Likewise, you should be able to set up beam_size, fasttrack_beam_size and word_beam_size through the command line. The surprisal is computed on base 2. Usually the value for beam_size >> word_beam_size. The default value is the following one used in all previous experiments: beam_size=100, fasttrack_beam_size=5, word_beam_size=10.
Available in the
rnng-incremental
repo. Peng's instructions: