spaCy/spacy/tokens/doc.pyx

# coding: utf8
# cython: infer_types=True
# cython: bounds_check=False
from __future__ import unicode_literals

cimport cython
cimport numpy as np
import numpy
import numpy.linalg
import struct
import dill

from libc.string cimport memcpy, memset
from libc.math cimport sqrt

from .span cimport Span
from .token cimport Token
from .span cimport Span
from .token cimport Token
from .printers import parse_tree
from ..lexeme cimport Lexeme, EMPTY_LEXEME
from ..typedefs cimport attr_t, flags_t
from ..attrs import intify_attrs
from ..attrs cimport attr_id_t
from ..attrs cimport ID, ORTH, NORM, LOWER, SHAPE, PREFIX, SUFFIX, LENGTH, CLUSTER
from ..attrs cimport LENGTH, POS, LEMMA, TAG, DEP, HEAD, SPACY, ENT_IOB, ENT_TYPE
from ..attrs cimport SENT_START
from ..parts_of_speech cimport CCONJ, PUNCT, NOUN, univ_pos_t
from ..syntax.iterators import CHUNKERS
from ..util import normalize_slice
from ..compat import is_config
from .. import about


DEF PADDING = 5


cdef int bounds_check(int i, int length, int padding) except -1:
    if (i + padding) < 0:
        raise IndexError
    if (i - padding) >= length:
        raise IndexError


cdef attr_t get_token_attr(const TokenC* token, attr_id_t feat_name) nogil:
    if feat_name == LEMMA:
        return token.lemma
    elif feat_name == POS:
        return token.pos
    elif feat_name == TAG:
        return token.tag
    elif feat_name == DEP:
        return token.dep
    elif feat_name == HEAD:
        return token.head
    elif feat_name == SENT_START:
        return token.sent_start
    elif feat_name == SPACY:
        return token.spacy
    elif feat_name == ENT_IOB:
        return token.ent_iob
    elif feat_name == ENT_TYPE:
        return token.ent_type
    else:
        return Lexeme.get_struct_attr(token.lex, feat_name)


cdef class Doc:
    """A sequence of Token objects. Access sentences and named entities, export
    annotations to numpy arrays, losslessly serialize to compressed binary strings.
    The `Doc` object holds an array of `TokenC` structs. The Python-level
    `Token` and `Span` objects are views of this array, i.e. they don't own
    the data themselves.

    EXAMPLE: Construction 1
        >>> doc = nlp(u'Some text')

        Construction 2
        >>> from spacy.tokens import Doc
        >>> doc = Doc(nlp.vocab, words=[u'hello', u'world', u'!'], spaces=[True, False, False])
    """
    def __init__(self, Vocab vocab, words=None, spaces=None, orths_and_spaces=None):
        """Create a Doc object.

        vocab (Vocab): A vocabulary object, which must match any models you want
            to use (e.g. tokenizer, parser, entity recognizer).
        words (list or None): A list of unicode strings to add to the document
            as words. If `None`, defaults to empty list.
        spaces (list or None): A list of boolean values, of the same length as
            words. True means that the word is followed by a space, False means
            it is not. If `None`, defaults to `[True]*len(words)`
        RETURNS (Doc): The newly constructed object.
        """
        self.vocab = vocab
        size = 20
        self.mem = Pool()
        # Guarantee self.lex[i-x], for any i >= 0 and x < padding is in bounds
        # However, we need to remember the true starting places, so that we can
        # realloc.
        data_start = <TokenC*>self.mem.alloc(size + (PADDING*2), sizeof(TokenC))
        cdef int i
        for i in range(size + (PADDING*2)):
            data_start[i].lex = &EMPTY_LEXEME
            data_start[i].l_edge = i
            data_start[i].r_edge = i
        self.c = data_start + PADDING
        self.max_length = size
        self.length = 0
        self.is_tagged = False
        self.is_parsed = False
        self.sentiment = 0.0
        self.user_hooks = {}
        self.user_token_hooks = {}
        self.user_span_hooks = {}
        self.tensor = numpy.zeros((0,), dtype='float32')
        self.user_data = {}
        self._py_tokens = []
        self._vector = None
        self.noun_chunks_iterator = CHUNKERS.get(self.vocab.lang)
        cdef unicode orth
        cdef bint has_space
        if orths_and_spaces is None and words is not None:
            if spaces is None:
                spaces = [True] * len(words)
            elif len(spaces) != len(words):
                raise ValueError(
                    "Arguments 'words' and 'spaces' should be sequences of the "
                    "same length, or 'spaces' should be left default at None. "
                    "spaces should be a sequence of booleans, with True meaning "
                    "that the word owns a ' ' character following it.")
            orths_and_spaces = zip(words, spaces)
        if orths_and_spaces is not None:
            for orth_space in orths_and_spaces:
                if isinstance(orth_space, unicode):
                    orth = orth_space
                    has_space = True
                elif isinstance(orth_space, bytes):
                    raise ValueError(
                        "orths_and_spaces expects either List(unicode) or "
                        "List((unicode, bool)). Got bytes instance: %s" % (str(orth_space)))
                else:
                    orth, has_space = orth_space
                # Note that we pass self.mem here --- we have ownership, if LexemeC
                # must be created.
                self.push_back(
                    <const LexemeC*>self.vocab.get(self.mem, orth), has_space)
        # Tough to decide on policy for this. Is an empty doc tagged and parsed?
        # There's no information we'd like to add to it, so I guess so?
        if self.length == 0:
            self.is_tagged = True
            self.is_parsed = True

    def __getitem__(self, object i):
        """Get a `Token` or `Span` object.

        i (int or tuple) The index of the token, or the slice of the document to get.
        RETURNS (Token or Span): The token at `doc[i]]`, or the span at
            `doc[start : end]`.

        EXAMPLE:
            >>> doc[i]
            Get the `Token` object at position `i`, where `i` is an integer.
            Negative indexing is supported, and follows the usual Python
            semantics, i.e. `doc[-2]` is `doc[len(doc) - 2]`.

            >>> doc[start : end]]
            Get a `Span` object, starting at position `start` and ending at
            position `end`, where `start` and `end` are token indices. For
            instance, `doc[2:5]` produces a span consisting of tokens 2, 3 and 4.
            Stepped slices (e.g. `doc[start : end : step]`) are not supported,
            as `Span` objects must be contiguous (cannot have gaps). You can use
            negative indices and open-ended ranges, which have their normal
            Python semantics.
        """
        if isinstance(i, slice):
            start, stop = normalize_slice(len(self), i.start, i.stop, i.step)
            return Span(self, start, stop, label=0)

        if i < 0:
            i = self.length + i
        bounds_check(i, self.length, PADDING)
        if self._py_tokens[i] is not None:
            return self._py_tokens[i]
        else:
            return Token.cinit(self.vocab, &self.c[i], i, self)

    def __iter__(self):
        """Iterate over `Token`  objects, from which the annotations can be
        easily accessed. This is the main way of accessing `Token` objects,
        which are the main way annotations are accessed from Python. If faster-
        than-Python speeds are required, you can instead access the annotations
        as a numpy array, or access the underlying C data directly from Cython.

        EXAMPLE:
            >>> for token in doc
        """
        cdef int i
        for i in range(self.length):
            if self._py_tokens[i] is not None:
                yield self._py_tokens[i]
            else:
                yield Token.cinit(self.vocab, &self.c[i], i, self)

    def __len__(self):
        """The number of tokens in the document.

        RETURNS (int): The number of tokens in the document.

        EXAMPLE:
            >>> len(doc)
        """
        return self.length

    def __unicode__(self):
        return u''.join([t.text_with_ws for t in self])

    def __bytes__(self):
        return u''.join([t.text_with_ws for t in self]).encode('utf-8')

    def __str__(self):
        if is_config(python3=True):
            return self.__unicode__()
        return self.__bytes__()

    def __repr__(self):
        return self.__str__()

    @property
    def doc(self):
        return self

    def similarity(self, other):
        """Make a semantic similarity estimate. The default estimate is cosine
        similarity using an average of word vectors.

        other (object): The object to compare with. By default, accepts `Doc`,
            `Span`, `Token` and `Lexeme` objects.
        RETURNS (float): A scalar similarity score. Higher is more similar.
        """
        if 'similarity' in self.user_hooks:
            return self.user_hooks['similarity'](self, other)
        if self.vector_norm == 0 or other.vector_norm == 0:
            return 0.0
        return numpy.dot(self.vector, other.vector) / (self.vector_norm * other.vector_norm)

    property has_vector:
        """A boolean value indicating whether a word vector is associated with
        the object.

        RETURNS (bool): Whether a word vector is associated with the object.
        """
        def __get__(self):
            if 'has_vector' in self.user_hooks:
                return self.user_hooks['has_vector'](self)

            return any(token.has_vector for token in self)

    property vector:
        """A real-valued meaning representation. Defaults to an average of the
        token vectors.

        RETURNS (numpy.ndarray[ndim=1, dtype='float32']): A 1D numpy array
            representing the document's semantics.
        """
        def __get__(self):
            if 'vector' in self.user_hooks:
                return self.user_hooks['vector'](self)
            if self._vector is None:
                if len(self):
                    self._vector = sum(t.vector for t in self) / len(self)
                else:
                    return numpy.zeros((self.vocab.vectors_length,), dtype='float32')
            return self._vector

        def __set__(self, value):
            self._vector = value

    property vector_norm:
        """The L2 norm of the document's vector representation.

        RETURNS (float): The L2 norm of the vector representation.
        """
        def __get__(self):
            if 'vector_norm' in self.user_hooks:
                return self.user_hooks['vector_norm'](self)
            cdef float value
            cdef double norm = 0
            if self._vector_norm is None:
                norm = 0.0
                for value in self.vector:
                    norm += value * value
                self._vector_norm = sqrt(norm) if norm != 0 else 0
            return self._vector_norm

        def __set__(self, value):
            self._vector_norm = value

    @property
    def string(self):
        return self.text

    property text:
        """A unicode representation of the document text.

        RETURNS (unicode): The original verbatim text of the document.
        """
        def __get__(self):
            return u''.join(t.text_with_ws for t in self)

    property text_with_ws:
        """An alias of `Doc.text`, provided for duck-type compatibility with
        `Span` and `Token`.

        RETURNS (unicode): The original verbatim text of the document.
        """
        def __get__(self):
            return self.text

    property ents:
        """Iterate over the entities in the document. Yields named-entity `Span`
        objects, if the entity recognizer has been applied to the document.

        YIELDS (Span): Entities in the document.

        EXAMPLE: Iterate over the span to get individual Token objects, or access
            the label:

            >>> tokens = nlp(u'Mr. Best flew to New York on Saturday morning.')
            >>> ents = list(tokens.ents)
            >>> assert ents[0].label == 346
            >>> assert ents[0].label_ == 'PERSON'
            >>> assert ents[0].orth_ == 'Best'
            >>> assert ents[0].text == 'Mr. Best'
        """
        def __get__(self):
            cdef int i
            cdef const TokenC* token
            cdef int start = -1
            cdef attr_t label = 0
            output = []
            for i in range(self.length):
                token = &self.c[i]
                if token.ent_iob == 1:
                    assert start != -1
                elif token.ent_iob == 2 or token.ent_iob == 0:
                    if start != -1:
                        output.append(Span(self, start, i, label=label))
                    start = -1
                    label = 0
                elif token.ent_iob == 3:
                    if start != -1:
                        output.append(Span(self, start, i, label=label))
                    start = i
                    label = token.ent_type
            if start != -1:
                output.append(Span(self, start, self.length, label=label))
            return tuple(output)

        def __set__(self, ents):
            # TODO:
            # 1. Allow negative matches
            # 2. Ensure pre-set NERs are not over-written during statistical prediction
            # 3. Test basic data-driven ORTH gazetteer
            # 4. Test more nuanced date and currency regex
            cdef int i
            for i in range(self.length):
                self.c[i].ent_type = 0
                # At this point we don't know whether the NER has run over the
                # Doc. If the ent_iob is missing, leave it missing.
                if self.c[i].ent_iob != 0:
                    self.c[i].ent_iob = 2 # Means O. Non-O are set from ents.
            cdef attr_t ent_type
            cdef int start, end
            for ent_info in ents:
                if isinstance(ent_info, Span):
                    ent_id = ent_info.ent_id
                    ent_type = ent_info.label
                    start = ent_info.start
                    end = ent_info.end
                elif len(ent_info) == 3:
                    ent_type, start, end = ent_info
                else:
                    ent_id, ent_type, start, end = ent_info
                if ent_type is None or ent_type < 0:
                    # Mark as O
                    for i in range(start, end):
                        self.c[i].ent_type = 0
                        self.c[i].ent_iob = 2
                else:
                    # Mark (inside) as I
                    for i in range(start, end):
                        self.c[i].ent_type = ent_type
                        self.c[i].ent_iob = 1
                    # Set start as B
                    self.c[start].ent_iob = 3

    property noun_chunks:
        """Iterate over the base noun phrases in the document. Yields base
        noun-phrase #[code Span] objects, if the document has been syntactically
        parsed. A base noun phrase, or "NP chunk", is a noun phrase that does
        not permit other NPs to be nested within it – so no NP-level
        coordination, no prepositional phrases, and no relative clauses.

        YIELDS (Span): Noun chunks in the document.
        """
        def __get__(self):
            if not self.is_parsed:
                raise ValueError(
                    "noun_chunks requires the dependency parse, which "
                    "requires data to be installed. For more info, see the "
                    "documentation: \n%s\n" % about.__docs_models__)
            # Accumulate the result before beginning to iterate over it. This prevents
            # the tokenisation from being changed out from under us during the iteration.
            # The tricky thing here is that Span accepts its tokenisation changing,
            # so it's okay once we have the Span objects. See Issue #375
            spans = []
            for start, end, label in self.noun_chunks_iterator(self):
                spans.append(Span(self, start, end, label=label))
            for span in spans:
                yield span

    property sents:
        """Iterate over the sentences in the document. Yields sentence `Span`
        objects. Sentence spans have no label. To improve accuracy on informal
        texts, spaCy calculates sentence boundaries from the syntactic
        dependency parse. If the parser is disabled, the `sents` iterator will
        be unavailable.

        EXAMPLE:
            >>> doc = nlp("This is a sentence. Here's another...")
            >>> assert [s.root.text for s in doc.sents] == ["is", "'s"]
        """
        def __get__(self):
            if 'sents' in self.user_hooks:
                return self.user_hooks['sents'](self)

            if not self.is_parsed:
                raise ValueError(
                    "sentence boundary detection requires the dependency parse, which "
                    "requires data to be installed. For more info, see the "
                    "documentation: \n%s\n" % about.__docs_models__)
            cdef int i
            start = 0
            for i in range(1, self.length):
                if self.c[i].sent_start:
                    yield Span(self, start, i)
                    start = i
            if start != self.length:
                yield Span(self, start, self.length)

    cdef int push_back(self, LexemeOrToken lex_or_tok, bint has_space) except -1:
        if self.length == 0:
            # Flip these to false when we see the first token.
            self.is_tagged = False
            self.is_parsed = False
        if self.length == self.max_length:
            self._realloc(self.length * 2)
        cdef TokenC* t = &self.c[self.length]
        if LexemeOrToken is const_TokenC_ptr:
            t[0] = lex_or_tok[0]
        else:
            t.lex = lex_or_tok
        if self.length == 0:
            t.idx = 0
        else:
            t.idx = (t-1).idx + (t-1).lex.length + (t-1).spacy
        t.l_edge = self.length
        t.r_edge = self.length
        assert t.lex.orth != 0
        t.spacy = has_space
        self.length += 1
        self._py_tokens.append(None)
        return t.idx + t.lex.length + t.spacy

    @cython.boundscheck(False)
    cpdef np.ndarray to_array(self, object py_attr_ids):
        """Given a list of M attribute IDs, export the tokens to a numpy
        `ndarray` of shape `(N, M)`, where `N` is the length of the document.
        The values will be 32-bit integers.

        attr_ids (list[int]): A list of attribute ID ints.
        RETURNS (numpy.ndarray[long, ndim=2]): A feature matrix, with one row
            per word, and one column per attribute indicated in the input
            `attr_ids`.

        EXAMPLE:
            >>> from spacy.attrs import LOWER, POS, ENT_TYPE, IS_ALPHA
            >>> doc = nlp(text)
            >>> # All strings mapped to integers, for easy export to numpy
            >>> np_array = doc.to_array([LOWER, POS, ENT_TYPE, IS_ALPHA])
        """
        cdef int i, j
        cdef attr_id_t feature
        cdef np.ndarray[attr_t, ndim=2] output
        # Make an array from the attributes --- otherwise our inner loop is Python
        # dict iteration.
        cdef np.ndarray[attr_t, ndim=1] attr_ids = numpy.asarray(py_attr_ids, dtype=numpy.uint64)
        output = numpy.ndarray(shape=(self.length, len(attr_ids)), dtype=numpy.uint64)
        for i in range(self.length):
            for j, feature in enumerate(attr_ids):
                output[i, j] = get_token_attr(&self.c[i], feature)
        return output

    def count_by(self, attr_id_t attr_id, exclude=None, PreshCounter counts=None):
        """Count the frequencies of a given attribute. Produces a dict of
        `{attribute (int): count (ints)}` frequencies, keyed by the values of
        the given attribute ID.

        attr_id (int): The attribute ID to key the counts.
        RETURNS (dict): A dictionary mapping attributes to integer counts.

        EXAMPLE:
            >>> from spacy import attrs
            >>> doc = nlp(u'apple apple orange banana')
            >>> tokens.count_by(attrs.ORTH)
            {12800L: 1, 11880L: 2, 7561L: 1}
            >>> tokens.to_array([attrs.ORTH])
            array([[11880], [11880], [7561], [12800]])
        """
        cdef int i
        cdef attr_t attr
        cdef size_t count

        if counts is None:
            counts = PreshCounter()
            output_dict = True
        else:
            output_dict = False
        # Take this check out of the loop, for a bit of extra speed
        if exclude is None:
            for i in range(self.length):
                counts.inc(get_token_attr(&self.c[i], attr_id), 1)
        else:
            for i in range(self.length):
                if not exclude(self[i]):
                    attr = get_token_attr(&self.c[i], attr_id)
                    counts.inc(attr, 1)
        if output_dict:
            return dict(counts)

    def _realloc(self, new_size):
        self.max_length = new_size
        n = new_size + (PADDING * 2)
        # What we're storing is a "padded" array. We've jumped forward PADDING
        # places, and are storing the pointer to that. This way, we can access
        # words out-of-bounds, and get out-of-bounds markers.
        # Now that we want to realloc, we need the address of the true start,
        # so we jump the pointer back PADDING places.
        cdef TokenC* data_start = self.c - PADDING
        data_start = <TokenC*>self.mem.realloc(data_start, n * sizeof(TokenC))
        self.c = data_start + PADDING
        cdef int i
        for i in range(self.length, self.max_length + PADDING):
            self.c[i].lex = &EMPTY_LEXEME

    cdef void set_parse(self, const TokenC* parsed) nogil:
        # TODO: This method is fairly misleading atm. It's used by Parser
        # to actually apply the parse calculated. Need to rethink this.

        # Probably we should use from_array?
        self.is_parsed = True
        for i in range(self.length):
            self.c[i] = parsed[i]

    def from_array(self, attrs, array):
        if SENT_START in attrs and HEAD in attrs:
            raise ValueError(
                "Conflicting attributes specified in doc.from_array():\n"
                "(HEAD, SENT_START)\n"
                "The HEAD attribute currently sets sentence boundaries implicitly,\n"
                "based on the tree structure. This means the HEAD attribute would "
                "potentially override the sentence boundaries set by SENT_START.\n"
                "See https://github.com/spacy-io/spaCy/issues/235 for details and "
                "workarounds, and to propose solutions.")
        cdef int i, col
        cdef attr_id_t attr_id
        cdef TokenC* tokens = self.c
        cdef int length = len(array)
        # Get set up for fast loading
        cdef Pool mem = Pool()
        cdef int n_attrs = len(attrs)
        attr_ids = <attr_id_t*>mem.alloc(n_attrs, sizeof(attr_id_t))
        for i, attr_id in enumerate(attrs):
            attr_ids[i] = attr_id
        # Now load the data
        for i in range(self.length):
            token = &self.c[i]
            for j in range(n_attrs):
                Token.set_struct_attr(token, attr_ids[j], array[i, j])
        # Auxiliary loading logic
        for col, attr_id in enumerate(attrs):
            if attr_id == TAG:
                for i in range(length):
                    if array[i, col] != 0:
                        self.vocab.morphology.assign_tag(&tokens[i], array[i, col])
        set_children_from_heads(self.c, self.length)
        self.is_parsed = bool(HEAD in attrs or DEP in attrs)
        self.is_tagged = bool(TAG in attrs or POS in attrs)
        return self

    def to_disk(self, path):
        """Save the current state to a directory.

        path (unicode or Path): A path to a directory, which will be created if
            it doesn't exist. Paths may be either strings or `Path`-like objects.
        """
        raise NotImplementedError()

    def from_disk(self, path):
        """Loads state from a directory. Modifies the object in place and
        returns it.

        path (unicode or Path): A path to a directory. Paths may be either
            strings or `Path`-like objects.
        RETURNS (Doc): The modified `Doc` object.
        """
        raise NotImplementedError()

    def to_bytes(self):
        """Serialize, i.e. export the document contents to a binary string.

        RETURNS (bytes): A losslessly serialized copy of the `Doc`, including
            all annotations.
        """
        return dill.dumps(
            (self.text,
            self.to_array([LENGTH,SPACY,TAG,LEMMA,HEAD,DEP,ENT_IOB,ENT_TYPE]),
            self.sentiment,
            self.tensor,
            self.noun_chunks_iterator,
            self.user_data,
            (self.user_hooks, self.user_token_hooks, self.user_span_hooks)),
            protocol=-1)

    def from_bytes(self, data):
        """Deserialize, i.e. import the document contents from a binary string.

        data (bytes): The string to load from.
        RETURNS (Doc): Itself.
        """
        if self.length != 0:
            raise ValueError("Cannot load into non-empty Doc")
        cdef attr_t[:, :] attrs
        cdef int i, start, end, has_space
        fields = dill.loads(data)
        text, attrs = fields[:2]
        self.sentiment, self.tensor = fields[2:4]
        self.noun_chunks_iterator, self.user_data = fields[4:6]
        self.user_hooks, self.user_token_hooks, self.user_span_hooks = fields[6]

        start = 0
        cdef const LexemeC* lex
        cdef unicode orth_
        for i in range(attrs.shape[0]):
            end = start + attrs[i, 0]
            has_space = attrs[i, 1]
            orth_ = text[start:end]
            lex = self.vocab.get(self.mem, orth_)
            self.push_back(lex, has_space)

            start = end + has_space
        self.from_array([TAG,LEMMA,HEAD,DEP,ENT_IOB,ENT_TYPE],
            attrs[:, 2:])
        return self

    def merge(self, int start_idx, int end_idx, *args, **attributes):
        """Retokenize the document, such that the span at `doc.text[start_idx : end_idx]`
        is merged into a single token. If `start_idx` and `end_idx `do not mark
        start and end token boundaries, the document remains unchanged.

        start_idx (int): The character index of the start of the slice to merge.
        end_idx (int): The character index after the end of the slice to merge.
        **attributes: Attributes to assign to the merged token. By default,
            attributes are inherited from the syntactic root token of the span.
        RETURNS (Token): The newly merged token, or `None` if the start and end
            indices did not fall at token boundaries.
        """
        cdef unicode tag, lemma, ent_type
        if len(args) == 3:
            # TODO: Warn deprecation
            tag, lemma, ent_type = args
            attributes[TAG] = tag
            attributes[LEMMA] = lemma
            attributes[ENT_TYPE] = ent_type
        elif not args:
            if "label" in attributes and 'ent_type' not in attributes:
                if isinstance(attributes["label"], int):
                    attributes[ENT_TYPE] = attributes["label"]
                else:
                    attributes[ENT_TYPE] = self.vocab.strings[attributes["label"]]
            if 'ent_type' in attributes:
                attributes[ENT_TYPE] = attributes['ent_type']
        elif args:
            raise ValueError(
                "Doc.merge received %d non-keyword arguments. "
                "Expected either 3 arguments (deprecated), or 0 (use keyword arguments). "
                "Arguments supplied:\n%s\n"
                "Keyword arguments:%s\n" % (len(args), repr(args), repr(attributes)))

        # More deprecated attribute handling =/
        if 'label' in attributes:
            attributes['ent_type'] = attributes.pop('label')

        attributes = intify_attrs(attributes, strings_map=self.vocab.strings)

        cdef int start = token_by_start(self.c, self.length, start_idx)
        if start == -1:
            return None
        cdef int end = token_by_end(self.c, self.length, end_idx)
        if end == -1:
            return None
        # Currently we have the token index, we want the range-end index
        end += 1
        cdef Span span = self[start:end]
        # Get LexemeC for newly merged token
        new_orth = ''.join([t.text_with_ws for t in span])
        if span[-1].whitespace_:
            new_orth = new_orth[:-len(span[-1].whitespace_)]
        cdef const LexemeC* lex = self.vocab.get(self.mem, new_orth)
        # House the new merged token where it starts
        cdef TokenC* token = &self.c[start]
        token.spacy = self.c[end-1].spacy
        for attr_name, attr_value in attributes.items():
            if attr_name == TAG:
                self.vocab.morphology.assign_tag(token, attr_value) 
            else:
                Token.set_struct_attr(token, attr_name, attr_value)
        # Begin by setting all the head indices to absolute token positions
        # This is easier to work with for now than the offsets
        # Before thinking of something simpler, beware the case where a dependency
        # bridges over the entity. Here the alignment of the tokens changes.
        span_root = span.root.i
        token.dep = span.root.dep
        # We update token.lex after keeping span root and dep, since
        # setting token.lex will change span.start and span.end properties
        # as it modifies the character offsets in the doc
        token.lex = lex
        for i in range(self.length):
            self.c[i].head += i
        # Set the head of the merged token, and its dep relation, from the Span
        token.head = self.c[span_root].head
        # Adjust deps before shrinking tokens
        # Tokens which point into the merged token should now point to it
        # Subtract the offset from all tokens which point to >= end
        offset = (end - start) - 1
        for i in range(self.length):
            head_idx = self.c[i].head
            if start <= head_idx < end:
                self.c[i].head = start
            elif head_idx >= end:
                self.c[i].head -= offset
        # Now compress the token array
        for i in range(end, self.length):
            self.c[i - offset] = self.c[i]
        for i in range(self.length - offset, self.length):
            memset(&self.c[i], 0, sizeof(TokenC))
            self.c[i].lex = &EMPTY_LEXEME
        self.length -= offset
        for i in range(self.length):
            # ...And, set heads back to a relative position
            self.c[i].head -= i
        # Set the left/right children, left/right edges
        set_children_from_heads(self.c, self.length)
        # Clear the cached Python objects
        self._py_tokens = [None] * self.length
        # Return the merged Python object
        return self[start]

    def print_tree(self, light=False, flat=False):
        """Returns the parse trees in JSON (dict) format.

        light (bool): Don't include lemmas or entities.
        flat (bool): Don't include arcs or modifiers.
        RETURNS (dict): Parse tree as dict.

        EXAMPLE:
            >>> doc = nlp('Bob brought Alice the pizza. Alice ate the pizza.')
            >>> trees = doc.print_tree()
            >>> trees[1]
            {'modifiers': [
                {'modifiers': [], 'NE': 'PERSON', 'word': 'Alice', 'arc': 'nsubj',
                'POS_coarse': 'PROPN', 'POS_fine': 'NNP', 'lemma': 'Alice'},
                {'modifiers': [
                    {'modifiers': [], 'NE': '', 'word': 'the', 'arc': 'det',
                    'POS_coarse': 'DET', 'POS_fine': 'DT', 'lemma': 'the'}],
                'NE': '', 'word': 'pizza', 'arc': 'dobj', 'POS_coarse': 'NOUN',
                'POS_fine': 'NN', 'lemma': 'pizza'},
                {'modifiers': [], 'NE': '', 'word': '.', 'arc': 'punct',
                'POS_coarse': 'PUNCT', 'POS_fine': '.', 'lemma': '.'}],
                'NE': '', 'word': 'ate', 'arc': 'ROOT', 'POS_coarse': 'VERB',
                'POS_fine': 'VBD', 'lemma': 'eat'}
        """
        return parse_tree(self, light=light, flat=flat)


cdef int token_by_start(const TokenC* tokens, int length, int start_char) except -2:
    cdef int i
    for i in range(length):
        if tokens[i].idx == start_char:
            return i
    else:
        return -1


cdef int token_by_end(const TokenC* tokens, int length, int end_char) except -2:
    cdef int i
    for i in range(length):
        if tokens[i].idx + tokens[i].lex.length == end_char:
            return i
    else:
        return -1


cdef int set_children_from_heads(TokenC* tokens, int length) except -1:
    cdef TokenC* head
    cdef TokenC* child
    cdef int i
    # Set number of left/right children to 0. We'll increment it in the loops.
    for i in range(length):
        tokens[i].l_kids = 0
        tokens[i].r_kids = 0
        tokens[i].l_edge = i
        tokens[i].r_edge = i
    # Set left edges
    for i in range(length):
        child = &tokens[i]
        head = &tokens[i + child.head]
        if child < head:
            if child.l_edge < head.l_edge:
                head.l_edge = child.l_edge
            head.l_kids += 1

    # Set right edges --- same as above, but iterate in reverse
    for i in range(length-1, -1, -1):
        child = &tokens[i]
        head = &tokens[i + child.head]
        if child > head:
            if child.r_edge > head.r_edge:
                head.r_edge = child.r_edge
            head.r_kids += 1

    # Set sentence starts
    for i in range(length):
        if tokens[i].head == 0 and tokens[i].dep != 0:
            tokens[tokens[i].l_edge].sent_start = True