Abstract
Jointly extracting entities and their relations is a fundamental task in Information Extraction. Traditional pipeline methods often suffer from error propagation, where mistakes in entity recognition negatively impact relation extraction. This paper introduces a table-filling framework that treats entity and relation extraction as a single structured learning problem. By representing sentences in a 2D table and employing history-based search, the model effectively captures the mutual dependencies between entities and relations. Experimental results on the CoNLL 2004 dataset demonstrate that this joint approach significantly outperforms traditional pipeline models.
Introduction
Information Extraction (IE) typically involves two sub-tasks: Named Entity Recognition (NER) and Relation Extraction (RE). Traditionally, these are handled in a pipeline where entities are identified first, followed by relation classification. However, this approach is prone to error propagation—a mistake in entity typing can derail subsequent relation extraction. Furthermore, entities and relations are inherently linked; for instance, a "Live_in" relation implies the presence of a Person and a Location. This paper addresses these issues by modeling both tasks simultaneously using a unified table representation.
Methodology: The Entity and Relation Table
The central innovation of this work is the transformation of joint extraction into a table-filling task. For a sentence of $n$ words, the model constructs an $n \times n$ table (specifically focusing on the lower triangular portion).
- Diagonal Cells: These cells store entity labels for individual words using the BILOU (Begin, Inside, Last, Outside, Unit) scheme.
- Off-Diagonal Cells: These cells represent the relations between word pairs.
- Structured Learning: The model fills the table cell-by-cell. Because the label of a current cell depends on previously assigned labels (the "history"), the authors use history-based structured learning. Global features act as a bridge, ensuring that assigned relations are consistent with the identified entity types.
Search Orders and Decoding
Since the order of filling the table affects the context available for each decision, the authors explored several search strategies to determine the optimal sequence for label assignment.
The study found that the "Close-first" order—prioritizing relations between nearby words—yielded the best results. This suggests that local dependencies provide a more reliable foundation for resolving longer-distance, more ambiguous relations. To find the optimal global assignment, the model employs beam search, allowing it to explore multiple labeling paths before finalizing the extraction.
Experimental Results
The model was evaluated on the CoNLL 2004 corpus, which includes entity types like Location and Person, and relations such as "Live_in" and "Work_For." Unlike many previous studies, the authors performed the full task of detecting entity boundaries and types from scratch.
The joint model achieved an F1-score of 0.610, significantly outperforming the pipeline baseline of 0.577. While the joint approach drastically improved relation extraction, its impact on isolated entity recognition was more modest, suggesting that relation evidence is primarily useful for disambiguating the connections between entities rather than the entity types themselves. Additionally, the DCD-SSVM (Dual Coordinate Descent for Structural SVM) optimizer proved more robust than the standard Perceptron in this structured setting.
Conclusion
By reimagining joint extraction as a table-filling problem, this research provides a flexible framework for capturing complex linguistic dependencies. The results highlight that the order of structured search is a critical hyperparameter in joint modeling. Future work includes applying this representation to other structured NLP tasks, such as semantic parsing and coreference resolution, and exploring more sophisticated dynamic search policies.
Summary updated