Roadmap

EntityAlignment-Nexus is built to grow into the reference hub for entity-alignment models. The current release covers nine classic structural, relation-aware, attribute- and name-based methods. The next wave is transformer-based EA.

Vision

%%{init: {'theme':'base','themeVariables':{'fontSize':'14px','fontFamily':'Inter, sans-serif','lineColor':'#7d8590','primaryTextColor':'#e6edf3'}}}%%
flowchart LR
    subgraph PH1["Phase 1 - shipped (9 models)"]
        direction TB
        P1["NAEA · BootEA · AliNet · GCN-Align<br/>KECG · JAPE · DGMC · MRAEA · RREA"]:::done
    end
    subgraph PH2["Phase 2 - next"]
        direction TB
        P2["Transformer encoders<br/>PLM-initialised EA<br/>Dangling-aware EA"]:::next
    end
    subgraph PH3["Phase 3 - later"]
        direction TB
        P3["Multi-modal EA<br/>LLM-assisted EA<br/>Unsupervised / zero-seed EA"]:::later
    end
    PH1 ==> PH2 ==> PH3
    style PH1 fill:#0d1117,stroke:#3fb950,color:#e6edf3
    style PH2 fill:#0d1117,stroke:#58a6ff,color:#e6edf3
    style PH3 fill:#0d1117,stroke:#a371f7,color:#e6edf3
    classDef done  fill:#14532d,stroke:#3fb950,stroke-width:2px,color:#dcfce7;
    classDef next  fill:#0c2d6b,stroke:#58a6ff,stroke-width:2px,color:#dbeafe;
    classDef later fill:#3b0764,stroke:#a371f7,stroke-width:2px,color:#f3e8ff;

Planned: transformer-based EA

• Self-attention encoders

  ---

  Graph-transformer aggregators that replace fixed-hop GAT with global attention over the neighbourhood (e.g. relation-aware transformer layers on the KG).

• PLM-initialised alignment

  ---

  Initialise entity features with pre-trained multilingual language models (mBERT, XLM-R, LaBSE) instead of GloVe, in the spirit of BERT-INT / SelfKG.

• Dangling-aware EA

  ---

  Handle entities with no counterpart (the DBP2.0 / dangling setting), a more realistic open-world variant of the task.

• LLM-assisted EA

  ---

  Use large language models as candidate re-rankers or verifiers on top of a cheap structural retriever.

Candidate models (shortlist)

Model	Venue	Why it fits
BERT-INT	IJCAI 2020	BERT-based interaction model over names/descriptions/attributes
SelfKG	WWW 2022	self-supervised, (almost) no seed alignments
Dual-AMN	WWW 2021	proxy-attention, very fast and strong on DBP15K
TransEdge	ISWC 2019	edge-centric translational embeddings
EVA / MMEA	AAAI 2021	multi-modal (images + structure + attributes)

These are candidates, not commitments - priority follows community interest and reproducibility.

How to propose or contribute a model

1. Open an issue describing the method and its DBP15K numbers.
2. Add code/src/models/<your_model>.py (encoder + loss) following the existing pattern.
3. Add a trainer (or reuse one) in code/src/trainer.py and a configs/<your_model>.yaml.
4. Add a self-contained notebook and a docs page mirroring the others.

The shared data.py / metrics.py mean you mostly write the model itself. See About & contributing for the conventions.

Want a specific model next?

Tell us which transformer-based EA model you want first - community demand drives the order.