📚 Using Shapley Interactions to Understand How Models Use Structure

This repository provides code and scripts to reproduce experiments from the paper:

“Using Shapley Interactions to Understand How Models Use Structure”
Divyansh Singhvi, Diganta Misra, Andrej Erkelens, Raghav Jain, Isabel Papadimitriou, Naomi Saphra
arXiv:2403.13106v2

---

📝 Overview

Modern language and speech models learn rich hidden structures about syntax, semantics, and phonetics.

This repository shows how to use the Shapley-Taylor Interaction Index (STII) to quantify pairwise interactions:

• Text models: How do pairs of tokens interact beyond their individual effects?
• Speech models: How do acoustic frames interact near phoneme boundaries?

By doing so, you can test:

• How well models encode syntactic tree structures
• How they handle multiword expressions
• How speech models reflect phonetic coarticulation

---

📂 Repository Structure

---

🧮 How It Works

✅ STII for Text (ExperimentRunner):

• Load tagged sentences with multiword expressions (MWEs) and syntactic trees
• For token pairs:
  • Compute logits for 4 contexts: AB, A, B, φ (none)
  • Interaction = (AB - A - B + φ) and normalize by (φ) norms
• Analyze how interaction varies with:
  • Linear distance
  • Syntactic distance
  • Whether tokens belong to a strong or weak MWE

✅ STII for Speech (SpeechSTIIExperimentRunner):

• Load audio and phoneme time alignments
• Mask 20ms waveform slices to simulate ablations
• Compare interaction:
  • Consonant-vowel vs consonant-consonant
  • By manner of articulation (how vowel-like a consonant is)
  • The methodology is same for both Speech and Text

---

🚀 How to Run

1️⃣ Install Environment

# Using conda (recommended)
conda env create -f conda.yaml
conda activate shapley_llm