W2CSpase for Interpretable Language Modeling

Official code for "Constructing Word-Context-Coupled Space Aligned with Associative Knowledge Relations for Interpretable Language Modeling", Findings in ACL2023.

You can find our paper on Link.

Cite as: Fanyu Wang and Zhenping Xie. 2023. Constructing Word-Context-Coupled Space Aligned with Associative Knowledge Relations for Interpretable Language Modeling. In Findings of the Association for Computational Linguistics: ACL 2023, pages 8414–8427, Toronto, Canada. Association for Computational Linguistics.

Contact Information

Fanyu Wang: Personal Page

Zhenping Xie: Personal Page

ColeGroup (Chinese Only): Home Page

Lauchment

Our default config are stored in ./config/config.py

You can customize the settings in config.py, where "mode" refer to sentiment classification and spelling correction tasks.

0. initial_akn.py

You can use your personal dataset for AKN initialization or you can find our AKN weight in ./akn/akn_download.txt.

1. model_initial.py

Initialization py file for finetuning BERT model and training for mapping network.

2. context_cluster.py

Clustering py file for abstraction of context-level semantics.

3. senti_classify.py / correction.py

Tasks completion py files for sentiment classification and spelling correction tasks.

Datasets

CHNST

CHNST dataset for sentiment classification task. We do preprocessing operation.

SIGHAN15

SIGHAN15 dataset for spelling correction task. Evaluation only.

Weibo

Weibo dataset for sentiment classification task. We do preprocessing operation.

trainset_download.txt

We upload our preprocessed training dataset on Google Drive. The trainset are constructed based on HyBird and SIGHAN15-Trainset.