🕶️ ExMod

This repo contains a re-implementation of the original approach for Exemplification Modeling, i.e., the task of generating usage examples given one or more words along with their textual definitions, presented by Barba et al. (2021).

The repo currently focuses on k=1. A live demo is available at:

If you find our work useful, please cite us with:

@inproceedings{ijcai2021-520,
  title     = {Exemplification Modeling: Can You Give Me an Example, Please?},
  author    = {Barba, Edoardo and Procopio, Luigi and Lacerra, Caterina and Pasini, Tommaso and Navigli, Roberto},
  booktitle = {Proceedings of the Thirtieth International Joint Conference on
               Artificial Intelligence, {IJCAI-21}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},
  editor    = {Zhi-Hua Zhou},
  pages     = {3779--3785},
  year      = {2021},
  month     = {8},
  note      = {Main Track}
  doi       = {10.24963/ijcai.2021/520},
  url       = {https://doi.org/10.24963/ijcai.2021/520},
}

Installation

Requirements:

• Debian-based (e.g. Debian, Ubuntu, ...) system
• conda installed

Run the following command to quickly setup the env needed to run our code:

bash setup.sh

Training

This repo builds on top of the classy framework. To train on custom datasets, please refer to its documentation.

Checkpoints

A checkpoint trained on SemCor, WNE and Fews (both source and target language are English) is available for download.

Important: the model is released as is after its fine-tuning on the afore-mentioned datasets, without any further machinery but for beam decoding. Therefore, it may present bias issues and ethical concerns, and further mitigation strategies are needed.

Once you have downloaded the file, untar it inside the experiments/ folder.

# move file to experiments folder
mv ~/Downloads/bart-semcor-wne-fews.tar.gz experiments/
# untar
tar -xf experiments/bart-semcor-wne-fews.tar.gz -C experiments/
rm experiments/bart-semcor-wne-fews.tar.gz

Predict

There are two alternatives to perform predictions with the trained checkpoints:

• Directly through classy
• Through a Rest API

Library

As this repo is built on top of classy, we can use its functionalities to perform the prediction. First, store your data in the following format (important: the file must have exmj extension):

$ cat <path-to-input-file> | head -1 | jq
{
  "source_language": "en",
  "D": [
    [
      "dog",
      "a domestic animal"
    ]
  ],
  "target_language": "en"
}

D is a list of lemma-definition pairs (as mentioned above, for the time being, len(D) must always be 1).

Then:

$ classy predict file <path-to-ckpt> <path-to-input-file> -o <path-to-output-file> --prediction-params configurations/prediction-params/beam.yaml
$ cat <path-to-output-file> | head -1 | jq
{
  "source_language": "en",
  "D": [
    [
      "dog",
      "a domestic animal"
    ]
  ],
  "target_language": "en",
  "s": "The dog was lying on the floor, his head on his paws.",
  "phi": [
    [
      4,
      5,
      6
    ]
  ]
}

Here, s represents the generated example, while phi is a list with as many items as D such that phi[0] denotes the character positions where D[0] has been exemplified in s.

Rest API

To query via the Rest API, it's possible to either spawn the server directly or through Docker:

• Directly:

  $ classy serve <path-to-ckpt> --prediction-params configurations/prediction-params/beam.yaml
  $ curl -X 'POST' http://localhost:8000/ -H 'accept: application/json' -H 'Content-Type: application/json' -d '[{"source_language": "en", "D": [["dog", "a domestic animal"]], "target_language": "en"}]'
  [
    {
      "source_language": "en",
      "D": [
        [
          "dog",
          "a domestic animal"
        ]
      ],
      "target_language": "en",
      "s": "The dog was lying on the floor, his head on his paws.",
      "phi": [
        [
          4,
          5,
          6
        ]
      ]
    }
  ]

• Through Docker:

Acknowledgments

This work was supported by Sapienza University of Rome under the grant "Avvio alla Ricerca (Tipo 1) 2021".