GraphER: A Structure-aware Text-to-Graph Model for Entity and Relation Extraction

TL;DR

GraphER reframes joint entity and relation extraction as Graph Structure Learning, starting from an initial, imperfect text-derived graph and iteratively refining it through a Graph Token Transformer. It uses precise node/edge selection to form an initial scaffold, followed by a single-shot graph editing stage that yields a final, structure-aware IE graph, with nodes labeled for entities and edges labeled for relations. Training optimizes a multitask loss that combines node/edge selection, editing, and final classification losses, achieving strong results on ACE 05, CoNLL-2004, and SciERC, often outperforming MPGNN baselines. The approach demonstrates that explicit graph structure learning and global attention over graph tokens can better handle noisy, heterogeneous graphs, enabling robust joint IE with practical implications for knowledge graph construction and NLP pipelines.

Abstract

Information extraction (IE) is an important task in Natural Language Processing (NLP), involving the extraction of named entities and their relationships from unstructured text. In this paper, we propose a novel approach to this task by formulating it as graph structure learning (GSL). By formulating IE as GSL, we enhance the model's ability to dynamically refine and optimize the graph structure during the extraction process. This formulation allows for better interaction and structure-informed decisions for entity and relation prediction, in contrast to previous models that have separate or untied predictions for these tasks. When compared against state-of-the-art baselines on joint entity and relation extraction benchmarks, our model, GraphER, achieves competitive results.

Figures (5)

• Figure 1: GraphER architecture. Please refer to the text box in the figure for an explanation of the different steps.
• Figure 2: Graph transformer: GraphER first constructs an initial graph from the input text (b) (Section \ref{['graph_con']}). A transformer then processes nodes and edges of the graphs to refine their representation (d) (Section \ref{['trans_tok']}). Edge representation only uses node and type identifiers (without using edge-specific features) to enforce the transformer to use the graph structure for representation computation.
• Figure 3: Attention visualization in Graph Transformer. In this figure, we illustrate the attention for the input "Stephen Curry plays for the Warriors in San Francisco". The nodes are represented by their identifier and text spans and edges are represented by their corresponding (directed) pairs of node identifier. We illustrate only the top 3 nodes and top 3 edge for better visibility.
• Figure 4: Common Errors in Entity Recognition. This caption highlights prevalent model errors found within false positives.
• Figure 5: Entity and Relation confusion matrix.