Knowledge Graph Generation From Text

Description

Grapher is an end-to-end multi-stage Knowledge Graph (KG) construction system, that separates the overall generation process into two stages.

The graph nodes are generated first using pretrained language model, such as T5.The input text is transformed into a sequence of text entities. The features corresponding to each entity (node) is extracted and then sent to the edge generation module.

Edge construction, using generation (e.g.,GRU) or a classifier head. Blue circles represent the features corresponding to the actual graph edges (solid lines) and the white circles are the features that are decoded into ⟨NO_EDGE⟩ (dashed line).

Environment

Install miniconda and create a conda environment with the following command, which also sets the python version: conda create -n data_science python=3.10.6 Activate your conda environment with conda activate data_science Install the requirements.txt file pip install -r requirements.txt

Setup

Install dependencies

# clone project   
git clone [email protected]:GreenfishK/Grapher.git

# clone another external repositories for scoring the results
cd Grapher/src
git submodule add https://github.com/WebNLG/WebNLG-Text-to-triples.git WebNLG_Text_to_triples

Data

We pull the WebNLG 3.0 dataset from hugging face. It is done automatically in the script datasets.py.

How to train

If you run your training on a single, low resource GPU, set <script_set> to single_node_ti.If you run your training on two a40 GPUs set <script_set> to single_node_a40 If you run your training on two nodes, each with two a40 GPUs, using SLURM, set <script_set> to hpc. Feel free to change the parameters in the sripts.

# naviagate to scripts directory
cd src/scripts/<script_set>

There are two scripts to run two variantes of the model:

run Grapher with the edge generation head

Description from the paper: The advantage of generation is the ability to construct any edge sequence, including ones unseen during training, at the risk of not matching the target edge token sequence exactly.

./train_gen.sh

run Grapher with the classifier edge head

Description from the paper: if the set of possible relationships is fixed and known, the classification head is more efficient and accurate

./train_class.sh

How to test

Same logic as for training, just the the scripts are named test_class.sh and test_gen.sh in each of the script set directories.

Visualize Training Results

Results can be visualized in Tensorboard

tensorboard --logdir output --load_fast=false

How to run inference

# run inference classifier edge head (version 1) or alternatively edge generation head (version 2) using latest checkpoint last.ckpt
source .env && python3 main.py --run inference --version ${MODEL_VERSION} --default_root_dir ${STORAGE_DRIVE}/data/core/grapher/output --inference_input_text "Danielle Harris had a main role in Super Capers, a 98 minute long movie."

Citation

@inproceedings{grapher2022,
  title={Knowledge Graph Generation From Text},
  author={Igor Melnyk, Pierre Dognin, Payel Das},
  booktitle={Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (Findings of EMNLP)},
  year={2022}
}

Reproducibility study

• The scripts ./train_gen and ./train_class.sh both produce an error during execution due to several non-existing arguments passed during the call of main.py (see below).

main.py: error: unrecognized arguments: --max_epochs 100 --accelerator gpu --num_nodes 1 --num_sanity_val_steps 0 --fast_dev_run 0 --overfit_batches 0 --limit_train_batches 1.0 --limit_val_batches 1.0 --limit_test_batches 1.0 --accumulate_grad_batches 10 --detect_anomaly True --log_every_n_steps 100 --val_check_interval 1000

• Some additional libraries are required, such as SentencePie(See below).

T5Tokenizer requires the SentencePiece library but it was not found in your environment. Checkout the instructions on the installation page of its repo: https://github.com/google/sentencepiece#installation and follow the ones that match your environment. Please note that you may need to restart your runtime after installation.

• Instead of just cloning the two separately needed git repositories corpusreader and WebNLG-Text-to-triples, I added them as submodules (See Setup section) and included them in gitignore in order to not track the pycache changes.

• I added the parameter cache_dir to point to a directory with enough space, because the model training requires 2950.74 MB for T5.

• I updated two function names and interfaces in litgrapher.py (now called grapher_lightning.py) due to incompatibilities with pytorch_lightning >= 2.0.0:

validation_epoch_end -> on_validation_epoch_end
test_epoch_end -> on_test_epoch_end