Advanced Quantitative Text Analysis

This course will revisit topics in Introduction to Quantitative Text Analysis but goes much deeper into theoretical and technological foundations of quantitative text analysis to be able to develop complex analytic pipelines in research projects.  

Day 1
The lecture will advance your knowledge of supervised and unsupervised methods, focusing on their strengths and weaknesses. We will cover Wordscores and naïve Bayes classifiers, Random Forest, latent Dirichlet allocation (LDA), and Structural Topic Model (STM). Wordscores and naïve Bayes classifiers are simple supervised algorithms for document scaling and for document classification, respectively. Random Forest can be used for both purposes, but it has a more sophisticated algorithm. LDA and STM are unsupervised algorithms for topic classification, but the latter can take into account document-level variables. We learn how to apply these models in the seminar.

Day 2
We will explain a seeded LDA model as well as LSS and Newsmap that offers compromise between strengths and weaknesses of supervised and unsupervised methods. These models rely on exemplary words (seed words) as supervision to perform document scaling or classification tasks. Semi-supervised models can be used for similar purposes as both supervised and unsupervised models, but training semi-supervised models demands special attention to the semantics of seed words. We will learn how to use these models in the seminar. 

Day 3
We discuss the word-embeddings technique that helps us accurately estimate semantic proximity of words in a large corpus. Although there are few applications of this technique in political science research, recently developed models Word2vec and GloVe attracted the attention of many quantitative text analysts to the technique. We explore its potential in the seminar.

Day 4
In the lecture, you will learn how to derive extra information from syntactic structures in texts and how to use that information to perform fine-tuned analysis. In the seminar, we apply syntactic parser, which recognises part-of-speech and dependencies of words, to improve text pre-processing, and geographical parsing (geoparser.io), which is a combination of a syntactic parser and a geographic database, to identify places mentioned in texts.

Day 5
You should come to the lecture with concrete research ideas involving quantitative text analysis. Some of you will be asked to present your ideas to initiate a class-wide discussion on how to choose analytic methods in actual research projects. In the seminar, you will learn how to develop your own text analysis tools by combining NLP functions in R.

You should have experience in quantitative text analysis in R – textual data management and preprocessing.

Basic knowledge of programming (object types, control flow, loop etc.) is desirable.