Political press releases: More than just words

The word clouds were generated by machine learning analysis of the six political parties which were most likely to be represented in Parliament. They are not simple word counts but rather reflect the ranking of the words that consistently distinguished the press releases of the given party from its peers.

" (These wordclouds) reflect the ranking of the words that consistently distinguished the press releases of the given party from its peers."

The parties are in alphabetical order.

ACT

{CF_IMAGE}

GREEN

{CF_IMAGE}

LABOUR

{CF_IMAGE}

MAORI

{CF_IMAGE}

NATIONAL

{CF_IMAGE}

NZ FIRST

{CF_IMAGE}

Source: Political websites, ANZ Research

Note: The two models were a Naïve Bayes Multinomial and a Support Vector Machine classifier using TD-IDF vectorisation and word stemming (a representative word is used in the clouds). The words and phrases shown are those that were in the top 200 distinguishing features for each party across both models, and are sized according to the coefficients from the Naïve Bayes model. The two models achieved accuracy of over 85% on an unseen test set of press releases. The word clouds were generated using wordle.net.