BERTMap: A BERT-based Ontology Alignment System

Important Notices

• The relevant paper was accepted in AAAI-2022.
• Arxiv version is available at: https://arxiv.org/abs/2112.02682.
• Code will be re-implemented as an example model in Onto-PLM, which will be a package for ontology engineering.

About

BERTMap is a BERT-based ontology alignment system, which utilizes the textual knowledge of ontologies to fine-tune BERT and make prediction. It also incorporates sub-word inverted indices for candidate selection, and (graph-based) extension and (logic-based) repair modules for mapping refinement.

Essential dependencies

The following packages are necessary but not sufficient for running BERTMap:

conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch  # pytorch
pip install cython  # the optimized parser of owlready2 relies on Cython
pip install owlready2  # for managing ontologies
pip install tensorboard  # tensorboard logging (optional)
pip install transformers  # huggingface library
pip install datasets  # huggingface datasets

Running BERTMap

IMPORTANT NOTICE: BERTMap relies on class labels for training, but different ontologies have different annotation properties to define the aliases (synonyms), so preprocessing is required for adding all the synonyms to rdf:label before running BERTMap. The preprocessed ontologies involved in our paper together with their reference mappings are available in data.zip.

Clone the repository and run:

# fine-tuning and evaluate bertmap prediction 
python run_bertmap.py -c config.json -m bertmap

# mapping extension (-e specify which mapping set {src, tgt, combined} to be extended)
python extend_bertmap.py -c config.json -e src

# evaluate extended bertmap 
python eval_bertmap.py -c config.json -e src

# repair and evluate final outputs (-t specify best validation threshold)
python repair_bertmap.py -c config.json -e src -t 0.999

# baseline models (edit similarity and pretrained bert embeddings)
python run_bertmap.py -c config.json -m nes
python run_bertmap.py -c config.json -m bertembeds

The script skips data construction once built for the first time to ensure that all of the models share the same set of pre-processed data.

The fine-tuning model is implemented with huggingface Trainer, which by default uses multiple GPUs, for restricting to GPUs of specified indices, please run (for example):

# only device (1) and (2) are visible to the script
CUDA_VISIBLE_DEVICES=1,2 python run_bertmap.py -c config.json -m bertmap 

Configurations

Here gives the explanations of the variables used in config.json for customized BERTMap running.

• data:
  • task_dir: directory for saving all the output files.
  • src_onto: source ontology name.
  • tgt_onto: target ontology name.
  • task_suffix: any suffix of the task if needed, e.g. the LargeBio track has 'small' and 'whole'.
  • src_onto_file: source ontology file in .owl format.
  • tgt_onto_fil: target ontology file in .owl format.
  • properties: list of textual properties used for constructing semantic data , default is class labels: ["label"].
  • cut: threshold length for the keys of sub-word inverted index, preserve the keys only if their lengths > cut, default is 0.
• corpora:
  • sample_rate: number of (soft) negative samples for each positive sample generated in corpora (not the ultimate fine-tuning data).
  • src2tgt_mappings_file: reference mapping file for evaluation and semi-supervised learning setting in .tsv format with columns: "Entity1", "Entity2" and "Value".
  • ignored_mappings_file: file in .tsv format but stores mappings that should be ignored by the evaluator.
  • train_map_ratio: proportion of training mappings to used in semi-supervised setting, default is 0.2.
  • val_map_ratio: proportion of validation mappings to used in semi-supervised setting, default is 0.1.
  • test_map_ratio: proportion of test mappings to used in semi-supervised setting, default is 0.7.
  • io_soft_neg_rate: number of soft negative sample for each positive sample generated in the fine-tuning data at the intra-ontology level.
  • io_hard_neg_rate: number of hard negative sample for each positive sample generated in the fine-tuning data at the intra-ontology level.
  • co_soft_neg_rate: number of soft negative sample for each positive sample generated in the fine-tuning data at the cross-ontology level.
  • depth_threshold: classes of depths larger than this threshold will not considered in hard negative generation, default is null.
  • depth_strategy: strategy to compute the depths of the classes if any threshold is set, default is max, choices are max and min.
• bert
  • pretrained_path: real or huggingface library path for pretrained BERT, e.g. "emilyalsentzer/Bio_ClinicalBERT" (BioClinicalBERT).
  • tokenizer_path: real or huggingface library path for BERT tokenizer, e.g. "emilyalsentzer/Bio_ClinicalBERT" (BioClinicalBERT).
• fine-tune
  • include_ids: include identity synonyms in the positive samples or not.
  • learning: choice of learning setting ss (semi-supervised) or us (unsupervised).
  • warm_up_ratio: portion of warm up steps.
  • max_length: maximum length for tokenizer (highly important for large task!).
  • num_epochs: number of training epochs, default is 3.0.
  • batch_size: batch size for fine-tuning BERT.
  • early_stop: whether or not to apply early stopping (patience has been set to 10), default is false.
  • resume_checkpoint: path to previous checkpoint if any, default is null.
• map
  • candidate_limits: list of candidate limits used for mapping computation, suggested values are [25, 50, 100, 150, 200].
  • batch_size: batch size used for mapping computation.
  • nbest: number of top results to be considered.
  • string_match: whether or not to use string match before others.
  • strategy: strategy for classifier scoring method, default is mean.
• eval:
  • automatic: whether or not automatically evaluate the mappings.

Should you need any further customizaions especially on the evaluation part, please set eval: automatic to false and use your own evaluation script.

Acknolwedgements

The repair module is credited to Ernesto Jiménez Ruiz et al., and the code can be found here.