JAPE¶

TransE + attributes

Cross-lingual Entity Alignment via Joint Attribute-Preserving Embedding Zequn Sun, Wei Hu, Chengkai Li - ISWC 2017 (the paper that introduced DBP15K) Paper | models/jape.py | notebook

Idea in one sentence

Run TransE on a merged-seed graph (aligned seeds share one id, bridging the two KGs) and fuse the structure similarity with a cross-KG TF-IDF attribute similarity at scoring time.

Architecture¶

%%{init: {'theme':'base','themeVariables':{'fontSize':'14px','fontFamily':'Inter, sans-serif','lineColor':'#7d8590','primaryTextColor':'#e6edf3'}}}%%
flowchart LR
    M["merged-seed graph<br/>(aligned seeds share id)"] --> SE["Structure Embedding<br/>TransE ||h+r-t||"]
    A["attribute bags<br/>(sup_attr_pairs merges predicates)"] --> AE["TF-IDF cosine"]
    SE --> FUSE["sim = beta*SE + (1-beta)*AE"]
    AE --> FUSE
    FUSE --> CSLS["CSLS"]
    CSLS --> EVAL["MRR / Hit@k"]
    classDef b fill:#7c2d54,stroke:#f778ba,stroke-width:2px,color:#ffe4f0;
    class M,SE,A,AE,FUSE,CSLS,EVAL b;

Components¶

Structure Embedding (SE). Plain TransE over both KGs in one merged graph; because aligned seeds share ids, TransE propagates the alignment to the test entities.
Attribute Embedding (AE). A cross-KG, shared-vocabulary bag of attribute predicates; sup_attr_pairs merges the zh<->en predicates so the vocabularies overlap. TF-IDF cosine similarity refines SE.
Fusion + CSLS. \(\text{sim} = \beta\,\text{SE} + (1-\beta)\,\text{AE}\) with \(\beta = 0.9\) (SE-dominant), then CSLS on the fused matrix.

Results¶

DBP15K, 30% seed. This repo meets or beats the paper on all three pairs.

Pair	Hit@1 (paper)	Hit@1 (here)	Hit@10 (paper)	Hit@10 (here)	MRR (paper)	MRR (here)
zh_en	0.412	0.425	0.745	0.761	0.490	0.537
ja_en	-	0.368	-	0.735	-	0.490
fr_en	-	0.311	-	0.705	-	0.442

JAPE metrics — Test metrics over training (this repo, zh_en). SE saturates early; SE+AE is tracked.

Debugging lessons

Merged-seed format (use_mtranse_format: false) is essential: SE alone already reaches ~0.31 Hit@1 because the shared ids bridge the two KGs.
AE must be a refiner, not an equal: \(\beta = 0.5\) degrades SE (the TF-IDF and SE cosines are on different scales). \(\beta = 0.9\) realises the attribute gain (~+11 Hit@1).
CSLS on the fused matrix adds another ~+7 Hit@1.
Without merging predicates via sup_attr_pairs, the two attribute vocabularies barely overlap and the AE channel is useless.

References¶

Sun, Hu, Li, JAPE, ISWC 2017.
Bordes et al., TransE, NeurIPS 2013.
Lample et al., CSLS, ICLR 2018.