spaCy/spacy/matcher/matcher.pyx

# cython: infer_types=True, cython: profile=True
from libcpp.vector cimport vector
from libc.stdint cimport int32_t
from cymem.cymem cimport Pool
from murmurhash.mrmr cimport hash64

import re
import srsly
import warnings

from ..typedefs cimport attr_t
from ..structs cimport TokenC
from ..vocab cimport Vocab
from ..tokens.doc cimport Doc, get_token_attr_for_matcher
from ..tokens.span cimport Span
from ..tokens.token cimport Token
from ..attrs cimport ID, attr_id_t, NULL_ATTR, ORTH, POS, TAG, DEP, LEMMA

from ..schemas import validate_token_pattern
from ..errors import Errors, MatchPatternError, Warnings
from ..strings import get_string_id
from ..attrs import IDS


DEF PADDING = 5


cdef class Matcher:
    """Match sequences of tokens, based on pattern rules.

    DOCS: https://spacy.io/api/matcher
    USAGE: https://spacy.io/usage/rule-based-matching
    """

    def __init__(self, vocab, validate=True):
        """Create the Matcher.

        vocab (Vocab): The vocabulary object, which must be shared with the
            documents the matcher will operate on.
        RETURNS (Matcher): The newly constructed object.
        """
        self._extra_predicates = []
        self._patterns = {}
        self._callbacks = {}
        self._extensions = {}
        self._seen_attrs = set()
        self.vocab = vocab
        self.mem = Pool()
        self.validate = validate

    def __reduce__(self):
        data = (self.vocab, self._patterns, self._callbacks)
        return (unpickle_matcher, data, None, None)

    def __len__(self):
        """Get the number of rules added to the matcher. Note that this only
        returns the number of rules (identical with the number of IDs), not the
        number of individual patterns.

        RETURNS (int): The number of rules.
        """
        return len(self._patterns)

    def __contains__(self, key):
        """Check whether the matcher contains rules for a match ID.

        key (str): The match ID.
        RETURNS (bool): Whether the matcher contains rules for this match ID.
        """
        return self._normalize_key(key) in self._patterns

    def add(self, key, patterns, *_patterns, on_match=None):
        """Add a match-rule to the matcher. A match-rule consists of: an ID
        key, an on_match callback, and one or more patterns.

        If the key exists, the patterns are appended to the previous ones, and
        the previous on_match callback is replaced. The `on_match` callback
        will receive the arguments `(matcher, doc, i, matches)`. You can also
        set `on_match` to `None` to not perform any actions.

        A pattern consists of one or more `token_specs`, where a `token_spec`
        is a dictionary mapping attribute IDs to values, and optionally a
        quantifier operator under the key "op". The available quantifiers are:

        '!': Negate the pattern, by requiring it to match exactly 0 times.
        '?': Make the pattern optional, by allowing it to match 0 or 1 times.
        '+': Require the pattern to match 1 or more times.
        '*': Allow the pattern to zero or more times.

        The + and * operators are usually interpretted "greedily", i.e. longer
        matches are returned where possible. However, if you specify two '+'
        and '*' patterns in a row and their matches overlap, the first
        operator will behave non-greedily. This quirk in the semantics makes
        the matcher more efficient, by avoiding the need for back-tracking.

        As of spaCy v2.2.2, Matcher.add supports the future API, which makes
        the patterns the second argument and a list (instead of a variable
        number of arguments). The on_match callback becomes an optional keyword
        argument.

        key (str): The match ID.
        patterns (list): The patterns to add for the given key.
        on_match (callable): Optional callback executed on match.
        *_patterns (list): For backwards compatibility: list of patterns to add
            as variable arguments. Will be ignored if a list of patterns is
            provided as the second argument.
        """
        errors = {}
        if on_match is not None and not hasattr(on_match, "__call__"):
            raise ValueError(Errors.E171.format(arg_type=type(on_match)))
        if patterns is None or hasattr(patterns, "__call__"):  # old API
            on_match = patterns
            patterns = _patterns
        for i, pattern in enumerate(patterns):
            if len(pattern) == 0:
                raise ValueError(Errors.E012.format(key=key))
            if not isinstance(pattern, list):
                raise ValueError(Errors.E178.format(pat=pattern, key=key))
            if self.validate:
                errors[i] = validate_token_pattern(pattern)
        if any(err for err in errors.values()):
            raise MatchPatternError(key, errors)
        key = self._normalize_key(key)
        for pattern in patterns:
            try:
                specs = _preprocess_pattern(pattern, self.vocab.strings,
                    self._extensions, self._extra_predicates)
                self.patterns.push_back(init_pattern(self.mem, key, specs))
                for spec in specs:
                    for attr, _ in spec[1]:
                        self._seen_attrs.add(attr)
            except OverflowError, AttributeError:
                raise ValueError(Errors.E154.format())
        self._patterns.setdefault(key, [])
        self._callbacks[key] = on_match
        self._patterns[key].extend(patterns)

    def remove(self, key):
        """Remove a rule from the matcher. A KeyError is raised if the key does
        not exist.

        key (str): The ID of the match rule.
        """
        norm_key = self._normalize_key(key)
        if not norm_key in self._patterns:
            raise ValueError(Errors.E175.format(key=key))
        self._patterns.pop(norm_key)
        self._callbacks.pop(norm_key)
        cdef int i = 0
        while i < self.patterns.size():
            pattern_key = get_ent_id(self.patterns.at(i))
            if pattern_key == norm_key:
                self.patterns.erase(self.patterns.begin()+i)
            else:
                i += 1

    def has_key(self, key):
        """Check whether the matcher has a rule with a given key.

        key (string or int): The key to check.
        RETURNS (bool): Whether the matcher has the rule.
        """
        key = self._normalize_key(key)
        return key in self._patterns

    def get(self, key, default=None):
        """Retrieve the pattern stored for a key.

        key (str / int): The key to retrieve.
        RETURNS (tuple): The rule, as an (on_match, patterns) tuple.
        """
        key = self._normalize_key(key)
        if key not in self._patterns:
            return default
        return (self._callbacks[key], self._patterns[key])

    def pipe(self, docs, batch_size=1000, n_threads=-1, return_matches=False,
             as_tuples=False):
        """Match a stream of documents, yielding them in turn.

        docs (iterable): A stream of documents.
        batch_size (int): Number of documents to accumulate into a working set.
        return_matches (bool): Yield the match lists along with the docs, making
            results (doc, matches) tuples.
        as_tuples (bool): Interpret the input stream as (doc, context) tuples,
            and yield (result, context) tuples out.
            If both return_matches and as_tuples are True, the output will
            be a sequence of ((doc, matches), context) tuples.
        YIELDS (Doc): Documents, in order.
        """
        if n_threads != -1:
            warnings.warn(Warnings.W016, DeprecationWarning)

        if as_tuples:
            for doc, context in docs:
                matches = self(doc)
                if return_matches:
                    yield ((doc, matches), context)
                else:
                    yield (doc, context)
        else:
            for doc in docs:
                matches = self(doc)
                if return_matches:
                    yield (doc, matches)
                else:
                    yield doc

    def __call__(self, object doclike):
        """Find all token sequences matching the supplied pattern.

        doclike (Doc or Span): The document to match over.
        RETURNS (list): A list of `(key, start, end)` tuples,
            describing the matches. A match tuple describes a span
            `doc[start:end]`. The `label_id` and `key` are both integers.
        """
        if isinstance(doclike, Doc):
            doc = doclike
            length = len(doc)
        elif isinstance(doclike, Span):
            doc = doclike.doc
            length = doclike.end - doclike.start
        else:
            raise ValueError(Errors.E195.format(good="Doc or Span", got=type(doclike).__name__))
        if len(set([LEMMA, POS, TAG]) & self._seen_attrs) > 0 \
          and not doc.is_tagged:
            raise ValueError(Errors.E155.format())
        if DEP in self._seen_attrs and not doc.is_parsed:
            raise ValueError(Errors.E156.format())
        matches = find_matches(&self.patterns[0], self.patterns.size(), doclike, length,
                                extensions=self._extensions, predicates=self._extra_predicates)
        for i, (key, start, end) in enumerate(matches):
            on_match = self._callbacks.get(key, None)
            if on_match is not None:
                on_match(self, doc, i, matches)
        return matches

    def _normalize_key(self, key):
        if isinstance(key, basestring):
            return self.vocab.strings.add(key)
        else:
            return key


def unpickle_matcher(vocab, patterns, callbacks):
    matcher = Matcher(vocab)
    for key, specs in patterns.items():
        callback = callbacks.get(key, None)
        matcher.add(key, callback, *specs)
    return matcher


cdef find_matches(TokenPatternC** patterns, int n, object doclike, int length, extensions=None, predicates=tuple()):
    """Find matches in a doc, with a compiled array of patterns. Matches are
    returned as a list of (id, start, end) tuples.

    To augment the compiled patterns, we optionally also take two Python lists.

    The "predicates" list contains functions that take a Python list and return a
    boolean value. It's mostly used for regular expressions.

    The "extra_getters" list contains functions that take a Python list and return
    an attr ID. It's mostly used for extension attributes.
    """
    cdef vector[PatternStateC] states
    cdef vector[MatchC] matches
    cdef PatternStateC state
    cdef int i, j, nr_extra_attr
    cdef Pool mem = Pool()
    output = []
    if length == 0:
        # avoid any processing or mem alloc if the document is empty
        return output
    if len(predicates) > 0:
        predicate_cache = <char*>mem.alloc(length * len(predicates), sizeof(char))
    if extensions is not None and len(extensions) >= 1:
        nr_extra_attr = max(extensions.values()) + 1
        extra_attr_values = <attr_t*>mem.alloc(length * nr_extra_attr, sizeof(attr_t))
    else:
        nr_extra_attr = 0
        extra_attr_values = <attr_t*>mem.alloc(length, sizeof(attr_t))
    for i, token in enumerate(doclike):
        for name, index in extensions.items():
            value = token._.get(name)
            if isinstance(value, basestring):
                value = token.vocab.strings[value]
            extra_attr_values[i * nr_extra_attr + index] = value
    # Main loop
    cdef int nr_predicate = len(predicates)
    for i in range(length):
        for j in range(n):
            states.push_back(PatternStateC(patterns[j], i, 0))
        transition_states(states, matches, predicate_cache,
            doclike[i], extra_attr_values, predicates)
        extra_attr_values += nr_extra_attr
        predicate_cache += len(predicates)
    # Handle matches that end in 0-width patterns
    finish_states(matches, states)
    seen = set()
    for i in range(matches.size()):
        match = (
            matches[i].pattern_id,
            matches[i].start,
            matches[i].start+matches[i].length
        )
        # We need to deduplicate, because we could otherwise arrive at the same
        # match through two paths, e.g. .?.? matching 'a'. Are we matching the
        # first .?, or the second .? -- it doesn't matter, it's just one match.
        if match not in seen:
            output.append(match)
            seen.add(match)
    return output


cdef void transition_states(vector[PatternStateC]& states, vector[MatchC]& matches,
                            char* cached_py_predicates,
        Token token, const attr_t* extra_attrs, py_predicates) except *:
    cdef int q = 0
    cdef vector[PatternStateC] new_states
    cdef int nr_predicate = len(py_predicates)
    for i in range(states.size()):
        if states[i].pattern.nr_py >= 1:
            update_predicate_cache(cached_py_predicates,
                states[i].pattern, token, py_predicates)
        action = get_action(states[i], token.c, extra_attrs,
                            cached_py_predicates)
        if action == REJECT:
            continue
        # Keep only a subset of states (the active ones). Index q is the
        # states which are still alive. If we reject a state, we overwrite
        # it in the states list, because q doesn't advance.
        state = states[i]
        states[q] = state
        while action in (RETRY, RETRY_ADVANCE, RETRY_EXTEND):
            if action == RETRY_EXTEND:
                # This handles the 'extend'
                new_states.push_back(
                    PatternStateC(pattern=states[q].pattern, start=state.start,
                                  length=state.length+1))
            if action == RETRY_ADVANCE:
                # This handles the 'advance'
                new_states.push_back(
                    PatternStateC(pattern=states[q].pattern+1, start=state.start,
                                  length=state.length+1))
            states[q].pattern += 1
            if states[q].pattern.nr_py != 0:
                update_predicate_cache(cached_py_predicates,
                    states[q].pattern, token, py_predicates)
            action = get_action(states[q], token.c, extra_attrs,
                                cached_py_predicates)
        if action == REJECT:
            pass
        elif action == ADVANCE:
            states[q].pattern += 1
            states[q].length += 1
            q += 1
        else:
            ent_id = get_ent_id(state.pattern)
            if action == MATCH:
                matches.push_back(
                    MatchC(pattern_id=ent_id, start=state.start,
                            length=state.length+1))
            elif action == MATCH_DOUBLE:
                # push match without last token if length > 0
                if state.length > 0:
                    matches.push_back(
                        MatchC(pattern_id=ent_id, start=state.start,
                                length=state.length))
                # push match with last token
                matches.push_back(
                    MatchC(pattern_id=ent_id, start=state.start,
                            length=state.length+1))
            elif action == MATCH_REJECT:
                matches.push_back(
                    MatchC(pattern_id=ent_id, start=state.start,
                            length=state.length))
            elif action == MATCH_EXTEND:
                matches.push_back(
                    MatchC(pattern_id=ent_id, start=state.start,
                           length=state.length))
                states[q].length += 1
                q += 1
    states.resize(q)
    for i in range(new_states.size()):
        states.push_back(new_states[i])


cdef int update_predicate_cache(char* cache,
        const TokenPatternC* pattern, Token token, predicates) except -1:
    # If the state references any extra predicates, check whether they match.
    # These are cached, so that we don't call these potentially expensive
    # Python functions more than we need to.
    for i in range(pattern.nr_py):
        index = pattern.py_predicates[i]
        if cache[index] == 0:
            predicate = predicates[index]
            result = predicate(token)
            if result is True:
                cache[index] = 1
            elif result is False:
                cache[index] = -1
            elif result is None:
                pass
            else:
                raise ValueError(Errors.E125.format(value=result))


cdef void finish_states(vector[MatchC]& matches, vector[PatternStateC]& states) except *:
    """Handle states that end in zero-width patterns."""
    cdef PatternStateC state
    for i in range(states.size()):
        state = states[i]
        while get_quantifier(state) in (ZERO_PLUS, ZERO_ONE):
            is_final = get_is_final(state)
            if is_final:
                ent_id = get_ent_id(state.pattern)
                matches.push_back(
                    MatchC(pattern_id=ent_id, start=state.start, length=state.length))
                break
            else:
                state.pattern += 1


cdef action_t get_action(PatternStateC state,
        const TokenC* token, const attr_t* extra_attrs,
        const char* predicate_matches) nogil:
    """We need to consider:
    a) Does the token match the specification? [Yes, No]
    b) What's the quantifier? [1, 0+, ?]
    c) Is this the last specification? [final, non-final]

    We can transition in the following ways:
    a) Do we emit a match?
    b) Do we add a state with (next state, next token)?
    c) Do we add a state with (next state, same token)?
    d) Do we add a state with (same state, next token)?

    We'll code the actions as boolean strings, so 0000 means no to all 4,
    1000 means match but no states added, etc.

    1:
      Yes, final:
        1000
      Yes, non-final:
        0100
      No, final:
        0000
      No, non-final
        0000
    0+:
      Yes, final:
        1001
      Yes, non-final:
        0011
      No, final:
        1000 (note: Don't include last token!)
      No, non-final:
        0010
    ?:
      Yes, final:
        1000
      Yes, non-final:
        0100
      No, final:
        1000 (note: Don't include last token!)
      No, non-final:
        0010

    Possible combinations:  1000, 0100, 0000, 1001, 0110, 0011, 0010,

    We'll name the bits "match", "advance", "retry", "extend"
    REJECT = 0000
    MATCH = 1000
    ADVANCE = 0100
    RETRY = 0010
    MATCH_EXTEND = 1001
    RETRY_ADVANCE = 0110
    RETRY_EXTEND = 0011
    MATCH_REJECT = 2000 # Match, but don't include last token
    MATCH_DOUBLE = 3000 # Match both with and without last token

    Problem: If a quantifier is matching, we're adding a lot of open partials
    """
    cdef char is_match
    is_match = get_is_match(state, token, extra_attrs, predicate_matches)
    quantifier = get_quantifier(state)
    is_final = get_is_final(state)
    if quantifier == ZERO:
        is_match = not is_match
        quantifier = ONE
    if quantifier == ONE:
      if is_match and is_final:
          # Yes, final: 1000
          return MATCH
      elif is_match and not is_final:
          # Yes, non-final: 0100
          return ADVANCE
      elif not is_match and is_final:
          # No, final: 0000
          return REJECT
      else:
          return REJECT
    elif quantifier == ZERO_PLUS:
      if is_match and is_final:
          # Yes, final: 1001
          return MATCH_EXTEND
      elif is_match and not is_final:
          # Yes, non-final: 0011
          return RETRY_EXTEND
      elif not is_match and is_final:
          # No, final 2000 (note: Don't include last token!)
          return MATCH_REJECT
      else:
          # No, non-final 0010
          return RETRY
    elif quantifier == ZERO_ONE:
      if is_match and is_final:
          # Yes, final: 3000
          # To cater for a pattern ending in "?", we need to add
          # a match both with and without the last token
          return MATCH_DOUBLE
      elif is_match and not is_final:
          # Yes, non-final: 0110
          # We need both branches here, consider a pair like:
          # pattern: .?b string: b
          # If we 'ADVANCE' on the .?, we miss the match.
          return RETRY_ADVANCE
      elif not is_match and is_final:
          # No, final 2000 (note: Don't include last token!)
          return MATCH_REJECT
      else:
          # No, non-final 0010
          return RETRY


cdef char get_is_match(PatternStateC state,
        const TokenC* token, const attr_t* extra_attrs,
        const char* predicate_matches) nogil:
    for i in range(state.pattern.nr_py):
        if predicate_matches[state.pattern.py_predicates[i]] == -1:
            return 0
    spec = state.pattern
    if spec.nr_attr > 0:
        for attr in spec.attrs[:spec.nr_attr]:
            if get_token_attr_for_matcher(token, attr.attr) != attr.value:
                return 0
    for i in range(spec.nr_extra_attr):
        if spec.extra_attrs[i].value != extra_attrs[spec.extra_attrs[i].index]:
            return 0
    return True


cdef char get_is_final(PatternStateC state) nogil:
    if state.pattern[1].nr_attr == 0 and state.pattern[1].attrs != NULL:
        id_attr = state.pattern[1].attrs[0]
        if id_attr.attr != ID:
            with gil:
                raise ValueError(Errors.E074.format(attr=ID, bad_attr=id_attr.attr))
        return 1
    else:
        return 0


cdef char get_quantifier(PatternStateC state) nogil:
    return state.pattern.quantifier


cdef TokenPatternC* init_pattern(Pool mem, attr_t entity_id, object token_specs) except NULL:
    pattern = <TokenPatternC*>mem.alloc(len(token_specs) + 1, sizeof(TokenPatternC))
    cdef int i, index
    for i, (quantifier, spec, extensions, predicates) in enumerate(token_specs):
        pattern[i].quantifier = quantifier
        # Ensure attrs refers to a null pointer if nr_attr == 0
        if len(spec) > 0:
            pattern[i].attrs = <AttrValueC*>mem.alloc(len(spec), sizeof(AttrValueC))
        pattern[i].nr_attr = len(spec)
        for j, (attr, value) in enumerate(spec):
            pattern[i].attrs[j].attr = attr
            pattern[i].attrs[j].value = value
        if len(extensions) > 0:
            pattern[i].extra_attrs = <IndexValueC*>mem.alloc(len(extensions), sizeof(IndexValueC))
        for j, (index, value) in enumerate(extensions):
            pattern[i].extra_attrs[j].index = index
            pattern[i].extra_attrs[j].value = value
        pattern[i].nr_extra_attr = len(extensions)
        if len(predicates) > 0:
            pattern[i].py_predicates = <int32_t*>mem.alloc(len(predicates), sizeof(int32_t))
        for j, index in enumerate(predicates):
            pattern[i].py_predicates[j] = index
        pattern[i].nr_py = len(predicates)
        pattern[i].key = hash64(pattern[i].attrs, pattern[i].nr_attr * sizeof(AttrValueC), 0)
    i = len(token_specs)
    # Even though here, nr_attr == 0, we're storing the ID value in attrs[0] (bug-prone, thread carefully!)
    pattern[i].attrs = <AttrValueC*>mem.alloc(2, sizeof(AttrValueC))
    pattern[i].attrs[0].attr = ID
    pattern[i].attrs[0].value = entity_id
    pattern[i].nr_attr = 0
    pattern[i].nr_extra_attr = 0
    pattern[i].nr_py = 0
    return pattern


cdef attr_t get_ent_id(const TokenPatternC* pattern) nogil:
    # There have been a few bugs here. We used to have two functions,
    # get_ent_id and get_pattern_key that tried to do the same thing. These
    # are now unified to try to solve the "ghost match" problem.
    # Below is the previous implementation of get_ent_id and the comment on it,
    # preserved for reference while we figure out whether the heisenbug in the
    # matcher is resolved.
    #
    #
    #     cdef attr_t get_ent_id(const TokenPatternC* pattern) nogil:
    #         # The code was originally designed to always have pattern[1].attrs.value
    #         # be the ent_id when we get to the end of a pattern. However, Issue #2671
    #         # showed this wasn't the case when we had a reject-and-continue before a
    #         # match.
    #         # The patch to #2671 was wrong though, which came up in #3839.
    #         while pattern.attrs.attr != ID:
    #             pattern += 1
    #         return pattern.attrs.value
    while pattern.nr_attr != 0 or pattern.nr_extra_attr != 0 or pattern.nr_py != 0 \
            or pattern.quantifier != ZERO:
        pattern += 1
    id_attr = pattern[0].attrs[0]
    if id_attr.attr != ID:
        with gil:
            raise ValueError(Errors.E074.format(attr=ID, bad_attr=id_attr.attr))
    return id_attr.value


def _preprocess_pattern(token_specs, string_store, extensions_table, extra_predicates):
    """This function interprets the pattern, converting the various bits of
    syntactic sugar before we compile it into a struct with init_pattern.

    We need to split the pattern up into three parts:
    * Normal attribute/value pairs, which are stored on either the token or lexeme,
        can be handled directly.
    * Extension attributes are handled specially, as we need to prefetch the
        values from Python for the doc before we begin matching.
    * Extra predicates also call Python functions, so we have to create the
        functions and store them. So we store these specially as well.
    * Extension attributes that have extra predicates are stored within the
        extra_predicates.
    """
    tokens = []
    for spec in token_specs:
        if not spec:
            # Signifier for 'any token'
            tokens.append((ONE, [(NULL_ATTR, 0)], [], []))
            continue
        if not isinstance(spec, dict):
            raise ValueError(Errors.E154.format())
        ops = _get_operators(spec)
        attr_values = _get_attr_values(spec, string_store)
        extensions = _get_extensions(spec, string_store, extensions_table)
        predicates = _get_extra_predicates(spec, extra_predicates)
        for op in ops:
            tokens.append((op, list(attr_values), list(extensions), list(predicates)))
    return tokens


def _get_attr_values(spec, string_store):
    attr_values = []
    for attr, value in spec.items():
        if isinstance(attr, basestring):
            attr = attr.upper()
            if attr == '_':
                continue
            elif attr == "OP":
                continue
            if attr == "TEXT":
                attr = "ORTH"
            if attr == "IS_SENT_START":
                attr = "SENT_START"
            attr = IDS.get(attr)
        if isinstance(value, basestring):
            value = string_store.add(value)
        elif isinstance(value, bool):
            value = int(value)
        elif isinstance(value, int):
            pass
        elif isinstance(value, dict):
            continue
        else:
            raise ValueError(Errors.E153.format(vtype=type(value).__name__))
        if attr is not None:
            attr_values.append((attr, value))
        else:
            # should be caught in validation
            raise ValueError(Errors.E152.format(attr=attr))
    return attr_values


# These predicate helper classes are used to match the REGEX, IN, >= etc
# extensions to the matcher introduced in #3173.

class _RegexPredicate(object):
    operators = ("REGEX",)

    def __init__(self, i, attr, value, predicate, is_extension=False):
        self.i = i
        self.attr = attr
        self.value = re.compile(value)
        self.predicate = predicate
        self.is_extension = is_extension
        self.key = (attr, self.predicate, srsly.json_dumps(value, sort_keys=True))
        if self.predicate not in self.operators:
            raise ValueError(Errors.E126.format(good=self.operators, bad=self.predicate))

    def __call__(self, Token token):
        if self.is_extension:
            value = token._.get(self.attr)
        else:
            value = token.vocab.strings[get_token_attr_for_matcher(token.c, self.attr)]
        return bool(self.value.search(value))


class _SetMemberPredicate(object):
    operators = ("IN", "NOT_IN")

    def __init__(self, i, attr, value, predicate, is_extension=False):
        self.i = i
        self.attr = attr
        self.value = set(get_string_id(v) for v in value)
        self.predicate = predicate
        self.is_extension = is_extension
        self.key = (attr, self.predicate, srsly.json_dumps(value, sort_keys=True))
        if self.predicate not in self.operators:
            raise ValueError(Errors.E126.format(good=self.operators, bad=self.predicate))

    def __call__(self, Token token):
        if self.is_extension:
            value = get_string_id(token._.get(self.attr))
        else:
            value = get_token_attr_for_matcher(token.c, self.attr)
        if self.predicate == "IN":
            return value in self.value
        else:
            return value not in self.value

    def __repr__(self):
        return repr(("SetMemberPredicate", self.i, self.attr, self.value, self.predicate))


class _ComparisonPredicate(object):
    operators = ("==", "!=", ">=", "<=", ">", "<")

    def __init__(self, i, attr, value, predicate, is_extension=False):
        self.i = i
        self.attr = attr
        self.value = value
        self.predicate = predicate
        self.is_extension = is_extension
        self.key = (attr, self.predicate, srsly.json_dumps(value, sort_keys=True))
        if self.predicate not in self.operators:
            raise ValueError(Errors.E126.format(good=self.operators, bad=self.predicate))

    def __call__(self, Token token):
        if self.is_extension:
            value = token._.get(self.attr)
        else:
            value = get_token_attr_for_matcher(token.c, self.attr)
        if self.predicate == "==":
            return value == self.value
        if self.predicate == "!=":
            return value != self.value
        elif self.predicate == ">=":
            return value >= self.value
        elif self.predicate == "<=":
            return value <= self.value
        elif self.predicate == ">":
            return value > self.value
        elif self.predicate == "<":
            return value < self.value


def _get_extra_predicates(spec, extra_predicates):
    predicate_types = {
        "REGEX": _RegexPredicate,
        "IN": _SetMemberPredicate,
        "NOT_IN": _SetMemberPredicate,
        "==": _ComparisonPredicate,
        "!=": _ComparisonPredicate,
        ">=": _ComparisonPredicate,
        "<=": _ComparisonPredicate,
        ">": _ComparisonPredicate,
        "<": _ComparisonPredicate,
    }
    seen_predicates = {pred.key: pred.i for pred in extra_predicates}
    output = []
    for attr, value in spec.items():
        if isinstance(attr, basestring):
            if attr == "_":
                output.extend(
                    _get_extension_extra_predicates(
                        value, extra_predicates, predicate_types,
                        seen_predicates))
                continue
            elif attr.upper() == "OP":
                continue
            if attr.upper() == "TEXT":
                attr = "ORTH"
            attr = IDS.get(attr.upper())
        if isinstance(value, dict):
            for type_, cls in predicate_types.items():
                if type_ in value:
                    predicate = cls(len(extra_predicates), attr, value[type_], type_)
                    # Don't create a redundant predicates.
                    # This helps with efficiency, as we're caching the results.
                    if predicate.key in seen_predicates:
                        output.append(seen_predicates[predicate.key])
                    else:
                        extra_predicates.append(predicate)
                        output.append(predicate.i)
                        seen_predicates[predicate.key] = predicate.i
    return output


def _get_extension_extra_predicates(spec, extra_predicates, predicate_types,
        seen_predicates):
    output = []
    for attr, value in spec.items():
        if isinstance(value, dict):
            for type_, cls in predicate_types.items():
                if type_ in value:
                    key = (attr, type_, srsly.json_dumps(value[type_], sort_keys=True))
                    if key in seen_predicates:
                        output.append(seen_predicates[key])
                    else:
                        predicate = cls(len(extra_predicates), attr, value[type_], type_,
                                        is_extension=True)
                        extra_predicates.append(predicate)
                        output.append(predicate.i)
                        seen_predicates[key] = predicate.i
    return output


def _get_operators(spec):
    # Support 'syntactic sugar' operator '+', as combination of ONE, ZERO_PLUS
    lookup = {"*": (ZERO_PLUS,), "+": (ONE, ZERO_PLUS),
              "?": (ZERO_ONE,), "1": (ONE,), "!": (ZERO,)}
    # Fix casing
    spec = {key.upper(): values for key, values in spec.items()
            if isinstance(key, basestring)}
    if "OP" not in spec:
        return (ONE,)
    elif spec["OP"] in lookup:
        return lookup[spec["OP"]]
    else:
        keys = ", ".join(lookup.keys())
        raise ValueError(Errors.E011.format(op=spec["OP"], opts=keys))


def _get_extensions(spec, string_store, name2index):
    attr_values = []
    if not isinstance(spec.get("_", {}), dict):
        raise ValueError(Errors.E154.format())
    for name, value in spec.get("_", {}).items():
        if isinstance(value, dict):
            # Handle predicates (e.g. "IN", in the extra_predicates, not here.
            continue
        if isinstance(value, basestring):
            value = string_store.add(value)
        if name not in name2index:
            name2index[name] = len(name2index)
        attr_values.append((name2index[name], value))
    return attr_values