OntoEMMA ontology matcher

This ontology matcher can be used to generate alignments between knowledgebases.

Installation

Go to the base git directory and run: ./setup.sh

This will create an ontoemma conda environment and install all required libraries.

Train OntoEmma

To train an alignment model, use train_ontoemma.py. The wrapper takes the following arguments:

• -p <model_type> (lr = logistic regression, nn = neural network)
• -m <model_path>
• -c <configuration_file>

Example usage:

python train_ontoemma.py -p nn -m model_path -c configuration_file.json

This script will then use the train function in OntoEmma.py to train the model.

OntoEmma module

The module OntoEmma is used for accessing the training and alignment capabilities of OntoEmma.

Train mode

In training mode, the OntoEmma module can use the OntoEmmaLRModel logistic regression module or AllenNLP to train the model:

NN with AllenNLP:

• Training data is formatted according to Data format: OntoEmma training data
• Train model using AllenNLP; example configuration file given in: config/example_ontoemma_config.json
• Save model to specified serialization directory

Configuration file:

• Training and validation data are outputs of extract_training_data_from_umls.py
• Enriched training/validation data have been enriched with DBPedia and MeSH terms, and Wikipedia summary sentences, see scripts/enrich_match_data.py

When training other models with OntoEmmaModel, the module performs the following:

• Load training data from file (training data is extracted from UMLS, see UMLS training data for more information)
• Iterate through KB pairs in training data, loading KBs, calculating features for each pair
• Train OntoEmmaModel using features and labels calculated from training data
• Save OntoEmmaModel to disk

Align mode

In alignment mode, the OntoEmma module performs the following:

• Load source and target ontologies specified by source_kb_path and target_kb_path (OntoEmma is able to handle KnowledgeBase json and pickle files, as well as KBs in OBO, OWL, TTL, and RDF formats to the best of its ability. It can also load an ontology from a web URI.)
• Initialize CandidateSelection module for source and target KBs

If using NN model with AllenNLP:

• Write candidates to file
• Call AllenNLP OntoEmma predictor on data
• Read predictor output from file

If using logistic regression model:

• Load LR model
• Initialize FeatureGenerator module
• For each candidate pair, generate features using FeatureGenerator and make alignment predictions using OntoEmmaModel

For all models following predictions:

• If gold_file_path is specified, evaluate alignment against gold standard
• If output_file_path is specified, save alignment to file

Candidate selection module

The module CandidateSelection is used to select candidate matched pairs from the source and target KBs.

CandidateSelection is initialized with the following inputs:

• source_kb source KB as KnowledgeBase object
• target_kb target KB as KnowledgeBase object

The module builds the following token maps:

• s_ent_to_tokens mapping entities in source KB to word and n-gram character tokens
• t_ent_to_tokens mapping entities in target KB to word and n-gram character tokens
• s_token_to_ents mapping tokens found in source KB to entities in source KB containing those tokens
• t_token_to_ents mapping tokens found in target KB to entities in target KB containing those tokens
• s_token_to_idf mapping tokens found in source KB to their IDF in source KB
• t_token_to_idf mapping tokens found in target KB to their IDF in target KB

Candidates are accessed through the select_candidates method, which takes an input research_entity_id from the source KB and returns an ordered list of candidates from the target KB. The candidates are ordered by the sum of their token IDF scores.

The output of the CandidateSelection module is evaluated using the eval method, which takes as input:

• gold_mappings a list of tuples defining the gold standard mappings
• top_ks a list of k's (top k from candidate list to return)

The eval method compares the candidates generated against the gold standard mappings, returning the following:

• cand_count total candidate yield
• precisions precision value associated with each k in top_ks
• recalls recall value associated with each k in top_ks

Feature generation module

The module FeatureGenerator is used to generate features from a candidate pair.

FeatureGenerator is initialized with the following inputs:

• s_kb source KB as KnowledgeBase object
• t_kb target KB as KnowledgeBase object

The module generates word and character-based n-gram tokens of entity aliases and the canonical names of entity parents and children. The module also uses a nltk stemmer and lemmatizer to produce stemmed and lemmatized version of canonical name tokens.

The calculate_features method is the core of this module. It generates a set of pairwise features between two input entities given by their respective entity ids from the source and target KBs. This is returned as the feature vector used in the OntoEmmaModel.

UMLS training data

The module extract_training_data_from_umls is used to extract KB and concept mapping data from UMLS for use in ontology matching training and evaluation.

extract_training_data_from_umls takes as inputs:

• UMLS_DIR directory to UMLS META RRF files
• OUTPUT_DIR directory to write KnowledgeBases and mappings

UMLS data subsets are currently located at /net/nfs.corp/s2-research/scigraph/data/ontoemma/2017AA_OntoEmma/. OUTPUT_DIR defaults to /net/nfs.corp/s2-research/scigraph/data/ontoemma/umls_output/.

extract_training_data_from_umls produces as output:

• KnowledgeBase objects written to json files in OUTPUT_DIR/kbs/. One json object is generated for each KnowledgeBase. Currently, KBs specified in kb_name_list are read from the UMLS subset and converted into KnowledgeBase objects.
• Pairwise KB mapping files are written to tsv files in OUTPUT_DIR/mappings/. Mappings are derived as pairs of raw_ids that are mapped under the same concept ID in UMLS.
• Negative mappings are sampled from UMLS as negative training data. The full training data including negatives are written to tsv files in OUTPUT_DIR/training/.

Data downloads

You will need to download a copy of the UMLS Metathesaurus by following these instructions: https://www.nlm.nih.gov/research/umls/knowledge_sources/metathesaurus/index.html

Contexts extracted from the Semantic Scholar API for some UMLS KBs are available in this S3 bucket.

aws s3 cp s3://ai2-s2-ontoemma/contexts/ data/kb_contexts/ --recursive

Once you have downloaded both datasets, update the corresponding path variables in emma/paths.py to point to the appropriate directories.

Sampling negative data

Hard negatives are sampled using the CandidateSelection module, selecting from candidate pairs that are negative matches. Easy negatives are sampled randomly from the rest of the KB. Currently, 5 hard negatives and 5 easy negatives are sampled for each positive match.

Mapping file format

Mapping files are of the format described in Data format: KB alignment file. For UMLS positive mappings, the provenance is given as <UMLS_header>:<CUI>; for UMLS negative mappings, as <UMLS_header>. Example data consisting of two positive and two negative mappings:

CPT:90281 DRUGBANK:DB00028 1 UMLS2017AA:C0358321 CPT:90283 DRUGBANK:DB00028 1 UMLS2017AA:C0358321 CPT:83937 DRUGBANK:DB00426 0 UMLS2017AA CPT:1014233 DRUGBANK:DB05907 0 UMLS2017AA

Run OntoEmma (align two KBs using trained model)

To run OntoEmma, use run_ontoemma.py. The wrapper implements the following arguments:

• -p <model_type> (lr = logistic regression, nn = neural network)
• -m <model_path>
• -s <source_ont>
• -t <target_ont>
• -i <input_alignment>
• -o <output_file>
• -g <cuda_device>

Example usage:

python run_ontoemma.py -p nn -m model_path -s source_ont.owl -t target_ont.owl -i input_alignment.tsv -o output_alignment.tsv -g 0

This script assumes that the model has been pre-trained, and uses align functions in OntoEmma.py accordingly.

Human annotations for evaluation

Available here

Other

String processing utilities are found in string_utils.py.

Constants used by OntoEmma are found in constants.py. Input training data and training model parameters are specified in the training configuration files in config/.

This script will then use the train function in OntoEmma.py to train the model. If no GPU is specified, the program defaults to CPU.