/TransLIST-H22

Primary LanguagePythonApache License 2.0Apache-2.0

Steps to run TransLIST on H22

The processed data files for collection1+collection2 can be found here: link , You have to extract the zip file into V0/. You have to use the following command if you are using the processed data files python flat_main_bigram.py --status train --batch 4 --data_create yes

Run the following commands

  • cd TransLIST-H22
  • conda env create -f tlat0.yml
  • conda activate tlat0
  • bash requirements.sh [PATH to your conda environments] (e.g. ~/anaconda3/envs)
  • python embedding_gen.py
  • python ngram_lat_gen.py
  • cp fastNLP-H22_files/dataset.py [PATH to your conda environments]/home/vp.shivasan/miniconda3/envs/tlat0/lib/python3.7/site-packages/fastNLP/core (example : cp fastNLP-H22_files/dataset.py /home/vp.shivasan/miniconda3/envs/tlat0/lib/python3.7/site-packages/fastNLP/core)
  • cd V0
  • python flat_main_bigram.py --status train --batch 4 --data_create yes

TransLIST : A Transformer-Based Linguistically Informed Sanskrit Tokenizer

Official code for the paper "TransLIST : A Transformer-Based Linguistically Informed Sanskrit Tokenizer. If you use this code please cite our paper.

Requirements

  • Python 3.7.3
  • Pytorch 1.5.0
  • CUDA 9.2

Installation

Create an environment using the tlat0.yml file given in the root directory.

 conda env create -f tlat0.yml
 conda activate tlat0
 bash requirements.sh [PATH to your conda environments] (e.g. ~/anaconda3/envs)

In case you are not able to run requirements.sh; complete the following steps manually. Replace the files tester.py, dataset.py, metrics.py and crf.py as per the files given in fastnlp-copy directory inside the root directory. You can find the paths of these files at:

tlat0/lib/python3.7/site-packages/fastNLP/core --> 'tester.py', 'dataset.py', 'metrics.py'
tlat0/lib/python3.7/site-packages/fastNLP/modules/decoder --> 'crf.py'

Datasets

We have used two datasets: (1) SIGHUM (2) Hackathon

Setup

Download these required files and keep them in root directory, then run bash setup.sh

How to train Models?

We propose 2 model variants:SHR and Ngram on two datasets: SIGHUM and Hackathon. The total of 4 settings are as follows

  1. Training "SIGHUM Translist Ngram" :

    Run bash set_sighum_ngram.sh [PATH to your conda environments] in the root directory or do the following modifications manually,

    • Place SIGHUM data in sktWS folder ( from root/sktWS dir run : cp SIGHUM/* . )
    • tlat0/.../fastNLP/core/dataset.py : line 310 : setting = "sighum-ngram"
    • Use the embeddings in the SIGHUM_embeds folder. (from root dir run : cp SIGHUM_embeds/* .)

    Next:

    • From V0 directory run : python flat_main_bigram.py --status train --batch 4
    • Rename saved model in V0/saved_model

  2. Training "SIGHUM Translist SHR"

    Run bash set_sighum_shr.sh [PATH to your conda environments] in the root directory or do the following modifications manually,

    • Place SIGHUM data in sktWS folder ( from root/sktWS dir run : cp SIGHUM/* . )
    • tlat0/.../fastNLP/core/dataset.py : line 310 : setting = "sighum-shr"
    • Use the embeddings in the SIGHUM_embeds folder. (from root dir run : cp SIGHUM_embeds/* .)

    Next:

    • From V0 directory run : python flat_main_bigram.py --status train --batch 8
    • Rename saved model in V0/saved_model

  3. Training "Hackathon Translist Ngram"

    Run bash set_hack_ngram.sh [PATH to your conda environments] in the root directory or do the following modifications manually,

    • Place Hackathon data in sktWS ( from root/sktWS dir run : cp hackathon/* . )
    • tlat0/.../fastNLP/core/dataset.py : line 310 : setting = "hack-ngram"
    • Use the embeddings in the Hackathon_data/embeds folder. (from root dir run : cp Hackathon_data/embeds/* .)

    Next:

    • From V0 directory run : python flat_main_bigram.py --status train --batch 4
    • Rename saved model in V0/saved_model

  4. Training "Hackathon Translist SHR"

    Run bash set_hack_shr.sh [PATH to your conda environments] or do manually...

    • Place Hackathon data in sktWS ( from root/sktWS dir run : cp hackathon/* . )
    • tlat0/.../fastNLP/core/dataset.py : line 310 : setting = "hack-shr"
    • Use the embeddings in the Hackathon_data/embeds folder. (from root dir run : cp Hackathon_data/embeds/* .)

    Next:

    • From V0 directory run : python flat_main_bigram.py --status train --batch 8
    • Rename saved model in V0/saved_model

Inference

Note that we also call our Path Ranking for Corrupted Predictions (PRCP) module as Constrained Inference (CI) module. Please note that for all the 4 settings; we share our pretrained models in saved_models.

  1. Testing "SIGHUM Translist Ngram" :

    Run bash set_sighum_ngram.sh [PATH to your conda environments]

    • From V0 run python flat_main_bigram.py --status test --test_model best_sighum_ngram2
    • Don't run CI (not applicable in this case)

  2. Testing "SIGHUM Translist SHR" :

    Run bash set_sighum_shr.sh [PATH to your conda environments]

    • From V0 run python flat_main_bigram.py --status test --test_model best_sighum_shr2
    • Run CI from root dir : python constrained_inference.py --dataset sighum

  3. Testing "Hackathon Translist Ngram" :

    Run bash set_hack_ngram.sh [PATH to your conda environments]

    • From V0 run python flat_main_bigram.py --status test --test_model best_hack_ngram2
    • don't run CI (not applicable in this case)

  4. Testing "Hackathon Translist SHR" :

    Run bash set_hack_shr.sh [PATH to your conda environments]

    • From V0 run python flat_main_bigram.py --status test --test_model best_hack_shr2
    • run CI from root dir : python constrained_inference.py --dataset hackathon

Interactive Mode

In interactive mode, the user can provide any sentence and the associated graphml file will be generated by scraping the SHR on the fly. The graphml file will then be converted to lattice and a pretrained model best_sighum_shr2 will give its prediction on the sentence. Finally, Constrained Inference will be applied and it will output the segmented sentence. From root/V0 directory, run the following script.

python interactive_module.py --sentence="[YOUR SENTENCE HERE]" (Sentence expected to be in SLP format)

Demo example is illustrated below.

python interactive_module.py --sentence="ahaM sOBapateH senAm AyasEr BujagEr iva"
Output : 
inp_data_i :  ['ahaM', 'sOBapateH', 'senAm', 'AyasEr', 'BujagEr', 'iva']
Model prediction :  ['aham', 'sOBa_pateH', 'senAm', 'AyasEH', 'BujagEH', 'iva']
Final Segmentation :  aham sOBa pateH senAm AyasEH BujagEH iva

Citation

If you use our tool, we'd appreciate if you cite our paper:

@misc{sandhan_TransLIST,
  doi = {10.48550/ARXIV.2210.11753},
  url = {https://arxiv.org/abs/2210.11753},
  author = {Sandhan, Jivnesh and Singha, Rathin and Rao, Narein and Samanta, Suvendu and Behera, Laxmidhar and Goyal, Pawan},
  keywords = {Computation and Language (cs.CL), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {TransLIST: A Transformer-Based Linguistically Informed Sanskrit Tokenizer},
  publisher = {arXiv},
  year = {2022},
  copyright = {Creative Commons Attribution 4.0 International}
}

License

This project is licensed under the terms of the Apache license 2.0.

Acknowledgements

We build our system on top of the codebase released by Flat-Lattice.