Distributed Asymmetric Allocation: A Topic Model for Large Imbalanced Corpora in Social Sciences
By: Kohei Watanabe
Social scientists employ latent Dirichlet allocation (LDA) to find highly specific topics in large corpora, but they often struggle in this task because (1) LDA, in general, takes a significant amount of time to fit on large corpora; (2) unsupervised LDA fragments topics into sub-topics in short documents; (3) semi-supervised LDA fails to identify specific topics defined using seed words. To solve these problems, I have developed a new topic model called distributed asymmetric allocation (DAA) that integrates multiple algorithms for efficiently identifying sentences about important topics in large corpora. I evaluate the ability of DAA to identify politically important topics by fitting it to the transcripts of speeches at the United Nations General Assembly between 1991 and 2017. The results show that DAA can classify sentences significantly more accurately and quickly than LDA thanks to the new algorithms. More generally, the results demonstrate that it is important for social scientists to optimize Dirichlet priors of LDA to perform content analysis accurately.
Similar Papers
Quantifying consistency and accuracy of Latent Dirichlet Allocation
Computation and Language
Finds real topics in messy text data.
Analyzing Political Text at Scale with Online Tensor LDA
Machine Learning (CS)
Lets computers understand huge amounts of text fast.
Topic Analysis with Side Information: A Neural-Augmented LDA Approach
Machine Learning (CS)
Helps computers understand topics using extra clues.