These are a few scripts for the generation, examination and comparison of LSA and LDA models. The current iteration works entirely in system memory. This is something that we will be looking to rectify in the near future so that these tests can be applied to large corpora such as Wikipedia. A brief summary of the scripts provided and their functionality are described below, although their use should be fairly obvious from their names.

Takes as input one or more text files and generates either an LSA or an LDA model from them. Command line arguments allow you to select the type of model (LDA by default), number of topics (2 by default) and name the output dictionary and model files (out.dict and out.model respectively by default).

A more comprehensive range of settings are available in the docs/config.py file. This gives you full access to the range of options provided by the gensim API. It also has a setting for the comparison script (described later).

Takes as input a dictionary and a model so that the model may be examined. At present, just prints the top terms for each topic in the model. Will be made more comprehensive later.

Takes as input two LSA or LDA models. At present, it identifies pairings for each of the most similar topics between the two models. However, in future this information will be expanded to compute the similarity of the two models and hence give a measure for their stability.

Similarity is measured using the average Jaccard distance, a metric proposed by Greene, O'Callaghan and Cunningham in their paper "How Many Topics? Stability Analysis for Topic Models"