Abstract
Extracting n-ary relations from unstructured text remains a significant challenge in Information Extraction (IE). This paper introduces DARE (Domain Adaptive Relation Extraction), a bootstrapping framework that utilizes a bottom-up, compositional approach to learn complex relation extraction rules from minimal seed data. By leveraging dependency parsing and rule projections, DARE effectively identifies multi-argument relations in both verbal and nominal contexts, achieving high precision with minimal supervision.
Introduction
Traditional IE systems often struggle with two primary issues: linguistic complexity and precise role assignment. Many systems are "verb-centered," failing to capture relations embedded in compound nouns (e.g., "the 2005 Nobel Peace Prize"). Furthermore, distinguishing between similar semantic roles—such as a person entering versus leaving a corporate position—is notoriously difficult for standard pattern-matching approaches. DARE addresses these challenges by employing a "seed-driven" bootstrapping method that learns to compose complex rules from simpler linguistic fragments.
The DARE Architecture
The DARE framework operates through a continuous bootstrapping loop designed to refine its extraction capabilities over time.
The process begins with linguistic annotation, where documents are processed using Named Entity Recognition (NER) and dependency parsing (e.g., MINIPAR). This transforms sentences into structural trees. The system then uses initial "seeds" (known relation instances) to identify relevant patterns within these trees. To ensure quality, DARE employs a Domain Relevance (DR) score, filtering out patterns that lack specificity to the target domain.
Bottom-Up Rule Composition
The core innovation of DARE is its bottom-up, compositional rule representation. Rather than attempting to match a single, exhaustive pattern for an n-ary relation, the system identifies "projections"—smaller k-ary subsets of the target relation.
By learning rules for these smaller components first, DARE can reconstruct complex events even when they are spread across different linguistic structures. For example, it might learn a rule for a [Year] and [Prize Name] pair and another for a [Person] and [Prize Name] pair, eventually synthesizing them into a complete quaternary extraction rule. This approach allows the system to capture relations in both active sentences and complex noun phrases.
Experimental Results
The framework was validated across two distinct domains:
- Nobel Prize Awards: Extracting a 4-argument relation (
<recipient, prize, area, year>). With only a single seed, DARE achieved 71.6% precision and 50.7% recall on a large news corpus. - Management Succession: Extracting a 4-argument relation (
<Person_In, Person_Out, Position, Organisation>) from the MUC-6 dataset. DARE achieved 62% precision and 48% recall using 55 seeds.
A key finding was DARE's ability to handle role precision. It successfully distinguished between Person_In and Person_Out roles, a significant improvement over previous models that often conflated participants in succession events.
Conclusion
DARE demonstrates that a bottom-up approach to relation extraction is both scalable and effective. By moving beyond verb-only patterns and utilizing informational redundancy across documents, the system captures complex relations with minimal supervision. Future work will explore using the Web as a vast, redundant corpus to further enhance pattern acquisition and system robustness.
Summary updated