Abstract
Entity alignment (EA), critical for knowledge graph (KG) integration, identifies equivalent entities across different KGs. Traditional methods often face challenges in semantic understanding and scalability. The rise of language models (LMs), particularly large language models (LLMs), has provided powerful new strategies. This paper systematically reviews LM-driven EA methods, proposing a novel taxonomy that categorizes methods in three key stages: data preparation, feature embedding, and alignment. We further summarize key benchmarks, evaluation metrics, and discuss future directions. This paper aims to provide researchers and practitioners with a clear and comprehensive understanding of how language models reshape the field of entity alignment.
Introduction: The Challenge of Information Silos
Knowledge Graphs (KGs) such as DBpedia, YAGO, and Freebase serve as the backbone for fact-based AI. However, these graphs are frequently constructed in isolation, leading to "information silos." A single real-world entity, like the city of "Rome," might exist in one KG with geographical data and in another with cultural or culinary details.
Entity Alignment (EA) is the process of identifying these equivalent entities across disparate KGs to create a unified knowledge base. Traditional EA methods relied heavily on Knowledge Representation Learning (KRL), using structural patterns through models like TransE or Graph Neural Networks (GNNs). While effective in some contexts, these structural approaches often struggle with sparse graphs, noisy data, or cross-lingual differences.
The emergence of Language Models (LMs) and Large Language Models (LLMs) has introduced a paradigm shift. By leveraging the vast "world knowledge" and deep semantic understanding inherent in these models, researchers can now overcome the limitations of purely structural methods.
Methodology: A Three-Phase Taxonomy
The survey proposes a systematic taxonomy for LM-driven EA, dividing the process into three distinct phases: Data Preparation, Feature Embedding, and Alignment.
1. Data Preparation Phase
In this initial stage, LMs are employed to refine the input data or generate "pseudo-seed pairs." Because EA typically requires labeled examples for training, LMs can automatically identify high-confidence matches, effectively expanding training sets without manual human intervention. This phase also includes graph filtering and unification to ensure the KGs are in a compatible format.
2. Feature Embedding Phase
This phase focuses on representing entities in a high-dimensional space where equivalent entities are close to each other.
- Semantic Features: Models like BERT or RoBERTa encode entity names and descriptions into dense vectors.
- Attribute Features: Specific traits (e.g., population, founding date) are directly encoded to provide additional context.
- Verbalization: A significant innovation where graph structures are "translated" into natural language sentences (e.g., "Rome is the capital of Italy"). This allows LMs to process structural information as text, utilizing their pre-trained linguistic capabilities.
3. Alignment Phase
The final stage involves the actual matching of entities.
- Similarity-based Alignment: This traditional approach calculates the distance (e.g., cosine similarity) between the generated feature vectors.
- Reasoning-based Alignment: Modern LLMs (like GPT-4) are used as reasoning engines. Through techniques like Chain-of-Thought prompting, the system asks the LLM to decide if two entities are identical based on the provided textual context and attributes.
Benchmarks and Performance
The survey evaluates these methods across several standard benchmarks, including DBP15K (cross-lingual), SRPRS (degree distribution), and large-scale sets like DBP1M.
Recent LM-driven methods have achieved near-perfect accuracy on these benchmarks. For instance, on the DBP15K (ZH-EN) subset, models like FGWEA and DERA achieve Hits@1 scores exceeding 0.98. Furthermore, reasoning-based models like Seg-Align demonstrate remarkable robustness in cross-lingual tasks without requiring specific cross-lingual training data, signaling that many traditional benchmarks are nearing saturation.
Key Insights and Future Directions
The shift from structure-only to semantic-heavy alignment marks a turning point for the field. LMs provide a "semantic anchor" that allows for successful alignment even when the surrounding graph structures differ significantly. However, several challenges remain for future research:
- Multimodal EA: Integrating images, audio, and video associated with entities.
- Efficiency and Scalability: Reducing the computational cost of LLM-based reasoning for billion-node graphs, perhaps through Parameter-Efficient Fine-Tuning (PEFT) like LoRA.
- Dynamic and Continual Alignment: Developing methods that can update links in real-time as KGs evolve without retraining from scratch.
- Privacy-Preserving Alignment: Merging data sources while protecting sensitive information, a critical but under-explored area.
This survey underscores that while LMs have "solved" many traditional EA problems, they have also opened new frontiers in how we integrate and reason over global knowledge.
Summary updated