nlpathits / figet-hyperbolic-space

Code for the paper "Fine-Grained Entity Typing in Hyperbolic Space"

Fine-Grained Entity Typing in Hyperbolic Space

Code for the paper "Fine-Grained Entity Typing in Hyperbolic Space" published at RepL4NLP @ ACL 2019

Model overview:

Citation

The source code and data in this repository aims at facilitating the study of fine-grained entity typing. If you use the code/data, please cite it as follows:

@inproceedings{lopez-etal-2019-fine,
    title = "Fine-Grained Entity Typing in Hyperbolic Space",
    author = "L{\'o}pez, Federico  and
      Heinzerling, Benjamin  and
      Strube, Michael",
    booktitle = "Proceedings of the 4th Workshop on Representation Learning for NLP (RepL4NLP-2019)",
    month = aug,
    year = "2019",
    address = "Florence, Italy",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/W19-4319",
    pages = "169--180",
}

Dependencies

• PyTorch 1.1
• tqdm
• tensorboardX
• pyflann

A conda environment can be created as well from the environment.yml file.

To embed the graphs into the different metric spaces the library Hype was used.

Running the code

1. Download data

Download and uncompress Ultra-Fine dataset and GloVe word embeddings:

./scripts/figet.sh get_data

2. Preprocess data

The parameter freq-sym can be replaced to store different preprocessing configurations:

./scripts/figet.sh preprocess freq-sym

3. Train model

The name of the preprocessing used in the previous step must be given as a parameter.

./scripts/figet.sh train freq-sym

3. Do inference

./scripts/figet.sh inference freq-sym

Acknowledgements

We thank to Choi et al for the release of the Ultra-Fine dataset and their model.

License

MIT

figet-hyperbolic-space's Issues

Run on custom dataset

Hi,
Thank you for releasing such a nice and easy to use code.
I wanted to ask about procedure to get predictions on a custom dataset. From what I understand infer.py computes all f-scores on top of predictions using figet.Coach. I believe this line would allow me get fine grained predictions and I should dump them. Is it correct?

Also for dataset formatting, I should pass the sentences to a NER model and format it in form of left_context, mention_span, right_context right?

Speed-up the validate phase

Hi,

I'm training your model using a different target space.

I'm noticing a very long computation time in the validate-typing-dev-# phase (15 minutes per epoch when the target space has 63 labeled vectors, 48 minutes when the types are approximately 200).

I know that in this phase the pyFLANN library is used an I think that is the Nearest Neighbor algorithm that slows down this phase.

Since the original dataset presents more than 8k types in the target space I think that you speed up this phase (or maybe I'm missing something in the code, but I didn't touch anything in the validation code) so I am here to ask if there is some method to speed up the training phase.

Thanks

Strange strings in the enviroment.yml

Hi,

I think that the enviroment.yml has some errors in it