Fine-Grained Semantics to Entity Embeddings (FGS2EE)

This repository contains code and data for the ACL 2020 paper,

@inproceedings{hou_2020_,
  title={Improving Entity Linking through Semantic Reinforced Entity Embeddings},
  author={Feng Hou and Ruili Wang and Jun He and Yi Zhou},
  booktitle={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
  year={2020}
}

The entity embeddings are tested on the following two linking models:

• the wnel model
• the mulrel-nel model

You can download the entity embeddings directly and test them on your model.

Generic steps

if you want to reproduce the entity embeddings, please follow the following steps:

• Download dataset
• Generate the semantic entity embeddings
• Generate the aggregated entity embeddings
• Test the entity embeddings on the linking models

Download dataset

Download our data

Download our data from Googledrive

1. the entities_with_types_wikitext.zip contains [entity-name, entity types, wikipedia article] from wikipedia-dump, unzip this file to directory entities_types_texts.
2. the type_dict.type (not include som OOVs) is a dictionary for type_word embeddings type_vec.npy, which is extracted from Word2Vec.
3. type_list.ndjson is the originally selected type words, type_list_OOVs_remap.ndjson is remapping OOV words.

Download ganea's entity embeddings data

Download from links in repository wenl and mulrel-nel, get the following two files:

1. the entity dictionary (ganea)dict.entity
2. the entity embeddings (ganea)entity_embeddings.npy

Generate the semantic entity embeddings

run: python generate_semantic_embeddings.py <path of type_dict.type> <path of type_vec.npy> <directory of entities_types_texts> <saving_path> <value of T>

• saving_path is the directory for saving semantic entity embeddings

Generate aggregated entity embeddings

run: python generate_aggregated_embeddings.py <path of semantic_dict.entity> <path of semantic_entity_vec.npy> <path of ganea_dict.entity> <path of ganea_entity_vec.npy> <saving path> <value of alpha>