Causal Abstractions of Neural Natural Language Inference Models

This is the implementation for the experiments in the paper Causal Abstractions of Neural Natural Language Inference Models.

Setup and dependencies

See requirements.txt.

Repository Structure

Code folders

• intervention/ Basic infrastructure for defining computation graphs and performing interventions.
• compgraphs/ Computation graphs for Natural Language Inference causal models and neural models.
• causal_abstraction/ Interchange experiments and analysis.
• datasets/ Class definitions for datasets.
• modeling/ Neural models for NLI and training code.
• probing/ Probing experiments.
• experiment/ Utilities for launching experiments and automatically recording experiment results in databases.
• feature_importance/ Utilities for integrated gradients experiments.

Scripts

Training models

• train_bert.py and train_lstm.py. Train one instance of a model.
• train_manager.py. Utilities for interfacing with the experiment module and managing grid search training.

Interchange experiments

• interchange.py. Run one set of an interchange experiment on a given causal model intermediate node, and all neural model locations for that node. Analyze the success rates of interventions.
• graph_analysis.py. Composes the graph linking the examples after interchange experiments and finds cliques.
• interchange_manager.py. Utilities for interfacing with the experiment module and run large batches of interchange experiments on a computing cluster.

Probing experiments

• probe.py
• probe_manager.py