projection_count_vectorizer.py

from sklearn.feature_extraction.text import CountVectorizer

class ProjectionCountVectorizer(CountVectorizer):

    def __init__(self, projection_path, *args, **kwargs):
        self.projection_path = projection_path.split('/')
        super().__init__(*args, **kwargs)

    def build_preprocessor(self):
        built = super().build_preprocessor()

        def projection_and_preprocess(doc):
            return built(self.do_projection(doc))
        return projection_and_preprocess
        
    def do_projection(self, doc):
        for step in self.projection_path:
            if isinstance(doc, dict):
                doc = doc[step]
            elif isinstance(doc, (tuple, list)):
                if step.isdigit():
                    doc = doc[int(step)]
                else:   # only valid for namedtuples
                    doc = getattr(doc, step)
            else:
                raise ValueError('cant apply step %s' % step)
        return doc