Abstract
Enterprise relation extraction aims to detect pairs of enterprise entities and identify the business relations between them from unstructured or semi-structured text data. This task is crucial for real-world applications such as risk analysis, rating research, and supply chain security. However, previous work has primarily focused on extracting attribute information (e.g., personnel or corporate business) rather than inter-enterprise relations. To address this gap, we introduce CEntRE, a new large-scale dataset constructed from publicly available business news with rigorous human annotation and intelligent data processing. Extensive experiments with six state-of-the-art models demonstrate the significant challenges posed by this paragraph-level dataset.
Introduction
In the domains of finance and risk management, identifying an entity is only the first step; understanding the web of connections between entities is paramount. Questions such as "Are these companies competitors?", "Is one a subsidiary of the other?", or "Do they share a supply chain?" are vital for comprehensive analysis.
Most existing Relation Extraction (RE) research operates at the sentence level, where relationships are explicitly stated within a single line. However, real-world business intelligence is often fragmented across multiple sentences within a paragraph. Before CEntRE, there was a lack of a unified, supervised dataset for Chinese enterprise relations that accounted for this paragraph-level complexity and the specific nuances of business terminology.
Methods
The construction of CEntRE followed a rigorous five-step pipeline designed to ensure high data quality and relevance to the business domain:
- Input Plain Text: Over 20,000 news reports were crawled from major tech and finance websites.
- Relation Vocabulary Matching: A dictionary of 19 specific enterprise relations (e.g., equity participation, supply chain, litigation) was established based on domain expert consultation.
- Machine Auxiliary Filtering: Named Entity Recognition (NER) was used to filter for paragraphs containing at least two enterprise entities, narrowing the focus to potentially relational text.
- Human Annotation: Master’s students performed double-blind annotations. Disagreements were resolved by a third expert annotator to ensure a "gold standard."
- Verification and Standardization: Raw text was converted into structured triplets
<Subject, Relation, Object>, and dependency parsing identified the most relevant mentions for each entity.
The CEntRE Dataset
CEntRE is uniquely focused on the inter-enterprise niche at a paragraph level. The dataset consists of 3,018 paragraphs, 10,894 unique enterprise entities, and 11,018 relational triples. It covers 19 distinct relation types, ranging from common ones like "Subsidiary" to rarer ones like "Manage on behalf of."
A key characteristic of CEntRE is its complexity. It includes "overlapping" entities where a single company might be involved in multiple different relations within the same text. Most importantly, 47.75% of the relational triples in CEntRE span more than one sentence. This "multi-sentence" challenge means that models limited to individual sentence analysis will fail to capture nearly half of the available business intelligence.
Experimental Benchmarking
To evaluate the dataset, the authors benchmarked six state-of-the-art models, categorized into Pipeline models (performing NER then RE) and Joint models (performing both tasks simultaneously). The results highlight a significant performance gap between identifying entities and extracting their relationships.
Key Findings:
- NER vs. RE: While models performed well at identifying company names (NER F1-scores ~90-91%), identifying the actual relationships (RE) proved much harder. The best-performing model, Table-Sequence, achieved an F1-score of only 48.43%.
- Joint Learning Advantage: Joint models (like CasRel and Table-Sequence) consistently outperformed pipeline models. This is because joint models capture the interaction between entity identification and relational roles, preventing error propagation from the first step to the second.
- The Power of Pre-training: Models utilizing BERT significantly outperformed standard LSTM-based models, suggesting that large-scale linguistic pre-training provides essential external knowledge for understanding complex business contexts.
Conclusion
CEntRE reveals a substantial "reasoning gap" in current AI models. While pattern matching suffices for entity recognition, extracting relations from paragraph-level text requires the ability to synthesize information across sentence boundaries. Future work will focus on improving how models capture inter-sentence context and addressing the long-tail distribution of rare business relations to bridge the gap between AI performance and human-level understanding.
Summary updated