Code in this repo generated the models described in our ICWSM 2020 paper:
Hemphill, L., & Schöpke-Gonzalez, A. M. (2020). Two Computational Models for Analyzing Political Attention in Social Media. Proceedings of the AAAI Conference on Web and Social Media, 14(1), 260–271. https://www.aaai.org/ojs/index.php/ICWSM/article/view/7297
This directory provides the data that we used to train models and that we made predictions on with models. Due to Twitter terms of service limitations, we did not shared raw tweet data in this public repo. Instead, you will find our processed dataset of the 115th Congress' tweets at data/congress_115/congress115.csv.gz, with the file called data/congress_115/congress115_codebook.csv explaining column labels for this dataset.
Russell's raw data is not publicly available at this time. Processed datasets which we use to train and test our best-performing supervised models are available at data/russell/russell_processed_0.9.csv.gz. The file data/russell/russell_liwc_w2v_both.csv.gz describes the tweetset that we used for training and testing supervised classifiers with word2vec and LIWC2015 features added. See the following two papers for additional details on Russell's raw data:
Russell, A. (2017). U.S. Senators on Twitter: Asymmetric Party Rhetoric in 140 Characters. American Politics Research, 46(4). https://doi.org/10.1177/1532673X17715619
Russell, A. (2018). The Politics of Prioritization: Senators’ Attention in 140 Characters. The Forum, 16(2), 331–356. https://doi.org/10.1515/for-2018-0020
The directory data/unsupervised/model_source provides data that used to train our unsupervised model. The directory data/unsupervised/model_output provides data that our unsupervised models produced. The file data/unsupervised/model_output/unsup_codebook.csv provides annotator labels for unsupervised topics returned by our model. Note that though our model produced 50 topics, annotators grouped some of these topics together due to thematic similarity, resulting in a total of 35 unsupervised topics.
This directory provides jupyter notebooks that walk a user through the steps we took extract LIWC (liwc) and word2vec (w2v) features for use in classifier training and testing. We provide this code for transparency and reproducibility. The final datasets created by this code are available in the parent directory's "data" and "models" folders. Intermediate datasets produced by the code are not available. You will need to re-run the code to produce these. For those that would like to reproduce processes, we provide descriptions of each notebook's function here:
Notebook Name:
notebooks/models/create/feature_extraction/1.0-ams-build-original-w2v-model.ipynb
Notebook Function: Builds a word2vec vocabulary from Russell's tweets and trains an original word2vec model
Notebook Name:
notebooks/models/create/feature_extraction/2.0-ams-w2v-featureextraction-original.ipynb
Notebook Function: Extract original word2vec features from Russell's tweets using model from previous notebook
Notebook Name:
notebooks/models/create/feature_extraction/3.0-ams-w2v-featureextraction-pretrained.ipynb
Notebook Function: Extract pretrained word2vec features from Russell's tweets using Frederic Godin's model. The model is originally created by Fréderic Godin, but the creator's link to the w2v model is no longer active. You will need to download
word2vec_twitter_model.binfrom the link toexternal_repositories/word2vec-twitterin order to run this notebook.
This directory provides jupyter notebooks that walk a user through the steps we took to train and test supervised classifiers. For those that would like to reproduce processes, we provide descriptions of each notebook's function here:
Notebook Name:
notebooks/models/create/supervised/1.0-ams-create-best-classifiers.ipynb
Notebook Function: Creates best performing dummy, naive bayes, and logistic regression classifiers
Notebook Name:
notebooks/models/create/supervised/2.0-ams-create-w2vliwc-classifiers.ipynb
Notebook Function: Creates logistic regression and support vector machine classifiers with word2vec and LIWC2015 features
This directory provides a jupyter notebook that walks a user through the steps we took to build our unsupervised model. For those that would like to reproduce processes, we provide a description of this notebook's function here:
Notebook Name:
notebooks/models/create/unsupervised/1.0-ams-unigram-lda-model-MALLET.ipynbNotebook Function: Train unsupervised model (NOTE that some of this code must be implemented in a command line interface)
This directory provides a Jupyter notebook that walks a user through the steps we took to compare the outputs of our supervised and unsupervised models. For those that would like to reproduce processes, we provide a description of this notebook's function here:
Notebook Name:
notebooks/models/run/1.0-ams-label-comparisons.ipynbNotebook Function: Load supervised classifiers, label tweets with supervised classifier predictions, perform comparison between supervised and unsupervised model labeling patterns
This directory provides the best performing supervised and unsupervised models that we discuss in our paper under models/best. It also provides the original w2v vocabulary developed from Russell's tweetset in models/feature_extraction.
This directory provides raw figure and table files used in our paper.
This folder provides external repositories used in developing models, including the code necessary to use Frederic Godin's pretrained word2vec model's vocab and MALLET's LDA model wrapper. You will need to download word2vec_twitter_model.bin from https://www.kaggle.com/hachemsfar/cbigru/version/1 (i.e., Godin's vocabulary) to external_repositories/word2vec-twitter in order to run this word2vec code.