Elamur Research⌕
☰
home / papers / 2014 / Political Ideology Detection Using Recursive Neural Networks

Political Ideology Detection Using Recursive Neural Networks

Summary

The paper tackles sentence-level detection of political ideology (liberal vs. conservative in the U.S. sense). Earlier approaches relied on bags of words or hand-built wordlists and largely ignored syntax. The authors argue that bias often appears only when phrases combine: scare quotes or framing like "a lie about X" can flip the apparent leaning of loaded terms. Building on compositional sentiment models, they apply a supervised recursive neural network (RNN). The RNN builds phrase vectors bottom-up over a parse tree and classifies each node with a softmax layer.

They use two corpora:

  • Convote, with party labels propagated down to individual sentences.
  • The Ideological Books Corpus (IBC), filtered toward likely-biased sentences and newly crowdsourced with phrase- and sentence-level ideology labels.

The RNN outperforms existing baselines on both datasets. It gains further when trained with phrase-level annotations, and it handles syntactically complex cases correctly. The study shows that ideology detection, like sentiment, benefits from modeling composition, and it contributes a reusable fine-grained annotated resource.

Key points

  • Model:
    • Each parent vector is a tanh of left and right composition matrices applied to its children, plus a bias.
    • A softmax classifier is applied at every node, trained with a regularized cross-entropy loss.
    • Optimization uses L-BFGS with parameter averaging and backpropagation through structure.
  • Initialization: two schemes are compared.
    • Random initialization.
    • 300-dimensional word2vec skip-gram vectors (trained on about 100 billion Google News words), with W_L = W_R = 0.5I so that parents start as the average of their children.
  • Convote: 2005 U.S. Congressional floor debates. Labels are mapped Democrat → liberal and Republican → conservative. After bias filtering and party balancing, 7,816 sentences remain.
  • IBC (Gross et al., 2013):
    • Books and magazine articles from 2008–2012, labeled by experts with coarse (left/right/center) and fine-grained ideologies.
    • Out of over a million sentences, filtering keeps a subset of 55,932.
  • Bias filter: a sentence is kept if it either contains words from selected LIWC categories (e.g., Negative Emotion, Anger, Kill verbs) or contains one of the top-100 party-associated "sticky bigrams" ranked by log-likelihood ratio. The final pool is the union of these two subsets.

Interesting ideas

  • I1. Separating partisanship from ideology. Author-level party labels are only a proxy for ideology, which motivated dedicated ideology annotation. This is a good reminder to check whether distant labels actually measure the target concept.
  • I2. Cheap pre-filtering before annotation. Running lexicon categories and discriminative bigrams over a mostly neutral corpus raises the share of informative sentences before paying for crowdsourcing. The same trick works for any rare-phenomenon annotation project.
  • I3. Averaging-based initialization. Setting both composition matrices to half the identity means untrained phrase vectors start as averages of pretrained word vectors. This keeps the embedding geometry intact at the start of training and is simple to apply to any tree-structured composition model.
  • I4. Supervision at every tree node. Labeling and supervising phrases as well as whole sentences gives the model signal about where polarity shifts occur. Phrase-level labels measurably helped here, which suggests the approach for other compositional tasks such as stance or framing.
  • I5. Ranking n-grams with the trained model. Listing the phrases the trained model finds most liberal or most conservative serves as a qualitative sanity check and makes a neural classifier easier to interpret for domain experts.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Natural language processing

Cited by

93 papers of our library cite it (of the 140884 whose bibliographies we read); 7 of them have a page here, 86 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.3115/v1/p14-1105

@inproceedings{iyyer2014political,
  title = {{Political Ideology Detection Using Recursive Neural Networks}},
  author = {Iyyer, Mohit and Enns, Peter and Boyd-Graber, Jordan and Resnik, Philip},
  booktitle = {Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2014},
  doi = {10.3115/v1/p14-1105},
  url = {https://doi.org/10.3115/v1/p14-1105}
}

References

32 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Amr Ahmed and Eric P Xing. 2010. Staying informed: supervised and semi-supervised multi-view topical analysis of ideological perspective. In EMNLP. find
  2. Peter F Brown, Peter V Desouza, Robert L Mercer, Vincent J Della Pietra, and Jenifer C Lai. 1992. Class-based n-gram models of natural language. Comp. Ling., 18(4):467–479. find
  3. Royce Carroll, Jeffrey B Lewis, James Lo, Keith T Poole, and Howard Rosenthal. 2009. Measuring bias and uncertainty in dw-nominate ideal point estimates via the parametric bootstrap. Political Analysis, 17(3):261–275. find
  4. Ronan Collobert and Jason Weston. 2008. A unified architecture for natural language processing: Deep neural networks with multitask learning. In ICML. find
  5. Frank E Dardis, Frank R Baumgartner, Amber E Boydstun, Suzanna De Boef, and Fuyuan Shen. 2008. Media framing of capital punishment and its impact on individuals’ cognitive responses. Mass Communication & Society, 11(2):115– 140. find
  6. Matthew Gentzkow and Jesse M Shapiro. 2010. What drives media slant? evidence from us daily newspapers. Econometrica, 78(1):35–71. find
  7. Sean Gerrish and David M Blei. 2011. Predicting legislative roll calls from text. In ICML. find
  8. Christoph Goller and Andreas Kuchler. 1996. Learning task-dependent distributed representations by backpropagation through structure. In Neural Networks, 1996., IEEE International Conference on, volume 1. find
  9. Stephan Greene and Philip Resnik. 2009. More than words: Syntactic packaging and implicit sentiment. In NAACL. find
  10. Tim Groseclose and Jeffrey Milyo. 2005. A measure of media bias. The Quarterly Journal of Economics, 120(4):1191– 1237. find
  11. Justin Gross, Brice Acree, Yanchuan Sim, and Noah A Smith. 2013. Testing the etch-a-sketch hypothesis: A computational analysis of mitt romney’s ideological makeover during the 2012 primary vs. general elections. In APSA 2013 Annual Meeting Paper. find
  12. Kazuma Hashimoto, Makoto Miwa, Yoshimasa Tsuruoka, and Takashi Chikayama. 2013. Simple customization of recursive neural networks for semantic relation classification. In EMNLP. find
  13. Karl Moritz Hermann and Phil Blunsom. 2013. The Role of Syntax in Vector Space Models of Compositional Semantics. In ACL. find
  14. George Lakoff. 2002. Moral Politics: How Liberals and Conservatives Think, Second Edition. University of Chicago Press. find
  15. Wei-Hao Lin, Eric Xing, and Alexander Hauptmann. 2008. A joint topic and perspective model for ideological discourse. In Machine Learning and Knowledge Discovery in Databases, pages 17–32. Springer. find
  16. Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781. find
  17. Viet-An Nguyen, Jordan Boyd-Graber, and Philip Resnik. 2013. Lexical and hierarchical topic regression. In NIPS, pages 1106–1114. find
  18. David Niven. 2003. Objective evidence on media bias: Newspaper coverage of congressional party switchers. Journalism & Mass Communication Quarterly, 80(2):311–326. find
  19. Bo Pang and Lillian Lee. 2008. Opinion mining and sentiment analysis. Foundations and trends in information retrieval, 2(1-2). find
  20. James W. Pennebaker, Martha E. Francis, and Roger J. Booth. 2001. Linguistic inquiry and word count [computer software]. Mahwah, NJ: Erlbaum Publishers. find
  21. Marta Recasens, Cristian Danescu-Niculescu-Mizil, and Dan Jurafsky. 2013. Linguistic models for analyzing and detecting biased language. find our card
  22. Asad B Sayeed, Jordan Boyd-Graber, Bryan Rusk, and Amy Weinberg. 2012. Grammatical structures for word-level sentiment detection. In NAACL. find
  23. Burr Settles. 2011. Closing the loop: Fast, interactive semi-supervised annotation with queries on features and instances. In EMNLP. find
  24. Yanchuan Sim, Brice Acree, Justin H Gross, and Noah A Smith. 2013. Measuring ideological proportions in political speeches. In EMNLP. find
  25. Richard Socher, Eric H. Huang, Jeffrey Pennington, Andrew Y. Ng, and Christopher D. Manning. 2011a. Dynamic Pooling and Unfolding Recursive Autoencoders for Paraphrase Detection. In NIPS. find
  26. Richard Socher, Jeffrey Pennington, Eric H. Huang, Andrew Y. Ng, and Christopher D. Manning. 2011b. Semi-Supervised Recursive Autoencoders for Predicting Sentiment Distributions. In EMNLP. find our card
  27. Richard Socher, John Bauer, Christopher D. Manning, and Andrew Y. Ng. 2013a. Parsing With Compositional Vector Grammars. In ACL. find
  28. Richard Socher, Alex Perelygin, Jean Y Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013b. Recursive deep models for semantic compositionality over a sentiment treebank. In EMNLP. find
  29. Matt Thomas, Bo Pang, and Lillian Lee. 2006. Get out the vote: Determining support or opposition from Congressional floor-debate transcripts. In EMNLP. find our card
  30. Janyce Wiebe, Theresa Wilson, Rebecca Bruce, Matthew Bell, and Melanie Martin. 2004. Learning subjective language. Comp. Ling., 30(3):277–308. find our card
  31. Theresa Wilson, Janyce Wiebe, and Paul Hoffmann. 2005. Recognizing contextual polarity in phrase-level sentiment analysis. In EMNLP. find our card
  32. Tae Yano, Philip Resnik, and Noah A Smith. 2010. Shedding (a thousand points of) light on biased language. In Proceedings of the NAACL HLT 2010 Workshop on Creating Speech and Language Data with Amazon’s Mechanical Turk, pages 152–158. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/iyyer-2014-political-ideology-detection-recursive-neural/