domain-relevance

We propose a hierarchical core-fringe learning framework to measure fine-grained domain relevance of terms – the degree that a term is relevant to a broad (e.g., computer science) or narrow (e.g., deep learning) domain.

Requirements

See requirements.txt

To install torch_geometric, please follow the instruction on pytorch_geometric

Reproduction

To reproduce the results in the paper (using word2vec embeddings)

Download data from Google Drive, unzip and put all the folders in the root directory of this repo (details about data are described below)

For broad domains (e.g., CS)

python run.py --domain cs --method cfl

For narrow domains (e.g., ML)

python run.py --domain cs --method hicfl --narrow

For narrow domains (PU setting) (e.g., ML)

python run.py --domain cs --method hicfl --narrow --pu

All experiments are run on an NVIDIA Quadro RTX 5000 with 16GB of memory under the PyTorch framework. The training of CFL for the CS domain can finish in 1 minute.

Query

To handle user query (using compositional GloVe embeddings as an example)

Download data from Google Drive, unzip and put all the folders in the root directory of this repo

Download GloVe embeddings from https://nlp.stanford.edu/projects/glove/, save the file to features/glove.6B.100d.txt

Example:

python query.py --domain cs --method cfl

The first run will train a model and save the model to model/. For the follow-up queries, the trained model can be loaded for prediction.

You can use the model either in a transductive or in an inductive setting (i.e., whether to include the query terms in training).

Options

You can check out the other options available using:

python run.py --help

Data

Data can be downloaded from Google Drive:

term-candidates/: list of seed terms. Format: term <tab> frequency

features/: features of terms (term embeddings trained by word2vec). To use compositional GloVe embeddings as features, you can download GloVe embeddings from https://nlp.stanford.edu/projects/glove/. To load the features, refer to utils.py for more details.

wikipedia/: Wikipedia search results for constructing the core-anchored semantic graph / automatic annotation

  • core-categories/: categories of core terms collected from Wikipedia. Format: term <tab> catogory <tab> ... category

  • gold-subcategories/: gold-subcategories for each domain collected from Wikipedia. Format: level#Category

  • ranking-results/: Wikipedia search results. 0 means using exact match, 1 means without exact match. Format: term <tab> result_1 <tab> ... result_k.

    The results are collected by the following script:

    # https://pypi.org/project/wikipedia/
import wikipedia
def get_wiki_search_result(term, mode=0):
    if mode==0:
        return wikipedia.search(f"\"{term}\"")
    else:
        return wikipedia.search(term)

train-valid-test/: train/valid/test split for evaluation with core terms

manual-data/:

  • ml2000-test.csv: manually created test set for ML
  • domain-relevance-comparison-pairs.csv: manually created test set for domain relevance comparison

Term lists

Several term lists with domain relevance scores produced by CFL/HiCFL are available on term-lists/

Format:

term <tab> domain relevance score <tab> core/fringe

Sample results for Machine Learning:

term-list-ml

Citation

The details of this repo are described in the following paper. If you find this repo useful, please kindly cite it:

@inproceedings{huang2021measuring,
  title={Measuring Fine-Grained Domain Relevance of Terms: A Hierarchical Core-Fringe Approach},
  author={Huang, Jie and Chang, Kevin Chen-Chuan and Xiong, Jinjun and Hwu, Wen-mei},
  booktitle={Proceedings of ACL-IJCNLP},
  year={2021}
}

GitHub

https://github.com/jeffhj/domain-relevance