Inductive entity representations from text via link prediction

This repository contains the code used for the experiments in the paper "Inductive entity representations from text via link prediction", presented at The Web Conference, 2021. To refer to our work, please use the following:

@inproceedings{daza2021inductive,
    title = {Inductive Entity Representations from Text via Link Prediction},
    author = {Daniel Daza and Michael Cochez and Paul Groth},
    booktitle = {Proceedings of The Web Conference 2021},
    year = {2021},
    doi = {10.1145/3442381.3450141},
}

In this work, we show how a BERT-based text encoder can be fine-tuned with a link prediction objective, in a graph where entities have an associated textual description. We call the resulting model BLP. There are three interesting properties of a trained BLP model:

It can predict a link between entities, even if one or both were not present during training.
It produces useful representations for a classifier, that don't require retraining the encoder.
It improves an information retrieval system, by better matching entities and questions about them.

Usage

Please follow the instructions next to reproduce our experiments, and to train a model with your own data.

1. Install the requirements

Creating a new environment (e.g. with conda) is recommended. Use requirements.txt to install the dependencies:

conda create -n blp python=3.7
conda activate blp
pip install -r requirements.txt

2. Download the data

Download the required compressed datasets into the data folder:

Download link	Size (compressed)
UMLS (small graph for tests)	121 KB
WN18RR	6.6 MB
FB15k-237	21 MB
Wikidata5M	1.4 GB
GloVe embeddings	423 MB
DBpedia-Entity	1.3 GB

Then use tar to extract the files, e.g.

tar -xzvf WN18RR.tar.gz

Note that the KG-related files above contain both transductive and inductive splits. Transductive splits are commonly used to evaluate lookup-table methods like ComplEx, while inductive splits contain entities in the test set that are not present in the training set. Files with triples for the inductive case have the ind prefix, e.g. ind-train.txt.

2. Reproduce the experiments

Link prediction

To check that all dependencies are correctly installed, run a quick test on a small graph (this should take less than 1 minute on GPU):

./scripts/test-umls.sh

The table below is a adapted from our paper. The "Script" column contains the name of the script that reproduces the experiment for the corresponding model and dataset. For example, if you want to reproduce the results of BLP-TransE on FB15k-237, run

./scripts/blp-transe-fb15k237.sh

Pre-trained models

We have additionally uploaded weights of pre-trained models for most of these experiments, so that it's possible to skip training and obtain the evaluation results from them. Download links can be found in the table, on each evaluation script.

We also provide evaluation scripts for these pre-trained models, which have the same names as training scripts, appended with -pretrained. For example, to reproduce the results of BLP-TransE on FB15k-237, first download the weights by following the link in the table, untar the file, and run

./scripts/blp-transe-fb15k237-pretrained.sh

	WN18RR		FB15k-237		Wikidata5M
Model	MRR	Script + weights link	MRR	Script + weights link	MRR	Script + weights link
GlovE-BOW	0.170	glove-bow-wn18rr.sh	0.172	glove-bow-fb15k237.sh	0.343	glove-bow-wikidata5m.sh
BE-BOW	0.180	bert-bow-wn18rr.sh	0.173	bert-bow-fb15k237.sh	0.362	bert-bow-wikidata5m.sh
GloVe-DKRL	0.115	glove-dkrl-wn18rr.sh	0.112	glove-dkrl-fb15k237.sh	0.282	glove-dkrl-wikidata5m.sh
BE-DKRL	0.139	bert-dkrl-wn18rr.sh	0.144	bert-dkrl-fb15k237.sh	0.322	bert-dkrl-wikidata5m.sh
BLP-TransE	0.285	blp-transe-wn18rr.sh	0.195	blp-transe-fb15k237.sh	0.478	blp-transe-wikidata5m.sh
BLP-DistMult	0.248	blp-distmult-wn18rr.sh	0.146	blp-distmult-fb15k237.sh	0.472	blp-distmult-wikidata5m.sh
BLP-ComplEx	0.261	blp-complex-wn18rr.sh	0.148	blp-complex-fb15k237.sh	0.489	blp-complex-wikidata5m.sh
BLP-SimplE	0.239	blp-simple-wn18rr.sh	0.144	blp-simple-fb15k237.sh	0.493	blp-simple-wikidata5m.sh

Entity classification

After training for link prediction, a tensor of embeddings for all entities is computed and saved in a file with name ent_emb-[ID].pt where [ID] is the id of the experiment in the database (we use Sacred to manage experiments). Another file called ents-[ID].pt contains entity identifiers for every row in the tensor of embeddings.

To ease reproducibility, we provide these tensors, which are required in the entity classification task. Click on the ID, download the file into the output folder, and decompress it. An experiment can be reproduced using the following command:

python train.py node_classification with checkpoint=ID dataset=DATASET

where DATASET is either WN18RR or FB15k-237. For example:

python train.py node_classification with checkpoint=199 dataset=WN18RR

	WN18RR		FB15k-237
Model	Acc.	ID	Acc. Bal.	ID
GloVe-BOW	55.3	219	34.4	293
BE-BOW	60.7	218	28.3	296
GloVe-DKRL	55.5	206	26.6	295
BE-DKRL	48.8	207	30.9	294
BLP-TransE	81.5	199	42.5	297
BLP-DistMult	78.5	200	41.0	298
BLP-ComplEx	78.1	201	38.1	300
BLP-SimplE	83.0	202	45.7	299

Information retrieval

This task runs with a pre-trained model saved from the link prediction task. For example, if the model trained is blp with transe and it was saved as model.pt, then run the following command to run the information retrieval task:

python retrieval.py with model=blp rel_model=transe \
checkpoint='output/model.pt'

Using your own data

If you have a knowledge graph where entities have textual descriptions, you can train a BLP model for the tasks of inductive link prediction, and entity classification (if you also have labels for entities).

To do this, add a new folder inside the data folder (let's call it my-kg). Store in it a file containing the triples in your KG. This should be a text file with one tab-separated triple per line (let's call it all-triples.tsv).

To generate inductive splits, you can use data/utils.py. If you run

python data/utils.py drop_entities --file=my-kg/all-triples.tsv

this will generate ind-train.tsv, ind-dev.tsv, ind-test.tsv inside my-kg (see Appendix A in our paper for details on how these are generated). You can then train BLP-TransE with

python train.py with dataset='my-kg'

Alternative implementations

Contextual Knowledge Bases by Raphael Sourty

2391134843/blp