NL4Opt Subtask 2

This repository contains the source code of "Named Entity-Based Enrichment with BART" for the second subtask of NL4Opt. Refer to the NL4Opt page for details about the subtask.

This repository is based on the baseline code for the second subtask which can be found at the NL4Opt's official implementation. Check their README.md for details.

Environment Setup

Use environment.yml to setup the environment:

conda env create -f environment.yml -n <ENV_NAME>
conda activate <ENV_NAME>

Verify that it was installed:

conda env list

Running the Pipeline

Download the Dataset

First, copy the dataset files train.jsonl and dev.jsonl to the data subdirectory. The dataset files can be found under the generation_task subdirectory in the dataset repository.

Training Configurations

The config files for training are present in the configs subdirectory.

• baseline.json: Config for the baseline model.
• default.json: Config for our approach that we used for the final submission.

Running the Pipeline

The training and testing pipeline can be run using train_and_evaluate.sh. This script expects Miniconda installed under ~/miniconda3/ and test.jsonl present under the data subdirectory.

bash train_and_evaluate.sh

Training and Evaluating the Model

To train the model, run the following:

python train.py --config configs/default.json --seed 42

The important parameters in the training are:

• use_copy uses a copy mechanism that computes $P_\text{copy}$ over the input tokens.
• per_declaration controls each training data sample to correspond to a single declaration of a given LP problem instead of the entire formulation (i.e. all declarations in the problem).
• enrich_ner controls if the named entity information should be added to the input before feeding it to the model.

To evaluate the model, run the following:

python test.py --gpu <gpu id> --checkpoint <checkpoint.mdl> --test-file <test.jsonl> --batch-size <test_batch_size> --beam-size <beam_size>

Contact

For any queries, feel free to reach out to gangwar2 [at] illinois [dot] edu.