NPEL(Neural Paired Entity Linking) is the entity linking method proposed by "Neural Paired Entity Linking in Web Tables"
There are three datasets used in NPEL experiment. Web_Manual_fixed_2020(with 347 tables), Wiki_Links_Large(with 168 tables),Wiki_Links_Random_2020(with 2335 tables).
Director Web_Manual_Fixed_2020 has two subdirectors. One of them is table data in json format and another director is the labeled entity correspondent to table data in the same position.
Director Wiki_Links_Large and Wiki_Links_Random_2020 share the same struct of files. They have only one tables director. All table data and labeled are in json type files. Each table mention labeled entity is annotate as target in surfaceLinks
code contains the mention-entity semantics model and pair-linking algorithm implementation.
The mention-entity semantics model mainly included in nn_models and pair-linking algorithm is implemented in pair_linking.py
Both neural network model and pair-linking algorithm required input data format each mention in a table contains all it's embedding, candidates, candidates' embedding, candidates' abstract. These elements are directly accessed as dict in model or algorithm.
Besides, we utilized this work to generate prior data.
This project required enwiki as knowledge base, which can download from this. All enwiki articles will be used to train entities and words embedding. All entity abstract need to be extracted for entity encoding. All anchor text will be used to calculate the prior of mention to entities.
data_generate.py/gen_abstract is used to get the necessary abstract of tables' entities.
data_generate.py/gen_embedding is used to get the necessary embedding of tables' entities and mentions/words.
data_generate.py/dump_mention_abstract_emb merge entity abstract and embedding to one file.
data_generate.py/rebuild_? is used for prepare format data to deep learning model and pair-linking algorithm.
pair_linking.py implement algorithm of pair-linking, it required format table