Stylesignal uses social media to enhance accuracy in fashion trend forecasting.

Bit of both - the sources are "curated" and then we automatically discover other sources. A big part of what the software does is to establish source quality and authority.

The taxonomy automatically evolves - so for example during Halloween, words related to Halloween rise into the statistics even though we never manually seeded those words into our system.

The paper describing it is well over 7000 words(!), but basically through natural language processing and machine learning.