This is the source code for our arXiv User-Guided Aspect Classification for Domain-Specific Texts.

The preprint version can be accessed at: https://arxiv.org/abs/2004.14555

• Model
• Training
  • Required Inputs
  • Dependencies
  • Command
• Citation

• Raw Corpus TXT File for Training Embedding
  • Example: restaurant/restaurant_corpus.linked.txt
    • One review per line.
    • Must be named as dataset_corpus.linked.txt.
• Training CSV File
  • Example: restaurant/restaurant_train.csv
    • Only one column.
    • One review per line.
    • Must be named as dataset_train.csv.
• K Aspect Test CSV File
  • Example: restaurant/restaurant_test.csv
    • Two columns (segment, aspect) per line.
    • Aspects NOT including miscellaneous aspect.
    • Must be named as dataset_test.csv.
• K+1 Aspect Test CSV File
  • Example: restaurant/restaurant_test_kplus.csv
    • Two columns (segment, aspect) per line.
    • Aspects including miscellaneous aspect.
    • Must be named as dataset_test_kplus.csv
• Word Embedding File
  • Example: restaurant/restaurant.200d.txt
    • The word embedding files have been provided but you can also train your own.
    • Must be named as dataset.200d.csv
• Aspects TXT File
  • Example: restaurant/restaurant_aspects.txt
    • One aspect per line.
    • Must be named as dataset_aspects.txt
• Seed Words Dictionary
  • Example: restaurant/restaurant_seeds.txt
    • Two columns (seedword, aspect) per line.
    • Separated by tab.
    • Must be named as dataset_seeds.txt

This project is based on python==3.6.9. The dependencies are as follow:

pytorch==1.3.1
torchtext==0.4.0
scikit-learn==0.22
spacy==2.2.3
scipy=1.4.1
gensim==3.8.1
numpy==1.17.4

To train the word2vec embedding on the specific corpus, please run python embedding.py --dataset dataset_name.

To train ARYA, please run python train.py --dataset dataset_name --quantile quantile_to_decide_hnorm_threshold --score_threshold score_threshold_for_seed_selection --seedword_limit max_num_seedwords --no_filtering --no_tuning

The --quantile is the for deciding hnorm threshold for computing Pmisc.

The --score_threshold is the threshold for seed word selection in the seed expansion process.

The --no_filtering and --no_tuning are two options for ablation studies.

Please cite our paper if you are using our framework. Thank you!

@article{li2020user,
  title={User-Guided Aspect Classification for Domain-Specific Texts},
  author={Li, Peiran and Guo, Fang and Shang, Jingbo},
  journal={arXiv preprint arXiv:2004.14555},
  year={2020}
}