Set up reliable automated benchmarks

[etiennedi]

Background

There is no "one" way for a contextionary to perform better or worse, but we can approximate all the features we expect through several benchmarks. It should be a "one-click" operation to benchmark one contextionary for all possible scenarios.

Goals

• Push runs a travis (or similiar CI) job
• The job runs a combination of settings, e.g.
  • different k values for kNN classification
  • different cut-off parameters for contextual classification
• The job can constantly be expanded with new benchmarks
• The benchmark results are stored in an easily accessible format (e.g. a HTML page is copied into a GCS bucket which acts as a webserver)

[etiennedi]

Dataset idea for German language classifications: https://tblock.github.io/10kGNAD/