structural exclusions, and configuration metadata. Plot functions only render these returned tables; they do not change matching or scores. See the separate New section 4 assessment plots and tables section at the end of test_workflow.ipynb for relations and events, notebook-style titles, confidence/length diagnostics, and review summaries.

Convert downloaded OpenAI batches to CSV

Use the submitted cleaned inventory (or the exact submitted subset) to restore chunk IDs. The function accepts a JSONL path or download_batch_results() output:

CSV saving is automatic: omitted or None output_csv writes llm_relations.csv or llm_events.csv in the current working directory. Pass a path to choose another destination. Reruns overwrite the destination; both DataFrames are still returned.

from text2rel import process_batch_results

relations, relation_audit = process_batch_results(
    "batch_output_relations.jsonl", submitted_inventory,
    extraction_type="relations", output_csv="llm_relations.csv",
)
events, event_audit = process_batch_results(
    "batch_output_events.jsonl", submitted_inventory,
    extraction_type="events", output_csv="llm_events.csv",
)
assessment.load_llm_relations("llm_relations.csv")
assessment.load_llm_events("llm_events.csv")

Taxonomy can be applied to the returned tables before assessment. These are the ungrouped load_llm_* schemas, not historical load_chatgpt_* grouped CSVs. Errors or missing responses raise by default; pass strict=False to inspect partial results and their audit. Keep the audit, which also records successful empty extractions. Custom schemas must retain s/r/t or S/V; extra fields are preserved. For historical req_N IDs, pass an explicit mapping from each request ID to its original chunk metadata (including chunk_id).

Download Labelbox annotations to human CSVs

Use download_labelbox_results(project_ids, allocation, tasks_dir, output_dir) with project_ids={"relations": ["project_id"], "events": ["project_id"]}. It reuses the Done-only exporter and saves human_relations.csv, human_events.csv, and coverage CSVs. Each returned dictionary value is a (facts, coverage) pair. Supply api_key and optionally author_map.

Keep the original allocation JSON and uploaded TXT files unchanged: offsets in the TXT locate each annotation; allocation row IDs recover original chunk IDs. For manual JSON/JSONL exports use process_labelbox_results(exports, allocation, tasks_dir, "relations", output_csv) (or "events"). Ambiguous mappings raise before saving that CSV. Coverage records empty completed annotations separately; the current assessment does not automatically score these empty chunks. Downloads include only Done tasks and do not establish that every assigned task has been completed.

Labelbox event construction modes

labelbox_events_to_dataframe(..., construction_mode="graph") is the default. It follows section 2.4.3.2 of the 260801 notebook: connected annotations are rebuilt around one Verb, safe reversed arrows are corrected, and incomplete or ambiguous structures remain auditable. Use construction_mode="legacy_simple" only when explicitly reproducing the package's earlier permissive conversion. Relation conversion uses its audited endpoint joins separately.