CAP-ACL-2026

Code and artifacts for the BioNLP 2026 paper:

CAP: A Source-Grounded Proposition Scaffold for Faithful Clinical Dialogue-to-Note Generation

This work was conducted at AITRICS.

This repository is organized so that readers can:

• Inspect the end-to-end pipeline implementation (X -> C -> E -> Y).
• Inspect prompt-based reimplementations of baselines (Direct, Cluster2Sent-inspired, MEDSUM-ENT-inspired).
• Reproduce paper tables/figures from the included precomputed outputs without rerunning LLM inference.

Repository Layout

• code/: Main runner scripts and shared utilities.
• prompts/: Selected prompt text files; most prompts are defined as Python string constants in code/.
• results/: Precomputed outputs (CAP extraction + rendering/evaluation) used to build the quantitative tables and figures.
• examples/: Small qualitative bundles used in the paper (generated notes for selected cases).
• data/: No raw benchmark data is redistributed here; see data/README.md.

Quickstart (Reproducing Tables/Figures From Included Results)

The precomputed results are stored under:

• results/problem_state_tracking_full207_v11_shared_gemma3/
• results/template_rendering_full207_v11_main_ablation_eval/

Key files:

• aggregate_metrics.csv
• case_metrics.csv

Running The Pipeline (Optional)

Running the full pipeline requires access to LLM endpoints (local or hosted) and API keys. We recommend using explicit --api-base-url / --judge-api-base-url arguments rather than relying on any defaults.

Main entrypoints:

• code/run_problem_state_tracking_experiments.py (CAP extraction)
• code/run_template_rendering_experiments.py (note rendering + evaluation)
• tools/run_deployment_cost_subset_e2e.py (deployment-path latency/token analysis on subset: X->Y, X->C->Y, X->C->E->Y)

Example (deployment-path cost analysis on a 30-case subset):

python3 tools/run_deployment_cost_subset_e2e.py --subset-size 30 --seed 57 --force-regenerate

Prompts

Prompts are implemented as Python constants inside the runner scripts (search for PROMPT in code/). For reviewer convenience, we export prompt snapshots to prompts/ with an index:

• prompts/INDEX.md

To regenerate these snapshots from code/:

python3 tools/export_prompts.py

Notes

This repository intentionally excludes:

• .env files / API keys
• personal identifiers
• legacy backups and experimental scratch files