Context Convergence Improves Answering Inferential Questions

🧠 Why do LLMs struggle with inferential questions? Because not all context is equally helpful—some sentences guide reasoning, others just add noise.

💡 What if we could measure how useful a sentence is for reasoning? This work shows that convergence—how much a sentence narrows down possible answers—plays a key role in improving inferential QA.

🌟 Overview

Large Language Models (LLMs) are powerful—but they still struggle with inferential questions 🤔 (those where answers must be reasoned, not directly found).

💡 In this project, we introduce convergence as a signal that measures how well a sentence (hint) narrows down possible answers.

🔍 What we show:

• ✅ High-convergence sentences → better QA performance
• 📊 Convergence > cosine similarity for passage selection
• 🧠 Ordering sentences by convergence → even better results

🗂️ Repository Structure

├── dataset                                  # Data preparation and evaluation utilities
│   ├── compute_similarities.py              # Computes cosine similarity scores
│   ├── dataset_final.tar.gz                 # Ready-to-use final dataset for experiments
│   ├── make.sh                              # Rebuilds the dataset pipeline from scratch
│   ├── make_dataset.py                      # Creates dataset with convergence annotations
│   ├── merge.py                             # Merges intermediate outputs into final dataset
│   ├── qa.py                                # Runs QA evaluation pipeline
│
└── experiments                              # Experiment scripts used in the paper
    ├── convergence_vs_cosine.py             # Compares convergence vs cosine similarity
    ├── order.py                             # Tests effect of sentence ordering

📦 Dataset

📍 Preprocessed dataset included:

dataset/dataset_final.tar.gz

• ✅ Recommended: Use this directly
• ⚠️ Optional: Rebuild from scratch if needed

🧩 What’s inside the dataset?

The dataset is derived from hint-based QA data (TriviaHG) and is designed for inferential question answering. Unlike standard QA datasets, the answer must be inferred by combining hints, not extracted from a single sentence.

📊 Convergence in the dataset

Each hint has a convergence score, measuring how well it narrows down candidate answers:

• 🟢 High → strongly filters incorrect answers
• 🟡 Medium → partially informative
• 🔴 Low → weak or ambiguous

🧪 Running Experiments

⚖️ Convergence vs Cosine

python experiments/convergence_vs_cosine.py

🔢 Sentence Ordering

python experiments/order.py

🔁 Reproducibility

You can reproduce the paper in two ways:

✅ Option A — Reproduce using the provided dataset

This is the easiest and recommended way.

Step 1 — Get the code

Follow the following steps:

git clone https://github.com/DataScienceUIBK/Context-Convergence-Inferential-QA.git
cd Context-Convergence-Inferential-QA
pip install termcolor

You do not need to recreate the dataset for the experiments.

Step 2 — Run the experiments

python experiments/convergence_vs_cosine.py
python experiments/order.py

This reproduces the main experimental setup using the prepared data.

⚠️ Option B — Rebuild the dataset from scratch

Use this only if you want to regenerate the dataset yourself.

Step 1 — Complete setup first

Before rebuilding anything, finish all the following steps:

Make sure HintEval is installed correctly: 👉 https://hinteval.readthedocs.io/

Then:

git clone https://github.com/DataScienceUIBK/Context-Convergence-Inferential-QA.git
cd Context-Convergence-Inferential-QA
pip install -r requirements.txt

Step 2 — Go to the dataset directory

cd dataset

Step 3 — Run the full dataset pipeline

bash make.sh

This rebuilds the dataset step by step using the scripts in dataset/.

Step 4 — Return to the repository root

cd ..

Step 5 — Run the experiments on the rebuilt data

python experiments/convergence_vs_cosine.py
python experiments/order.py

📌 Notes for exact reproduction

• Use the provided dataset if you want the closest match to the reported results.

• Rebuilding the dataset is mainly for transparency and regeneration.

• Make sure HintEval is installed correctly before rebuilding.

• The experiments in this repository correspond to the two main studies in the paper:

  • convergence vs cosine similarity
  • sentence ordering by convergence

🧠 Key Findings

• 🟢 Convergence is a strong relevance signal
• 📈 High-convergence passages → better accuracy
• ❌ Cosine similarity is not reliable
• 🔝 Ordering by convergence improves performance
• 🧭 LLMs prioritize earlier information

📚 Citation

📜 License

MIT License — see LICENSE.