aqweteddy / LeverageJustAFewKeywords

Unofficial implementation for EMNLP2019 <Leveraging Just a Few Keywords for Fine-Grained Aspect Detection Through Weakly Supervised Co-Training>

Leveraging Just a Few Keywords for Fine-Grained Aspect Detection Through Weakly Supervised Co-Traing

• original paper
• I can't find the official implementation or any unofficial implementation.

Dataset

• OPOSUM

requirements

• pytorch >= 1.5
• numpy
• h5py
• click

How to Run

preprocess data

• preprocess data follow OPOSUM
  • hdf5 (train data and test data)
  • seed words
• extract preprocessed hdf5 data through through extract_data.py

python extract_data.py --source oposum/data/preprocessed/BOOTS.hdf5 --output data/boots_train.json
python extract_data.py --source oposum/data/preprocessed/BOOTS_TEST.hdf5 --output data/boots_test.json

train

• you can set config in config.py or using arguments.
  • notice that the general aspect index is not same in every datasets.
• start training
  • seed_words is no weight.

python parser.py --train_file ./data/boots_train.json --test_file ./data/boots_test.json --save_dir ./ckpt/boots --aspect_init_file ./data/seed_words.txt --epochs 3

• python parser.py --help to see detail.

Benchmark

OPOSUM

• Bags: 0.59 (RandomSampler 50000 data run 3 epochs.)
• TV:

Some difference between the paper and this implementation

RandomSampler

• random sample 50000 data every epochs.

Teacher

• In paper sec.(3.1),

If no seed word appears in s, then the teacher predicts the "General" aspect by setting $q_i^k = 1$

but in this implementation,

If no seed word appears in s, I let the teacher predicts like the seedword of general aspect appear one time.