Skip to content

Ontology Matching with BERTMap and BERTMapLt


Paper for BERTMap: BERTMap: A BERT-based Ontology Alignment System (AAAI-2022).

  title={BERTMap: a BERT-based ontology alignment system},
  author={He, Yuan and Chen, Jiaoyan and Antonyrajah, Denvar and Horrocks, Ian},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},

This page gives the tutorial for \(\textsf{BERTMap}\) family including the summary of the models and how to use them.


Figure 1. Pipeline illustration of BERTMap.

The ontology matching (OM) pipeline of \(\textsf{BERTMap}\) consists of following steps:

  1. Load the source and target ontologies and build annotation indices from them based on selected annotation properties.
  2. Construct the text semantics corpora including intra-ontology (from input ontologies), cross-ontology (optional, from input mappings), and auxiliary (optional, from auxiliary ontologies) sub-corpora.
  3. Split samples in the form of (src_annotation, tgt_annotation, synonym_label) into training and validation sets.
  4. Fine-tune a BERT synonym classifier on the samples and obtain the best checkpoint on the validation split.
  5. Predict mappings for each class \(c\) of the source ontology \(\mathcal{O}\) by:

    • Selecting plausible candidates \(c'\)s in the target ontology \(\mathcal{O'}\) based on idf scores w.r.t. the sub-word inverted index built from the target ontology annotation index. For \(c\) and a candidate \(c'\), first check if they can be string-matched (i.e., share a common annotation, or equivalently the maximum edit similarity score is \(1.0\)); if not, consider all combinations (cartesian product) of their respective class annotations, compute a synonym score for each combination, and take the average of synonym scores as the mapping score.
    • \(N\) best scored mappings (no filtering) will be preserved as raw predictions which should have relatively higher recall and lower precision.
  6. Extend the raw predictions using an iterative algorithm based on the locality principle. To be specific, if \(c\) and \(c'\) are matched with a relatively high mapping score (\(\geq \kappa\)), then search for plausible mappings between the parents (resp. children) of \(c\) and the parents (resp. children) of \(c'\). This process is iterative because there would be new highly scored mappings at each round. Terminate mapping extension when there is no new mapping with score \(\geq \kappa\) found or it exceeds the maximum number of iterations. Note that \(\kappa\) is set to \(0.9\) by default, as in the original paper.

  7. Truncate the extended mappings by preserving only those with scores \(\geq \lambda\). In the original paper, \(\lambda\) is supposed to be tuned on validation mappings – which are often not available. Also, \(\lambda\) is not a sensitive hyperparameter in practice. Therefore, we manually set \(\lambda\) to \(0.9995\) as a default value which usually yields a higher F1 score. Note that both \(\kappa\) and \(\lambda\) are made available in the configuration file.

  8. Repair the rest of the mappings with the repair module built in LogMap (BERTMap does not focus on mapping repair). In short, a minimum set of inconsistent mappings will be removed (further improve precision).

Steps 5-8 are referred to as the global matching process which computes OM mappings from two input ontologies. \(\textsf{BERTMapLt}\) is the light-weight version without BERT training and mapping refinement. The mapping filtering threshold for \(\textsf{BERTMapLt}\) is \(1.0\) (i.e., string-matched).

In addition to the traditional OM procedure, the scoring modules of \(\textsf{BERTMap}\) and \(\textsf{BERTMapLt}\) can be used to evaluate any class pair given their selected annotations. This is useful in ranking-based evaluation.


The \(\textsf{BERTMap}\) family rely on sufficient class annotations for constructing training corpora of the BERT synonym classifier, especially under the unsupervised setting where there are no input mappings and/or external resources. It is very important to specify correct annotation properties in the configuration file.


To use \(\textsf{BERTMap}\), a configuration file and two input ontologies to be matched should be imported.

from deeponto.onto import Ontology
from deeponto.align.bertmap import BERTMapPipeline

config_file = "path_to_config.yaml"
src_onto_file = "path_to_the_source_ontology.owl"  
tgt_onto_file = "path_to_the_target_ontology.owl" 

config = BERTMapPipeline.load_bertmap_config(config_file)
src_onto = Ontology(src_onto_file)
tgt_onto = Ontology(tgt_onto_file)

BERTMapPipeline(src_onto, tgt_onto, config)

The default configuration file can be loaded as:

from deeponto.align.bertmap import BERTMapPipeline, DEFAULT_CONFIG_FILE

config = BERTMapPipeline.load_bertmap_config(DEFAULT_CONFIG_FILE)

The loaded configuration is a CfgNode object supporting attribute access of dictionary keys. To customise the configuration, users can either copy the DEFAULT_CONFIG_FILE, save it locally using BERTMapPipeline.save_bertmap_config method, and modify it accordingly; or change it in the run time.

from deeponto.align.bertmap import BERTMapPipeline, DEFAULT_CONFIG_FILE

config = BERTMapPipeline.load_bertmap_config(DEFAULT_CONFIG_FILE)

# save the configuration file
BERTMapPipeline.save_bertmap_config(config, "path_to_saved_config.yaml")

# modify it in the run time
# for example, add more annotation properties for synonyms

If using \(\textsf{BERTMap}\) for scoring class pairs instead of global matching, disable automatic global matching and load class pairs to be scored.

from deeponto.onto import Ontology
from deeponto.align.bertmap import BERTMapPipeline

config_file = "path_to_config.yaml"
src_onto_file = "path_to_the_source_ontology.owl"  
tgt_onto_file = "path_to_the_target_ontology.owl" 

config = BERTMapPipeline.load_bertmap_config(config_file)
config.global_matching.enabled = False
src_onto = Ontology(src_onto_file)
tgt_onto = Ontology(tgt_onto_file)

bertmap = BERTMapPipeline(src_onto, tgt_onto, config)

class_pairs_to_be_scored = [...]  # (src_class_iri, tgt_class_iri)
for src_class_iri, tgt_class_iri in class_pairs_to_be_scored:
    # retrieve class annotations
    src_class_annotations = bertmap.src_annotation_index[src_class_iri]
    tgt_class_annotations = bertmap.tgt_annotation_index[tgt_class_iri]
    # the bertmap score
    bertmap_score = bertmap.mapping_predictor.bert_mapping_score(
        src_class_annotations, tgt_class_annotations
    # the bertmaplt score
    bertmaplt_score = bertmap.mapping_predictor.edit_similarity_mapping_score(
        src_class_annotations, tgt_class_annotations


The implemented \(\textsf{BERTMap}\) by default searches for each source ontology class a set of possible matched target ontology classes. Because of this, it is recommended to set the source ontology as the one with a smaller number of classes for efficiency.

Note that in the original paper, the model is expected to match for both directions src2tgt and tgt2src, and also consider the combination of both results. However, this does not usually bring better performance and consumes significantly more time. Therefore, this feature is discarded and the users can choose which direction to match.


Occasionally, the fine-tuning loss may not be converging and the validation accuracy is not improving; in that case, set to a different random seed can usually fix the problem.


The default configuration file looks like:

model: bertmap  # bertmap or bertmaplt

output_path: null  # if not provided, the current path "." is used

  -  # rdfs:label

# additional corpora 
known_mappings: null  # cross-ontology corpus
auxiliary_ontos: [] # auxiliary corpus

# bert config
  pretrained_path: emilyalsentzer/Bio_ClinicalBERT  
  max_length_for_input: 128 
  num_epochs_for_training: 3.0
  batch_size_for_training: 32
  batch_size_for_prediction: 128
  resume_training: null

# global matching config
  enabled: true
  num_raw_candidates: 200 
  num_best_predictions: 10 
  mapping_extension_threshold: 0.9   
  mapping_filtered_threshold: 0.9995 
  for_oaei: false


By changing this parameter to bertmap or bertmaplt, users can switch between \(\textsf{BERTMap}\) and \(\textsf{BERTMapLt}\). Note that \(\textsf{BERTMapLt}\) does not use any training and mapping refinement parameters.

Annotation Properties ⚠

The IRIs stored in this parameter refer to annotation properties with literal values that define the synonyms of an ontology class. Many ontology matching systems rely on synonyms for good performance, including the \(\textsf{BERTMap}\) family. The default config.annotation_property_iris are in line with the Bio-ML dataset, which will be constantly updated. Users can append or delete IRIs for specific input ontologies.

Note that it is safe to specify all possible annotation properties regardless of input ontologies because the ones that are not used will be ignored.

Additional Training Data

The text semantics corpora by default (unsupervised setting) will consist of two intra-ontology sub-corpora built from two input ontologies (based on the specified annotation properties). To add more training data, users can opt to feed input mappings (cross-ontology sub-corpus) and/or a list of auxiliary ontologies (auxiliary sub-corpora).

Specify the path to input mapping file here; the input mapping file should be a .tsv or .csv file with three columns with headings: ["SrcEntity", "TgtEntity", "Score"]. Each row corresponds to a triple \((c, c', s(c, c'))\) where \(c\) is a source ontology class, \(c'\) is a target ontology class, and \(s(c, c')\) is the matching score. Note that in the BERTMap context, input mapppings are assumed to be gold standard (reference) mappings with scores equal to \(1.0\). Regardless of scores specified in the mapping file, the scores of the input mapppings will be adjusted to \(1.0\) automatically.
Specify a list of paths to auxiliary ontology files here. For each auxiliary ontology, a corresponding intra-ontology corpus will be created and thus produce more synonym and non-synonym samples.

BERT Settings

\(\textsf{BERTMap}\) uses the pre-trained Bio-Clincal BERT as specified in this parameter because it was originally applied on biomedical ontologies. For general purpose ontology matching, users can use pre-trained variants such as bert-base-uncased.
Batch size for BERT fine-tuning.
Batch size for BERT validation and mapping prediction.

Adjust these two parameters if users found an inappropriate GPU memory fit.

Set to true if the BERT training process is somehow interrupted and users wish to continue training.

Global Matching Settings

As mentioned in usage, users can disable automatic global matching by setting this parameter to false if they wish to use the mapping scoring module only.
Set the number of raw candidates selected in the mapping prediction phase.
Set the number of best scored mappings preserved in the mapping prediction phase. The default value 10 is often more than enough.
Set the score threshold of mappings used in the iterative mapping extension process. Higher value shortens the time but reduces the recall.
The score threshold of mappings preserved for final mapping refinement.
Set to false for normal use and set to true for the OAEI 2023 Bio-ML Track such that entities that are annotated as not used in alignment will be ignored during global matching.

Output Format

Running \(\textsf{BERTMap}\) will create a directory named bertmap or bertmaplt in the specified output path. The file structure of this directory is as follows:

├── data
│   ├──
│   └── text-semantics.corpora.json
├── bert
│   ├── tensorboard
│   ├── checkpoint-{some_number}
│   └── checkpoint-{some_number}
├── match
│   ├── logmap-repair
│   ├── raw_mappings.json
│   ├── repaired_mappings.tsv 
│   ├── raw_mappings.tsv
│   ├── extended_mappings.tsv
│   └── filtered_mappings.tsv
├── bertmap.log
└── config.yaml

It is worth mentioning that the match sub-directory contains all the global matching files:

The raw mapping predictions before mapping refinement. The .json one is used internally to prevent accidental interruption. Note that bertmaplt only produces raw mapping predictions (no mapping refinement).
The output mappings after applying mapping extension.
The output mappings after mapping extension and threshold filtering.
A folder containing intermediate files needed for applying LogMap's debugger.
The final output mappings after mapping repair.

Last update: September 3, 2023
Created: November 21, 2021
GitHub: @Lawhy   Personal Page: