Skip to content

Latest commit

 

History

History
204 lines (156 loc) · 11.3 KB

README.md

File metadata and controls

204 lines (156 loc) · 11.3 KB
Raw

EAkit

Entity Alignment toolkit (EAkit), a lightweight, easy-to-use and highly extensible PyTorch implementation of many entity alignment algorithms. The algorithm list is from Entity_Alignment_Papers.

Table of Contents

  1. Design
  2. Organization
  3. Usage
    1. Run an implemented model
      1. Semantic Matching Models
      2. GNN-based Models
      3. KE-based Models
      4. Results
    2. Write a new model
  4. Dataset
  5. Reqirements
  6. TODO
  7. Acknowledgement

Design

We sort out the existing entity alignment algorithms and modularizing the composition of them, and then define an abstract structure as 1 Encoder - N Decoder(s), where different modules are regarded as specific implementations of different encoders and decoders, so as to restore the structures of the algorithms.

Framework of EAkit

Organization

./EAkit
├── README.md                           # Doc of EAkit
├── _runs                               # Tensorboard log dir
├── data                                # Datasets. (unzip data.zip)
│   └── DBP15K
├── examples                            # Shell scripts of implemented algorithms
│   ├── Tensorboard.sh                  # Start Tensorboard visualization
│   ├── run_BootEA.sh
│   ├── run_ComplEx.sh
│   ├── run_ConvE.sh
│   ├── run_DistMult.sh
│   ├── run_GCN-Align.sh
│   ├── run_HAKE.sh
│   ├── run_KECG.sh
│   ├── run_MMEA.sh
│   ├── run_MTransE.sh
│   ├── run_NAEA.sh
│   ├── run_RotatE.sh
│   ├── run_TransE.sh
│   ├── run_TransEdge.sh
│   ├── run_TransH.sh
│   └── run_TransR.sh
├── load_data.py                        # Load datasets. (data adapter)
├── models.py                           # Encoders & Decoders
├── run.py                              # Main
├── semi_utils.py                       # Bootstrap strategy
└── utils.py                            # Sampling methods, ...

Usage

Run an implemented model

  1. Start TensorBoard for metrics visualization (run under examples/):
./Tensorboard.sh
  1. Modify and run a script as follow (examples are under examples/):
CUDA_VISIBLE_DEVICES=0 python3 run.py --log gcnalign \
                                    --data_dir "data/DBP15K/zh_en" \
                                    --rate 0.3 \
                                    --epoch 1000 \
                                    --check 10 \
                                    --update 10 \
                                    --train_batch_size -1 \
                                    --encoder "GCN-Align" \
                                    --hiddens "100,100,100" \
                                    --decoder "Align" \
                                    --sampling "N" \
                                    --k "25" \
                                    --margin "1" \
                                    --alpha "1" \
                                    --feat_drop 0.0 \
                                    --lr 0.005 \
                                    --train_dist "euclidean" \
                                    --test_dist "euclidean"

In detail, the following methods are currently implemented:

Semantic Matching Models

Method Encoder Decoder
MTransE from Chen et al. (IJCAI 2017) [sh], [origin] None TransE, MTransE_Align
BootEA from Sun et al. (IJCAI 2018) [sh], [origin] None AlignEA
TransEdge from Sun et al. (ISWC 2019) [sh], [origin] None TransEdge
MMEA from Shi et al. (EMNLP 2019) [sh], [origin] None MMEA

GNN-based Models

Method Encoder Decoder
GCN-Align from Wang et al. (EMNLP 2018) [sh], [origin] GCN-Align Align
NAEA from Zhu et al. (IJCAI 2019) [sh], [origin] NAEA [N_TransE], N_TransE, N_R_Align
KECG from Li et al. (EMNLP 2019) [sh], [origin] KECG TransE, Align

KE-based Models

Method Encoder Decoder
TransE from Bordes et al. (NIPS 2013) [sh], None TransE
TransH from Wang et al. (AAAI 2014) [sh], None TransH
TransR from Lin et al. (AAAI 2015) [sh], None TransR
RotatE from Sun et al. (ICLR 2019) [sh], None RotatE
HAKE from Zhang et al. (AAAI 2020) [sh], None HAKE
DistMult from Yang et al. (ICLR 2015) [sh], None DistMult
ComplEx from Trouillon et al. (ICML 2016) [sh], None ComplEx
ConvE from Dettmers et al. (AAAI 2018) [sh], None ConvE

Results

Results on DBP15K(zh_en, ja_en, fr_en).

Hits@1 Hits@10 MRR Hits@1 Hits@10 MRR Hits@1 Hits@10 MRR
MTransE 0.419 0.753 0.535 0.433 0.773 0.549 0.407 0.751 0.526
BootEA 0.490 0.793 0.593 0.499 0.813 0.605 0.515 0.838 0.623
TransEdge 0.519 0.813 0.621 0.526 0.825 0.632 0.397 0.824 0.543
MMEA 0.405 0.672 0.499 0.397 0.680 0.496 0.442 0.749 0.550
GCN-Align 0.410 0.756 0.527 0.442 0.810 0.566 0.430 0.813 0.557
NAEA 0.323 0.481 0.381 0.311 0.457 0.363 0.307 0.460 0.362
KECG 0.467 0.815 0.586 0.485 0.843 0.605 0.479 0.844 0.602
TransE 0.343 0.634 0.441 0.365 0.710 0.480 0.374 0.735 0.493
TransH 0.436 0.735 0.540 0.450 0.778 0.561 0.485 0.821 0.599
TransR 0.371 0.697 0.481 0.368 0.709 0.484 0.378 0.741 0.497
RotatE 0.423 0.754 0.534 0.448 0.785 0.561 0.439 0.800 0.560
HAKE 0.288 0.588 0.391 0.319 0.607 0.421 0.319 0.638 0.428
DistMult 0.180 0.400 0.255 0.058 0.179 0.099 0.095 0.285 0.157
ComplEx 0.115 0.265 0.166 0.063 0.251 0.146 0.141 0.332 0.206
ConvE 0.210 0.466 0.299 0.339 0.556 0.415 0.350 0.602 0.439

Write a new model

  1. Divide the algorithm at the abstract level to obtain the structure of 1 (or 0) Encoder and 1 (or more) Decoder(s).
  2. Register the modules and add extra parameters in the top-level encoder (class Encoder) and top-level decoder (class Decoder) in models.py.
  3. Implement the concrete encoding module (class Encoder_Instance) and decoding module(s) (class Decoder_Instance) according to the given template.
  4. Write an execution script (XXX.sh) with parameter settings to run the new model.
  5. (Adapt a new dataset in load_data.py, and add a new sampling strategy in utils.py.)

Example of writing a new model

Dataset

(Currently, EAkit only supports DBP15K, but it is easy to adapt to other datasets.)

  • DBP15K is from the "mapping" folder of JAPE(But need to combine "ref_ent_ids" and "sup_ent_ids" into a single file named "ill_ent_ids")

Here, you can directly unpack the data file after downloading:

unzip data.zip

Reqirements

  • Python3 (tested on 3.7.7)
  • PyTorch (tested on 1.4.0)
  • PyTorch Geometric (PyG) (tested on 1.4.3)
  • TensorBoard (tested on 2.0.2)
  • Numpy
  • Scipy
  • Scikit-learn
  • Graph-tool (if use bootstrapping)

TODO

  • Results of BootEA, TransEdge, MMEA, NAEA are not satisfactory, they need debug (maybe on the bootstrapping process).

There are still many algorithms that need to be implemented (integrated):

  • Semantic Matching Models: NTAM, AttrE, CEAFF, ...
  • GNN-based Models: AVR-GCN, AliNet, MRAEA, CG-MuAlign, RDGCN, HGCN, GMNN, ...
  • KE-based Models: TransD, CapsE, ...
  • GAN-based Models: SEA, AKE, ...
  • Other Models: OTEA, ...

Find algorithms from Entity_Alignment_Papers.

Pull requests for implementing algorithms & updating (reproducible) results with shell scripts are welcome!

Acknowledgement

We refer to some codes of the following repos, and we appreciate for their great contributions: PyTorch Geometric, BootEA, TransEdge, AliNet, TuckER. If we miss some, do please let us know in Issues.

This project is mainly contributed by Chengjiang Li, Lei Hou, Juanzi Li.