Source-linked AI summary
GraphIE: A Graph-Based Framework for Information Extraction
Yujie Qian, Enrico Santus, Zhijing Jin, Jiang Guo, Regina Barzilay
TL;DR
Sequential IE taggers capture local context but have limited ability to use non-local dependencies that can reduce tagging ambiguity. GraphIE propagates information across task-specific graphs, and it consistently outperforms SeqIE across textual, social-media, and visual information extraction.
Problem
Sequential IE taggers model local context but have limited ability to exploit non-local and non-sequential dependencies that can help reduce tagging ambiguity.
Source: §1 · p. 1Method
GraphIE combines recurrent local-context encoding with graph convolutions that propagate information between connected textual units, then feeds learned representations to a decoder for word-level tagging.
Source: §1 · p. 1Results
Takeaways & Limitations
Simple task-specific graphs can define input structure for IE without major processing or external resources across the evaluated tasks.
Source: §1 · p. 2Further non-local relations such as coreference may improve results, but require additional resources and are left for future work.
Source: §5.2 · p. 6
Abstract
from arXiv · showhide
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.
1 Introduction
Sequential taggers capture local context but have limited ability to use non-local dependencies that can reduce tagging ambiguity. GraphIE integrates graph-based context into sequence tagging and outperforms SeqIE across three information-extraction tasks.
1 Introduction
- Sequential taggers have limited access to non-local dependencies, while output-space constraints and engineered input features do not readily capture the richer structural relations among inputs.Co-reference links can help resolve an ambiguous mention even when the linked expression is not itself labeled as an entity.Source: §1 · p. 1
- GraphIE encodes local context recurrently, propagates information between graph neighbors with graph convolutions, and returns learned representations to a decoder for word-level tagging.Graph nodes represent words or sentences, and edges describe their relations.Source: §1 · p. 2
- GraphIE outperforms SeqIE by 0.5% on CONLL03 and 1.4% on chemical entity extraction, improves EDU-CATION extraction by 3.7%, and beats the visual baseline by 1.2%.The reported improvements span textual, social-media, and visual information extraction.Source: §1 · p. 2
2 Related Work
Prior work has incorporated non-local dependencies through output constraints or input-space features, while graph neural networks offer a way to learn from graph structure. GraphIE adapts graph convolutions as an intermediate context module for sequence tagging and extends graph-based IE to sentence-level graphs and varied relations.
2 Related Work
- Prior methods often constrain the output space, while input-space feature approaches encode non-local dependencies but require challenging, time-consuming feature design and offer limited modeling of graph-informed node interactions.The cited discussion notes that co-reference dependencies may not be readily captured by output constraints when linked expressions are not entities.Source: §2 · p. 2
- GraphIE adapts graph convolutional networks from node classification as an intermediate module, projecting learned non-local context to a decoder for sequence tagging.This differs from using graph convolutional representations directly for classification.Source: §2 · p. 2
- Unlike prior graph neural IE studies described here, GraphIE supports word- or sentence-level graphs and tests relations including coreference, followed-by links, and document layout.The passage contrasts earlier improvements using dependency trees with GraphIE’s evaluation on other relation types and tasks.Source: §2 · p. 2
3 Problem Definition
The paper formulates information extraction as sequence tagging while assuming that input data also has graph structure encoding dependencies between words or sentences. It defines sentence- and word-level graph variants for representing those dependencies.
3 Problem Definition
- The formulation treats information extraction as sequence tagging and assumes a graph can capture non-local, non-sequential dependencies between words or sentences.The input comprises sentences and an auxiliary graph with nodes and edges.Source: §3 · pp. 2–3
- A sentence-level graph uses sentences as nodes, whereas a word-level graph uses words as nodes and can connect co-referent tokens.Sentence-level edges encode sentence dependencies; word-level edges connect pairs of words.Source: §3 · p. 3
- Graph edges may be directed or undirected, and multiple edge types can represent distinct structural factors in task-specific input data.Source: §3 · p. 3
- The paper uses the BIO (Begin, Inside, Outside) tagging scheme for word-level labels.Source: §3 · p. 3
4 Method
GraphIE combines a recurrent encoder, a graph module, and a decoder to learn local and non-local dependencies over words or sentences. Graph convolutions propagate information between connected nodes to enrich representations for tagging.
4 Method
- GraphIE jointly learns local and non-local dependencies through graph propagation, using an encoder, graph module, and decoder for labeling.Source: §4 · p. 3
4.1 Encoder
- The encoder concatenates each word embedding with a character-level CNN feature, encodes each sentence independently with a bidirectional LSTM, and averages word states for sentence representations.Source: §4.1 · p. 4
4.2 Graph Module
- The GCN aggregates neighbor information, normalizes by node degree, stacks layers to reach more distant nodes, and assigns different weights to edge types or directions.Source: §4.2 · p. 4
4.3 Decoder
- Graph-contextualized sentence vectors initialize the decoder’s forward and backward LSTMs; a CRF models tag transitions, and input concatenation performs worse empirically.Source: §4.3 · p. 5
4.4 Adaptation to Word-level Graphs
- For word-level graphs, GCN outputs feed directly to the decoder, whose initial states are zero; the CRF remains unchanged and models sequential tag connections.Source: §4.4 · p. 5
5 Experimental Setup
The experimental setup evaluates GraphIE on three information-extraction tasks using simple, task-specific graph topologies designed to capture each input's structure.
5 Experimental Setup
- The evaluation uses simple, task-specific graph topologies across textual, social-media, and visual information extraction.The topologies are designed to capture each task's underlying input structure without major processing.Source: §5 · p. 5
5.1 Task 1: Textual Information Extraction
- Discourse-level NER uses CONLL-2003 and CHEMDNER data, with word nodes linked to neighboring words and repeated non-stop-word tokens.The setup treats repeated mentions as a soft consistency cue across a discourse.Source: §5.1 · pp. 5–6
5.2 Task 2: Social Media Information Extraction
- The social-media task extracts users' education and job organizations from tweets, using user ego-networks linked by follow relationships.EDUCATION contains 443,476 tweets from 7,208 users; JOB contains 176,043 tweets from 1,772 users.Source: §5.2 · p. 6
5.3 Task 3: Visual Information Extraction
- Visual extraction targets eight attributes in variable-layout adverse-event reports, representing each page as a graph of nearby text boxes.The corpus contains 25,200 reports averaging nine pages; box connections use horizontal or vertical overlap exceeding 50%.Source: §5.3 · pp. 6–7
5.4 Baseline and Our Method
- SeqIE is a two-layer BiLSTM-CRF baseline; for visual IE, horizontally aligned text boxes are concatenated to strengthen the sequential baseline.GraphIE shares the baseline's encoder and decoder, adding a graph module for the relevant task-level graph.Source: §5.4 · p. 7
5.5 Implementation Details
- Training uses Adam with a 0.001 learning rate, dropout of 0.1, and development-set early stopping; GCN depth varies by task.Task 1 uses a two-layer GCN, while Tasks 2 and 3 use one layer; Task 3 also adds positional encodings from bounding-box coordinates.Source: §5.5 · p. 7
6 Results
GraphIE outperforms the sequential baseline across textual, social-media, and visual information extraction, with gains varying by dataset and visual layout.
6.1 Task 1: Textual Information Extraction
- GraphIE significantly outperforms SeqIE on CONLL03 and improves CHEMDNER performance by 1.4% over the feature-based method.It achieves the best reported performance among methods not using ELMo or BERT.Source: §6.1 · pp. 7–8
- On CONLL03, random connections lower GraphIE F1 from 95.12% to 94.29%, while trivial neighbor-feature augmentation also performs poorly.The analysis attributes the random-graph drop to the benefit of task-specific graph structure and identifies learned GCN embeddings as necessary.Source: §6.1 · p. 8, Figure 3 · p. 8
- GraphIE reduces CONLL03 tagging errors from 1.78% to 1.67%, with most of the reduction among ambiguous words.Ambiguous-word errors fall from 1.16% to 1.06%, while unambiguous-word errors change from 0.62% to 0.61%.Source: §6.1 · p. 8
6.2 Task 2: Social Media Information Extraction
- GraphIE outperforms SeqIE on both EDUCATION and JOB, with larger improvements on EDUCATION (3.7% versus 0.3%).The authors relate the difference to higher neighbor affinity in EDUCATION (74.3 versus 14.5 for JOB).Source: §6.2 · p. 8
6.3 Task 3: Visual Information Extraction
- GraphIE outperforms SeqIE on most visual-extraction attributes and improves micro-average F1 by 1.2%.The GraphIE–SeqIE gap ranges from −0.5% for Patient Birthday to 2.4% for Event.Source: §6.3 · pp. 8–9
- In the visual-task ablation, removing separate edge-type weights, horizontal edges, vertical edges, and CRF corresponds to contributions of +0.8%, +3.1%, +5.4%, and +5.7%, respectively.These are micro-average F1 results on the development set.Source: §6.3 · p. 9, Table 7 · p. 9
- On visual extraction, GraphIE scores 2.8% higher than SeqIE on seen templates, and the gap reaches 20.3% on unseen templates.The unseen-template evaluation uses two new template types.Source: §6.3 · p. 9, Figure 4 · p. 9
7 Conclusions
GraphIE learns local and non-local contextual representations from task-specific graph structures, and its evaluation reports consistent gains over SeqIE across textual, social media, and visual information extraction. Future work includes automatically learning the input graph structure.
7 Conclusions
- GraphIE jointly models textual-unit representations and their dependencies, propagating information through neighboring graph nodes to support word-level tagging.The framework uses task-specific graph topologies to represent the underlying input structure.Source: §7 · p. 9
- GraphIE consistently enhances accuracy and outperforms the competitive SeqIE baseline across textual, social media, and visual information extraction.The evaluation reports that it efficiently models non-local and non-sequential context.Source: §7 · p. 9
- Future work includes exploring automatic learning of the input data’s underlying graphical structure.Source: §7 · p. 9
A.1 Social Media Information Extraction
The mock-up depicts the Social Media Information Extraction task using users as nodes and follow-by relations as edges.
A.1 Social Media Information Extraction
- In the Social Media Information Extraction mock-up, users are nodes and follow-by relations are edges.Source: Figure 1 · p. 12
A.2 Visual Information Extraction
The mock-up shows two forms with different layouts and graphical dependencies linking text in blue bounding boxes.
A.2 Visual Information Extraction
- The visual IE mock-up shows graphical dependencies as green lines connecting text in blue bounding boxes across two differently laid-out forms.Source: Figure 2 · p. 12