[Paper Review] GraphIE: A Graph-Based Framework for Information Extraction

Yujie Qian, Enrico Santus+3 more|arXiv (Cornell University)|Oct 31, 2018

Topic Modeling32 references65 citations

TL;DR

GraphIE introduces a graph convolutional network module over task-specific graphs to inject non-local context into sequence tagging, improving information extraction across textual, social media, and visual domains over a strong SeqIE baseline.

ABSTRACT

Most modern Information Extraction (IE) systems are implemented as sequential taggers and only model local dependencies. Non-local and non-sequential context is, however, a valuable source of information to improve predictions. In this paper, we introduce GraphIE, a framework that operates over a graph representing a broad set of dependencies between textual units (i.e. words or sentences). The algorithm propagates information between connected nodes through graph convolutions, generating a richer representation that can be exploited to improve word-level predictions. Evaluation on three different tasks --- namely textual, social media and visual information extraction --- shows that GraphIE consistently outperforms the state-of-the-art sequence tagging model by a significant margin.

Motivation & Objective

• Motivate information extraction models to leverage non-local and non-sequential context beyond local word sequences.
• Propose GraphIE, a graph-based encoder-decoder framework that propagates information via graph convolutions.
• Demonstrate the effectiveness of GraphIE on textual, social media, and visual IE tasks using task-specific graphs.

Proposed method

• Encoder generates local context representations for textual units (words or sentences) with a BiLSTM and CharCNN features.
• Graph module adapts a Graph Convolutional Network (GCN) to propagate information across graph-structured dependencies between textual units.
• Decoder is a BiLSTM+CRF tagger; the graph outputs initialize the decoder hidden states to inject contextual information.
• Support for both sentence-level graphs and word-level graphs, enabling non-local context to reduce tagging ambiguity.
• Edge types can be modeled with separate weights, and multiple layers of GCN expand the receptive field over the graph.
• Training optimizes CRF objective using Adam with dropout regularization.

Experimental results

Research questions

• RQ1 Can a graph-based module capture non-local dependencies to improve word-level tagging in IE?
• RQ2 Do task-specific graphs (word- or sentence-level) provide inductive biases that improve performance over sequential baselines?
• RQ3 How does GraphIE perform across textual, social media, and visual IE tasks compared to SeqIE baselines?
• RQ4 What is the contribution of graph structure versus simple feature augmentation or random connections to performance gains?
• RQ5 Does GraphIE generalize to unseen input layouts and layouts with non-local dependencies?

Key findings

• GraphIE consistently improves over SeqIE (BiLSTM+CRF) on three IE tasks: textual (CoNLL03 and Chemdner), social media (Education and Job), and visual IE (AECR).
• On CoNLL03, GraphIE achieves 91.74 F1, outperforming SeqIE’s 91.16.
• On Chemdner, GraphIE reaches 89.71 F1, better than SeqIE’s 88.28.
• In social media, GraphIE yields larger gains, notably 3.7% over SeqIE for Education extraction.
• In visual IE, GraphIE improves micro F1 by about 1.2% over SeqIE across attributes.
• Ablation shows edge-type awareness, horizontal/vertical edge contributions, and the CRF layer each meaningfully boost performance.