## A Joint Topic and Perspective Model for Ideological Discourse

Citations: | 9 - 2 self |

### BibTeX

@MISC{Lin_ajoint,

author = {Wei-hao Lin and Eric Xing and Er Hauptmann},

title = {A Joint Topic and Perspective Model for Ideological Discourse},

year = {}

}

### Years of Citing Articles

### OpenURL

### Abstract

Abstract. Polarizing discussions on political and social issues are common in mass and user-generated media. However, computer-based understanding of ideological discourse has been considered too difficult to undertake. In this paper we propose a statistical model for ideology discourse. By ideology we mean “a set of general beliefs socially shared by a group of people. ” For example, Democratic and Republican are two major political ideologies in the United States. The proposed model captures lexical variations due to an ideological text’s topic and due to an author or speaker’s ideological perspective. To cope with the non-conjugacy of the logistic-normal prior we derive a variational inference algorithm for the model. We evaluate the proposed model on synthetic data as well as a written and a spoken political discourse. Experimental results strongly support that ideological perspectives are reflected in lexical variations. 1

### Citations

2392 | Latent Dirichlet allocation
- BLEI, NG, et al.
- 2002
(Show Context)
Citation Context ... which may not be true and could benefit from an extra story alignment step as [12] did. We borrow statistically modeling and inference techniques heavily from research on topic modeling (e.g., [14], =-=[15]-=- and [16]). They focus mostly on modeling text collections that containing many different (latent) topics (e.g., academic conference papers, news articles, etc). In contrast, we are interested in mode... |

834 | An Introduction to Variational Methods for Graphical Models
- Jordan, Ghahramani, et al.
- 1999
(Show Context)
Citation Context ...f τ and {φv}, however, are computationally intractable because of the non-conjugacy of the logistic-normal prior. We thus approximate the posterior probability distribution using a variational method =-=[4]-=-, and estimate the parameters using variational expectation maximization [5]. By the Generalized Mean Field Theorem (GMF) [6], we can approximate the joint posterior probability distribution of τ and ... |

792 | Probabilistic latent semantic indexing
- Hofmann
- 1999
(Show Context)
Citation Context ...issue, which may not be true and could benefit from an extra story alignment step as [12] did. We borrow statistically modeling and inference techniques heavily from research on topic modeling (e.g., =-=[14]-=-, [15] and [16]). They focus mostly on modeling text collections that containing many different (latent) topics (e.g., academic conference papers, news articles, etc). In contrast, we are interested i... |

735 |
Foundations of statistical natural language processing
- Manning, Schütze
- 1999
(Show Context)
Citation Context ...tive model predicted words from unseen ideological discourse in terms of perplexity on a held-out set. Perplexity has been a popular metric to assess how well a statistical language model generalizes =-=[10]-=-. A model generalizes well if it achieves lower perplexity. We choose unigram as a baseline. Unigram is a special case of the joint topic and perspective model that assumes no lexical variations are d... |

237 | P.: The author-topic model for authors and documents
- Rosen-Zvi, Griffiths, et al.
- 2004
(Show Context)
Citation Context ...e are interested in modeling ideology texts that are mostly on the same topic but mainly differs in their ideological perspectives. There have been studies going beyond topics (e.g., modeling authors =-=[17]-=-). We are interested in modeling lexical variation collectively from multiple authors sharing similar beliefs, not lexical variations due to individual authors. 6 Conclusion We present a statistical m... |

190 | A variational Bayesian framework for graphical models, in
- Attias
- 2000
(Show Context)
Citation Context ...njugacy of the logistic-normal prior. We thus approximate the posterior probability distribution using a variational method [4], and estimate the parameters using variational expectation maximization =-=[5]-=-. By the Generalized Mean Field Theorem (GMF) [6], we can approximate the joint posterior probability distribution of τ and {φv} as the product of individual functions of τ and φv: P (τ,{φv}|{Pd}, {Wd... |

78 | Discriminative vs Informative Learning
- Rubinstein, Hastie
- 1997
(Show Context)
Citation Context ...odel for Ideological Discourse 31 the assumption of the underlying generative process on ideological text. In contrast, discriminative classifiers such as SVM do not model the data generation process =-=[13]-=-. However, our methods implicitly assume that documents are about the same news event or issue, which may not be true and could benefit from an extra story alignment step as [12] did. We borrow statis... |

55 | Which side are you on? identifying perspectives at the document and sentence levels
- Lin, Wilson, et al.
- 2006
(Show Context)
Citation Context ...ngagement: unilateral or coordinated?”). The website editors have labeled the ideological perspective of each published article. The bitterlemons corpus has been used to learn individual perspectives =-=[9]-=-, but the previous work was based on naive Bayes models and did not simultaneously model topics and perspectives. The 2000 and 2004 presidential debates corpus consists of the spoken transcripts of si... |

20 |
Ideology: A multidisciplinary approach
- Dijk
- 1998
(Show Context)
Citation Context ...ely our own ideas. In this paper we take a definition of ideology broader than the classic Marxists’ definition, but define ideology as “a set of general beliefs socially shared by a group of people” =-=[1]-=-. Groups whose members share similar goals or face similar problems usually share a set of beliefs that define membership, value judgment, and action. These collective beliefs form an ideology. For ex... |

17 |
Computer simulation of individual belief systems. A/nericun Eehoviorol Science
- Abelson, Carroll
- 1965
(Show Context)
Citation Context ... the last century, but the idea of learning ideology automatically from texts has been considered almost impossible. Abelson expressed a very pessimistic view on automatic learning approaches in 1965 =-=[2]-=-. We share Abelson’s vision but do not subscribe to his view. We believe that ideology can be statistically modeled and learned from a large number of ideological texts. – In this paper we develop a s... |

14 |
POLITICS: Automated ideological reasoning
- Carbonell
- 1978
(Show Context)
Citation Context ...beliefs of a right-wing politician as a set of English sentences (e.g., “Cuba subverts Latin America.”). Carbonell proposed a system, POLITICS, that can interpret text from two conflicting ideologies =-=[11]-=-. These early studies model ideology at a more sophisticated level (e.g., goals, actors, and action) than the proposed joint topic and perspective model, but require humans to manually construct a kno... |

6 | On tight approximate inference of logistic-normal admixture model
- Ahmed, Xing
- 2007
(Show Context)
Citation Context ...φv〉•(H(ˆτ ) •〈φv〉)(ˆτ •〈φv〉)) , v where ↓ is column-wise vector-matrix product, → is row-wise vector-matrix product. The Laplace approximation for the logistic-normal prior has been shown to be tight =-=[8]-=-. qφv in (3.2) can be approximated in a similar fashion as a multivariate normal distribution with a mean vector μ † and a variance matrix Σ † as follows, Σ † ( = Σ −1 φ + nTv 1〈τ〉 ↓H(〈τ〉•ˆ ) −1 φv) →... |

5 | On topic evolution
- Xing
- 2005
(Show Context)
Citation Context ... τ. Calculating the GMF message for τ from (4) is computationally intractable because of the non-conjugacy between multivariate normal and multinomial distributions. We follow the similar approach in =-=[7]-=-, and made a Laplace approximation of (4). We first represent the word likelihood {Wd,n} as the following exponential form: ( ∑ P ({Wd,n}|{Pd},τ,{〈φv〉}) =exp nv(〈φv〉•τ) − ∑ n T v 1C(〈φv〉•τ) ) (5) v wh... |

3 |
A generalized mean eld algorithm for variational inference in exponential families
- Xing, Jordan, et al.
- 2003
(Show Context)
Citation Context ...roximate the posterior probability distribution using a variational method [4], and estimate the parameters using variational expectation maximization [5]. By the Generalized Mean Field Theorem (GMF) =-=[6]-=-, we can approximate the joint posterior probability distribution of τ and {φv} as the product of individual functions of τ and φv: P (τ,{φv}|{Pd}, {Wd,n}; Θ) ≈ qτ (τ) ∏ v qφv (φv), (1) where qτ (τ) a... |

2 |
Detecting the bias in media with statistical learning methods
- Fortuna, Galleguillos, et al.
(Show Context)
Citation Context ...The knowledgeintensive approaches suffer from the “knowledge acquisition bottleneck.” We take a completely different approach and aim to automatically learn ideology from a large number of documents. =-=[12]-=- explored a similar problem of identifying media’s bias. They found that the sources of news articles can be successfully identified based on word choices using Support Vector Machines. They identifie... |

1 |
The Media At War: Communication and Conict
- Carruthers
- 2000
(Show Context)
Citation Context ...thor’s ideological perspective on an issue. “One man’s terrorist is another man’s freedom fighter.” Labeling a group as “terrorists” strongly reveal an author’s value judgement and ideological stance =-=[3]-=-. We illustrate lexical variations in an ideological text about the IsraeliPalestinian conflict (see Section 4.2). There were two groups of authors holding contrasting ideological perspectives (i.e., ... |

1 |
M.: Finding scientic topics
- Griffiths, Steyvers
- 2004
(Show Context)
Citation Context ...y not be true and could benefit from an extra story alignment step as [12] did. We borrow statistically modeling and inference techniques heavily from research on topic modeling (e.g., [14], [15] and =-=[16]-=-). They focus mostly on modeling text collections that containing many different (latent) topics (e.g., academic conference papers, news articles, etc). In contrast, we are interested in modeling ideo... |