BERT-INT

This is code and datasets for BERT-INT

Dependencies

• Python 3 (tested on 3.6.9)
• Pytorch (tested on 1.1.0)
• transformers (tested on 2.1.1)
• Numpy

How to Run

The model runs in two steps:

1. Fine-tune Basic BERT Unit

To fine-tune the Basic BERT Unit, use:

cd basic_bert_unit/
python main.py

Note that basic_bert_unit/Param.py is the config file.

The obtained Basic BERT Unit and some other data will be stored in: ../Save_model

2. Run BERT-based Interaction Model

(Note that when running the BERT-based Interaction model, the parameters of the Basic BERT Unit model will be fixed.)

To extract the similarity features and run the BERT-base Interaction Model, use:

cd ../interaction_model/
python clean_attribute_data.py
python get_entity_embedding.py
python get_attributeValue_embedding.py
python get_neighView_and_desView_interaction_feature.py
python get_attributeView_interaction_feature.py
python interaction_model.py

Or directly use:

cd ../interaction_model/
bash run.sh

Note that interaction_model/Param.py is the config file.

Dataset

Description data

• data/dbp15k/2016-10-des_dict: A dictionary storing entity descriptions, which can be loaded by pickle.load()

The description of the entity is extracted from DBpedia(https://wiki.dbpedia.org/downloads-2016-10)

DBP15K

Initial datasets are from JAPE(https://github.com/nju-websoft/JAPE).

There are three cross-lingual datasets in folder data/dbp15k/ , take the dataset DBP15K(ZH-EN) as an example, the folder data/dbp15k/zh_en contains:

• ent_ids_1: entity ids and entities in source KG (ZH)
• ent_ids_2: entity ids and entities in target KG (EN)
• ref_pairs: entity links encoded by ids (Test Set)
• sup_pairs: entity links encoded by ids (Train Set)
• rel_ids_1: relation ids and relations in source KG (ZH)
• rel_ids_2: relation ids and relations in target KG (EN)
• triples_1: relation triples encoded by ids in source KG (ZH)
• triples_2: relation triples encoded by ids in target KG (EN)
• zh_att_triples: attribute triples of source KG (ZH)
• en_att_triples: attribute triples of target KG (EN)