Bootstrapping Relation Extractors

Implementation of "Bootstrapping Relation Extractors using Syntactic Search by Examples".

Classification

Classification and Evaluation

You can find how to run the classification and evluation script in run_classification.sh.

CMD:

bash run_classification.sh

Generation

{"task": ["tacred"], "training_method": ["generation"], "relation_name": ["org:founded_by"], "num_positive_examples": [100], "ratio_negative_examples": [10], "seed": [1,2,3], "logging_steps": [100]}

Generation

Here I'm mostly using modified scripts of huggingface's transformers.

Preprocessing

In order to create the trainable examples run

python preprocess/create_tacred_datafiles.py --file_path ../datasets/tacred/data/json/train.json --save_to_file data/tacred/for_generation/train --src_and_tgt_one_file_with_go

Finetune

You should finetune on your dataset using a run_lm_finetuning.py or an easy to use bash script similar to the one used for TACRED tacred_generation.sh. This file is also an example of the arguments you should pass run_lm_finetuning.py.

Generation

After finetuning, pass the model alongside different hyperparameters to run_generation.py. This should also recieve a sentence in the prompet like the following: William married Kate Middleton. <|GO|>. Again, you can find an example of the arguments in the corresponding bash script tacred_generation.