Adaptive Ensembling: Unsupervised Domain Adaptation for Political Document Analysis

Code and datasets for our EMNLP 2019 paper "Adaptive Ensembling: Unsupervised Domain Adaptation for Political Document Analysis". If you found this project helpful, please consider citing our paper:

@inproceedings{desai2019unsupervised,
  author={Desai, Shrey and Sinno, Barea and Rosenfeld, Alex and Junyi Li, Jessy},
  title={Unsupervised Domain Adaptation for Political Document Analysis},
  booktitle={Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing},
  year={2019},
}

Datasets

Our Corpus of Contemporary American English (COCA) annotations are available in datasets/coca_train.csv and datasets/coca_test.csv, where they denote the train and test sets, respectively. Both files consist of the following structure:

docid,ag,pe,ir
152020,1,0,1
203568,0,0,0
164937,1,1,0

The docid field is the unique identifier given to each document in the COCA dataset. American Government (ag), Political Economy (pe), and International Relations (ir) are the category annotations we provide. Here, a 0 indicates the document does not have this label while a 1 indicates the document does have this label. Similarly, if a document does not have a 1 in any category, then it is labeled as non-political.

Next, we also provide a list of document identifiers that are labeled as "political" by our domain adaptation framework. This is available in datasets/coca_politics.txt.

For more information on our annotation process, label information, and domain adaptation methods, please see our paper for an in-depth discussion.