Auto-EM source code

This repo hosts the source code used in the paper "Auto-EM: End-to-end Fuzzy Entity-Matching using Pre-trained Deep Models and Transfer Learning", published in the Web conference (WWW) 2019.

What are included

• Source code for entity typing and entity matching models

Dependency Installation

Run pip install -r requirements.txt.

Data files

We leverage the entity-type and entity-synonym data extracted from Microsoft's properietary KB, to train Auto-EM. While we could not directly release the data due to the proprietary nature of the KB, similar data are readily available in open-source KBs. For example, in Wikidata, each entity has an attribute called also known as, which lists alias/synonyms for the given entity (e.g., the Wikidata page for 'Bill Gates' lists 'William Gates', 'William Henry Gates III' as synonyms. Similarly, the Freebase also has an alias attribute for entity synonyms that can be used to train Auto-EM models).

Data format

Although we could not release our training data, we provide sample files (under data/) to demonstrate the data format accepted by the code (so that data extracted from Wikidata or Freebase can be made into the same format and run).

For entity typing (given an entity name, predict its entity types -- e.g., 'Bill Gates' is of type 'person', 'politician', etc.)

• The format for each line in data/sample_data_type.txt is: entity_name lb1___lb2___lb3
• Here, entity_name and its corresponding entity_types are seperated by a tab character, with multiple type-ids are seperated by underscores.

For entity matching (given two entity names, predict whether they are a match or not -- e.g., 'Bill Gates' and 'William Gates' is a match, but 'Bill Gates' and 'Bill Gates Sr.' is not a match)

• The format for each line in data/sample_data_match.txt is: entity_name pos_alias neg_alias1___neg_alias2___neg_alias3
• Here, each line has three components, separated by tabs. The first component is one canonical entity_name. The second component is its positive alias/synonyms. The third component is its negative alias/synonyms.

Entity Typing Model Training and Testing

For entity typing, cd type and run python alias_type_train.py to train a entity typing model, with following input arguments

• --train-file for training data
• --dev-file for dev data
• --test-file for test data
• --save-model for saved model path
• --load-model for loading model checkpoint from the path

Once the model is trained, run python alias_type_train.py --test (with your saved model) to make predictions on the test set.

Also see type/alias_type_train.py for command-line arguments to specify training file, dev file, test file, model saving location, etc.

Entity Matching Model Training and Testing

For entity matching, cd matching and run python hybrid_train.py to train a entity matching model, with following input arguments

• --train-file for training data
• --dev-file for dev data
• --test-file for test data
• --save-model for saved model path
• --load-model for loading model checkpoint from the path

Once the model is trained, run python hybrid_train.py --test (with your saved model) to make predictions on the test set.

Also see matching/hybrid_train.py for command-line arguments to specify training file, dev file, test file, model saving location, etc.

Trademarks

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft’s Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party’s policies.

Contact

If you have questions, suggestions and bug reports, please email the authors in the paper.