Assertion Extraction from Biomedical Abstracts

This repository contains all code, configuration, and modules for an experimental pipeline to extract scientific assertions (subject–predicate–object triples) from biomedical literature. The core focus is on evaluating and comparing different LLMs for multi-stage information extraction.

---

Project Overview

The pipeline is divided into three main stages:

1. Finding Detection – Identify which sentences in an abstract contain scientific findings.
2. Completion and Resolution – Convert finding sentences into standalone factual statements.
3. Assertion Extraction – Convert standalone statements into structured triples: (subject, predicate, object), with optional conditions.

---

Directory Structure

Folder / File	Description
01_sentence_classification/	Sentence-level classification experiments for finding detection.
01.1_binary_classification/	Old binary classifier experiments (e.g. SciBERT vs Logistic) – deprecated.
02_extract_results_CITATIONS/	Tools to extract and prepare PubMed citation abstracts (CSV files, PMIDs).
02.1_exmine_results_tag_API/	Evaluation of OpenAI API tagging outputs (e.g., result/finding matching).
03_extract_article_summary/	Early experiments in summarization / simplification of article content.
04_classification_model/	Transformer-based sentence classification models (e.g., BioBERT). – deprecated
05_three_single_pipeline/	Main pipeline experiment. Includes finding → completion → assertion.
.gitignore	Specifies excluded file types (e.g. .npy, .pt).
environment.yml	Conda environment file with all required dependencies.

---

Current Main Experiment: 05_three_single_pipeline/

This folder includes the three sequential LLM pipelines:

• finding_pipeline.py – Run GPT/Claude/LLaMA to identify finding sentences.
• completion_pipeline.py – Expand sentences into standalone factual units.
• assertion_pipeline.py – Extract subject–predicate–object triples.

Additional utilities:

• schemas.py, json_utils.py, prompts.py – shared dataclass schemas, JSON parsing, and prompt templates.

• config.py – API keys and model configuration (use environment variables in production).

• eval/ – evaluation outputs and summary statistics.

---

Environment Setup

conda env create -f environment.yml
conda activate hf-hpc