Genre Linking Blink

🚀 GENRE

The GENRE (Generative ENtity REtrieval) system is a powerful tool for entity retrieval tasks, leveraging a sequence-to-sequence approach based on the fine - tuned BART architecture.

The GENRE (Generative ENtity REtrieval) system, as introduced in Autoregressive Entity Retrieval, is implemented in PyTorch. In essence, GENRE adopts a sequence - to - sequence method for entity retrieval (such as entity linking). It is built upon the fine - tuned BART architecture. GENRE conducts retrieval by generating the unique entity name conditioned on the input text, using constrained beam search to ensure only valid identifiers are generated. The model was initially released in the facebookresearch/GENRE repository with fairseq. The transformers models are obtained through a conversion script similar to this.

This model was trained on the full training set of BLINK, which consists of 9M datapoints for entity - disambiguation grounded on Wikipedia.

📚 Documentation

BibTeX entry and citation info

Please consider citing our works if you use code from this repository.

@inproceedings{decao2020autoregressive,
  title={Autoregressive Entity Retrieval},
  author={Nicola {De Cao} and Gautier Izacard and Sebastian Riedel and Fabio Petroni},
  booktitle={International Conference on Learning Representations},
  url={https://openreview.net/forum?id=5k8F6UU39V},
  year={2021}
}

💻 Usage Examples

Basic Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

# OPTIONAL: load the prefix tree (trie), you need to additionally download
# https://huggingface.co/facebook/genre-linking-blink/blob/main/trie.py and 
# https://huggingface.co/facebook/genre-linking-blink/blob/main/kilt_titles_trie_dict.pkl
# import pickle
# from trie import Trie
# with open("kilt_titles_trie_dict.pkl", "rb") as f:
#     trie = Trie.load_from_dict(pickle.load(f))

tokenizer = AutoTokenizer.from_pretrained("facebook/genre-linking-blink")
model = AutoModelForSeq2SeqLM.from_pretrained("facebook/genre-linking-blink").eval()

sentences = ["Einstein was a [START_ENT] German [END_ENT] physicist."]

outputs = model.generate(
    **tokenizer(sentences, return_tensors="pt"),
    num_beams=5,
    num_return_sequences=5,
    # OPTIONAL: use constrained beam search
    # prefix_allowed_tokens_fn=lambda batch_id, sent: trie.get(sent.tolist()),
)

tokenizer.batch_decode(outputs, skip_special_tokens=True)

This code outputs the following top - 5 predictions (using constrained beam search):

['Germans',
 'Germany',
 'German Empire',
 'Weimar Republic',
 'Greeks']

📄 License

Since no license information is provided in the original README, this section is skipped.

🔧 Technical Details

The GENRE system uses a sequence - to - sequence approach for entity retrieval. It is based on the fine - tuned BART architecture. The model was trained on the full training set of BLINK, which contains 9M datapoints for entity - disambiguation grounded on Wikipedia. It uses constrained beam search to generate valid entity identifiers conditioned on the input text.

Information Table