engleski

Finding the most representative Latent Dirichlet Allocation run for topic modelling

The number of research publications is growing exponentially making the extraction of meaningful information increasingly challenging. Natural language processing may provide a solution. Latent Dirichlet Allocation (LDA) is frequently used to detect topics in a corpus of documents. It relies on Monte Carlo methods for estimation, which introduces a replicability risk. Recently, an approach to stabilization of topic- term allocation was proposed and implemented in R LDAPrototype package. Stabilization is achieved by analysing the topic-term frequency matrices from a set of LDA replications and choosing the LDA replication that is the most representative for the set. Another approach might be to base the choice of the most representative LDA replication on the document-topic frequency matrices. The objective of this research is to compare the two approaches to stabilization of LDA results on a corpus of papers on learning analytics and educational data mining.

learning analytics ; topic modelling ; Latent Dirichlet Allocation ; stability of allocation

nije evidentirano