Source-linked AI summary
Cross-lingual Knowledge Graph Alignment via Graph Matching Neural Network
Kun Xu, Liwei Wang, Mo Yu, Yansong Feng, Yan Song, Zhiguo Wang, Dong Yu
TL;DR
Cross-lingual KG alignment can fail when entities have different facts or sparse structural context. This paper introduces topic entity graphs and an attentive graph-matching model that jointly encodes entitywise matches into graph-level vectors. Experiments show significant improvements over existing baselines on benchmark datasets.
Problem
Existing alignment methods mainly rely on monolingual structural entity embeddings, which may fail when corresponding entities have different facts or insufficient structural information.
Method
The paper represents each entity’s local KG context as a topic entity graph and uses attentive graph matching to combine entitywise matching information into a graph-level vector.
Results
Experimental results on benchmark datasets show that the model significantly outperforms existing baselines.
Takeaways & Limitations
Graph-level matching improves handling of entities whose neighboring facts differ across the two knowledge graphs.
Takeaways & Limitations
Adding relation labels as distinct graph nodes hurt both performance and efficiency, so the model retains relation direction while omitting relation labels.
Abstract
from arXiv · showhide
Previous cross-lingual knowledge graph (KG) alignment studies rely on entity embeddings derived only from monolingual KG structural information, which may fail at matching entities that have different facts in two KGs. In this paper, we introduce the topic entity graph, a local sub-graph of an entity, to represent entities with their contextual information in KG. From this view, the KB-alignment task can be formulated as a graph matching problem; and we further propose a graph-attention based solution, which first matches all entities in two topic entity graphs, and then jointly model the local matching information to derive a graph-level matching vector. Experiments show that our model outperforms previous state-of-the-art methods by a large margin.
1 Introduction
Cross-lingual KG alignment matches entities across multilingual graphs, but monolingual structural embeddings can miss entities with differing facts or sparse neighborhoods. The paper addresses this by representing local KG context as topic entity graphs and matching those graphs jointly.
- Cross-lingual KG alignment automatically matches entities across multilingual knowledge graphs that lack cross-lingual links.
- Monolingual structural embeddings can diverge across languages when entities have different facts, making entity matching difficult.
- Sparse structural information and omitted surface forms make entities with few neighbors especially difficult to match.
- The paper represents an entity’s KG context with a topic entity graph and formulates alignment as graph matching.
- The proposed model matches entities across two topic graphs, jointly propagates local matching information, and produces graph-level matching vectors.
2 Topic Entity Graph
A topic entity graph represents a topic entity together with its local KG context and directed relations among the topic entity’s one-hop neighbors.
- A topic entity graph represents relations among a given topic entity and its neighboring entities in a knowledge base.
- The graph nodes are the topic entity’s collected 1-hop neighbors, denoted {e1, ..., en}.
- For each directly related entity pair, the construction adds a directed edge corresponding to their KG relation.
3 Graph Matching Model
The graph matching model encodes two topic graphs, computes bidirectional attentive entity matches, propagates those local states, and predicts whether the graphs describe the same entity.
- The model retrieves topic graphs G1 and G2, then estimates the probability that they describe the same entity.
- Input Representation Layer: GCN1 learns embeddings for entities occurring in each topic graph by aggregating incoming and outgoing neighbor representations.
- Node-Level (Local) Matching Layer: The local matching layer compares every entity in one graph with all entities in the other graph in both directions.
- Node-Level (Local) Matching Layer: Cosine similarities weight an attentive vector formed by summing the other graph’s entity embeddings.
- Node-Level (Local) Matching Layer: A multi-perspective cosine function produces matching vectors for entities in both topic graphs.
- Graph-Level (Global) Matching Layer: GCN2 propagates local matching information across the graph, after which max and mean pooling create a fixed-length graph representation for prediction.
4 Experiments
Experiments evaluate the model on DBP15K and compare surface-form, structural, and graph-matching approaches. Results show benefits from combining contextual matching, with best performance at GCN2 hop size 3 and limitations from modeling relation labels.
- The model is evaluated on DBP15K datasets linking Chinese, Japanese, and French DBpedia entities to English, with 15,000 interlanguage links per dataset.The experiments use the same train/test split as previous works.
- The BASELINE significantly outperforms previous structure-focused methods without KG structural information, indicating that entity surface form is important for alignment.Its entity embedding is the sum of word embeddings in the entity surface form.
- NodeMatching improves over the BASELINE by additionally encoding KG structural information with GCN1.
- The graph matching method significantly outperforms all baselines, suggesting that global topic-entity context improves similarity estimation.The matching layer jointly models entitywise matching information across the topic graphs.
- The model performs best when GCN2 reaches hop size λ = 3 and better handles entities whose neighboring facts differ across the two KGs.The graph matching layer propagates the most relevant local matching information throughout the graph.
- Adding relation labels as distinct graph nodes hurts performance and efficiency.The paper attributes this to abstract relation symbols and larger topic entity graphs requiring bigger hop sizes and longer running time.
5 Conclusions
The paper addresses limitations of monolingual-structure-based entity embeddings by representing local KG context as topic entity graphs and aligning them through graph matching. Experiments on benchmark datasets show that the proposed model significantly outperforms existing baselines.
- Previous methods may fail when equivalent entities have different facts across languages because they mainly rely on monolingual KG structural information.
- The proposed topic entity graph represents an entity's contextual KG information, while the graph matching model jointly encodes entitywise matching information into a graph matching vector.
- Experimental results on benchmark datasets show that the model significantly outperforms existing baselines.
A Matching Function fm
The multi-perspective cosine matching function compares two vectors and returns an l-dimensional vector of perspective-specific matching values. Trainable perspective weights emphasize different dimensions of the underlying d-dimensional space.
- fm compares two vectors and returns an l-dimensional matching vector.Each element corresponds to one matching perspective.
- Each matching value mk is the cosine similarity between two vectors weighted according to perspective k.The k-th perspective produces one element of the returned matching vector.
- The trainable matrix W assigns different weights to dimensions of the d-dimensional vector space for each perspective.Wk, the k-th row of W, controls perspective k through element-wise multiplication.