Weakly Supervised Concept Map Generation through Task-Guided Graph Translation

• GT-D2G Model
• Prerequisites
• Data
• Reproduce Classification Results
• Training
• Case Study

---

GT-D2G Model

This is the codebase for reproducing results of TKDE paper: "Weakly Supervised Concept Map Generation through Task-Guided Graph Translation". (arXiv link)

Prerequisites

python==3.7.9

For library requirements, please refer to ./requirements.txt. (You may replace PyTorch and dgl to CPU version)

Data

Pre-processed Graphs

The NLP pipeline derived initial concept maps are available at NYT link, AMiner link, YELP link.
Put it under the project root directory and decompress it. Then three *.pickle.gz files would reside under ./data/. (No need to decompress *.pickle.gz files)

The expected ./data folder after downloading necessary resources:

./data
|-- dblp.txt     # dblp refer to AMiner corpus used in paper
|-- dblp.win5.pickle.gz
|-- nyt.txt
|-- nyt.win5.pickle.gz
|-- yelp.txt
|-- yelp.sentiment_centric.win5.pickle.gz

Pre-trained Word Embeddings

GT-D2G relies on several pre-trained word embeddings. By default, the scripts read pre-trained embeddings from ./.vector_cache folder.

• GloVe for NYT, AMiner: Download glove.840B.300d from https://nlp.stanford.edu/projects/glove/.
• Customed emb for Yelp: For yelp dataset, we get the best performan using a hybrid of GloVe and restaurant embedding, which can be download from link.

The expected ./.vector_cache folder:

./.vector_cache
|--glove.840B.300d.txt
|--glove.840B.restaurant.400d.vec

Reproduce Classification Results

Checkpoints

• GT-D2G-path: https://figshare.com/articles/dataset/GT-D2G_Data/16415802?file=30419121
• GT-D2G-neigh: https://figshare.com/articles/dataset/GT-D2G_Data/16415802?file=30419181
• GT-D2G-var: https://figshare.com/articles/dataset/GT-D2G_Data/16415802?file=30419157

Please download gziped checkpoint files using the above urls, and decompress them under ./checkpoints folder.

Example of running GT-D2G for reproducibility: sh run_test.sh.

Training

You can train your own GT-D2G by modifying provided examples run_train.sh.

Case Study

---

---

---