Awesome, thanks for following up! We also use coreference resolution - we specifically found Allen NLP's coreference to be the most accurate in an analysis of a few different offerings: http://docs.allennlp.org/v0.9.0/api/allennlp.models.corefere...