This repository provides implementations for Multi-Hop Graph Relation Networks (MHGRN) and other graph encoding models for knowledge-aware question answering. It targets researchers and practitioners in NLP and knowledge graph reasoning, offering a framework to integrate external knowledge into QA systems for improved performance.

How It Works

MHGRN leverages graph neural networks to reason over knowledge graphs, specifically ConceptNet, to answer complex questions. It extracts relevant subgraphs for each question-answer pair and encodes relational information through multi-hop reasoning, allowing the model to capture indirect relationships crucial for answering questions that require synthesizing information from multiple sources.

Quick Start & Requirements

• Install: Create a conda environment and install dependencies using provided commands.
• Prerequisites: Python >= 3.6, PyTorch == 1.1.0, CUDA 10.0, transformers == 2.0.0, dgl-cu100 == 0.3.1, networkx == 2.3, spacy == 2.1.6.
• Setup: Download data and preprocess using provided scripts. Preprocessing takes ~3 hours on a 40-core CPU.
• Resources: ConceptNet, download.sh

Highlighted Details

• Supports multiple text encoders (LSTM, GPT, BERT, XLNet, RoBERTa) and graph encoding models (RelationNet, R-GCN, KagNet, GConAttn, KVMem, MHGRN).
• Provides preprocessed ConceptNet data and various entity embeddings (TransE, NumberBatch, BERT-based).
• Includes scripts for hyperparameter search and training/evaluation of different models.

Maintenance & Community

The project is associated with EMNLP 2020. No specific community links or active maintenance signals are present in the README.

Licensing & Compatibility

The repository does not explicitly state a license. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The project relies on specific older versions of PyTorch (1.1.0) and transformers (2.0.0), which may pose compatibility challenges with current environments. The setup process involves downloading significant data and a lengthy preprocessing step.