OpenBG-CSK

Usage

Experiments

To reproduce the experiments, run python run.py to run.

F1 score, accuracy and the detailed results will be printed.

Training models

Running shell files: bash run.sh, and the contents of shell files are as follow:

DATA_DIR="data"

MODEL_DIR="bert-base-chinese"
OUTPUT_DIR="output/save_dict/"
PREDICT_DIR="data/"
MAX_LENGTH=128

python run.py \
    --data_dir=${DATA_DIR} \
    --model_dir=${MODEL_DIR} \
    --output_dir=${OUTPUT_DIR} \
    --do_train=True \
    --max_length=${MAX_LENGTH} \
    --batch_size=16 \
    --learning_rate=1e-5 \
    --epochs=10 \
    --seed=2021

Inference & generation of results

Running shell files: bash run.sh predict, and the contents of shell files are as follows:

DATA_DIR="data"

MODEL_DIR="bert-base-chinese"
OUTPUT_DIR="output/save_dict/"
PREDICT_DIR="data/"
MAX_LENGTH=128

python run.py \
    --data_dir=${DATA_DIR} \
    --model_dir=${MODEL_DIR} \
    --output_dir=${OUTPUT_DIR} \
    --do_train=False \
    --max_length=${MAX_LENGTH} \
    --batch_size=16 \
    --learning_rate=1e-5 \
    --epochs=10 \
    --seed=2021

Detailed Results

The sample of dataset are putted in the data repo: train_triple.jsonl,dev_triple.jsonl.

In train_triple.jsonl, the data format is {"triple_id" : "0579","subject":"瓶装水","object":"跑步","predicate":"品类_适用_场景","salience": 0}.

In dev_triple.jsonl, the data format is {"triple_id":"0579","subject":"瓶装水","object":"跑步","predicate":"适用"}.

Benchmark

We evaluate the several models on the experiment dataset. We use KG-BERT as the base model and report the baselines of the task.

Model	F1	Acc.
BERT-base	55.2	55.8
RoBERTa-wwm-ext-base	56.9	57.2

KG-BERT takes texts of h, r, t as input of bidirectional encoder such as BERT and computes scoring function of the triple with language model. In specific, the input of model is the concatenation of h, r, t, as [CLS] h [SEP] r [SEP] t [SEP]. The final hidden state C corresponding to [CLS] and the classification layer weights W are used to calculate the triple score.

We list hyper-parameters during the baseline experiments.

Common hyper-parameters

Param	Value
weight_decay	0.01
learning_rate	1e-5