DetIE: Multilingual Open Information Extraction Inspired by Object Detection

This repository contains the code for the paper DetIE: Multilingual Open Information Extraction Inspired by Object Detection by Michael Vasilkovsky, Anton Alekseev, Valentin Malykh, Ilya Shenbin, Elena Tutubalina, Dmitriy Salikhov, Mikhail Stepnov, Andrei Chertok and Sergey Nikolenko.

Disclaimers

All the results have been obtained using V100 GPU with CUDA 10.1.

Preparations

Download the files bundle from here. Each of them should be put into the corresponding directory:

1. folder version_243 (DetIE_LSOIE) should be copied to: results/logs/default/version_243;
2. folder version_263 (DetIE_IMoJIE) should be copied to: results/logs/default/version_263;
3. files imojie_train_pattern.json, lsoie_test10.json and lsoie_train10.json should be copied to data/wikidata.

We suggest that you use the provided Dockerfile to deal with all the dependencies of this project.

E. g. clone this repository, then

cd DetIE/
docker build -t detie .
nvidia-docker run  -p 8808:8808 -it detie:latest bash

Once this docker image starts, we're ready for work.

Taking a minute to read the configs

This project uses hydra library for storing and changing the systems' metadata. The entry point to the arguments list that will be used upon running the scripts is the config/config.yaml file.

defaults:
  - model: detie-cut
  - opt: adam
  - benchmark: carb

model leads to config/model/... subdirectory; please see detie-cut.yaml for the parameters description.

opt/adam.yaml and benchmark/carb.yaml are the examples of configurations for the optimizer and the benchmark used.

If you want to change some of the parameters (e.g. max_epochs), not modifying the *.yaml files, just run e.g.

PYTHONPATH=. python some_..._script.py model.max_epochs=2

Training

PYTHONPATH=. python3 modules/model/train.py

Inference time

PYTHONPATH=. python3 modules/model/test.py model.best_version=243

This yields time in seconds when running inference against modules/model/evaluation/oie-benchmark-stanovsky/raw_sentences/all.txt using batch size equal to 32.

Should be 708.6 sentences/sec. on NVIDIA Tesla V100 GPU.

Evaluation

English sentences

To apply the model to CaRB sentences, run

cd modules/model/evaluation/carb-openie6/
PYTHONPATH=<repo root> python3 detie_predict.py
head -5 systems_output/detie243_output.txt

This will save the predictions into the modules/model/evaluation/carb-openie6/systems_output/ directory. The same should be done with modules/model/evaluation/carb-openie6/detie_conj_predictions.py.

To reproduce the DetIE numbers from the Table 3 in the paper, run

cd modules/model/evaluation/carb-openie6/
./eval.sh

• detie243 is a codename for DetIE_{LSOIE}
• detie243conj is a codename for DetIE_{LSOIE} + IGL-CA
• detie263 is a codename for DetIE_{IMoJIE}
• detie263conj is a codename for DetIE_{IMoJIE} + IGL-CA

Synthetic data

To generate sentences using Wikidata's triplets, one can run the scripts

PYTHONPATH=. python3 modules/scripts/data/generate_sentences_from_triplets.py  wikidata.lang=<lang> 
PYTHONPATH=. python3 modules/scripts/data/download_wikidata_triplets.py  wikidata.lang=<lang>

Cite

Please cite the original paper if you use this code.

@inproceedings{Vasilkovsky2022detie,
   author    = {Michael Vasilkovsky, Anton Alekseev, Valentin Malykh, Ilya Shenbin, Elena Tutubalina, 
               Dmitriy Salikhov, Mikhail Stepnov, Andrei Chertok and Sergey Nikolenko},
   title     = {{DetIE: Multilingual Open Information Extraction Inspired by Object Detection}},
   booktitle = {
       {Proceedings of the 36th {AAAI} Conference on Artificial Intelligence}
   },
   year      = {2022}
 }

Contact

Michael Vasilkovsky waytobehigh (at) gmail (dot) com