Graph4NLP: An Open‑Source Graph Neural Network Library for Natural Language Processing

Graph4NLP, launched in June 2020 and publicly released in June 2021 (v0.4.1), is the first open‑source software package dedicated to applying Graph Neural Networks (GNNs) in Natural Language Processing (NLP). The project is a joint effort by researchers from Meta, IBM Research, Pinterest, and several universities.

The library’s architecture consists of four layers: a data layer (handling raw graph data and datasets), a module layer (graph construction, GNN models, prediction, loss, and evaluation), a model layer, and an application layer that supports various NLP tasks such as text classification, semantic parsing, machine translation, knowledge base expansion, and natural language generation.

Key features include flexible and easy‑to‑use high‑level APIs, rich learning resources (code examples, tutorials, videos, and survey papers), high efficiency and scalability thanks to the underlying DGL framework, and abundant code examples for different NLP scenarios.

Graph construction supports both static methods (e.g., dependency trees, AMR graphs) and dynamic, learnable approaches, allowing users to combine prior knowledge with data‑driven graph learning. The embedding module offers unified interfaces for single‑token and multi‑token representations using word2vec, BERT, BiLSTM, etc.

The learning module incorporates a variety of GNN algorithms, including directed‑edge variants (bi‑fuse, bi‑sep) that often outperform their undirected counterparts. Prediction APIs cover both classification and generation tasks, providing mechanisms such as attention, copy, and coverage, as well as higher‑level Graph2Seq and Graph2Tree interfaces.

Dataset APIs adapt to different input‑output formats required by NLP tasks, and the inference module enables straightforward deployment of trained models in online services.

Two demo applications—text classification and math word‑problem solving—illustrate how to assemble the pipeline using the library’s modular APIs.

Future plans aim to improve scalability (multi‑GPU, multi‑node training), expand customizable NLP tasks, add more pre‑trained language model interfaces, and incorporate additional GNN models.