spaCy/spacy/syntax/parser.pyx

125 lines
3.8 KiB
Cython
Raw Normal View History

2014-12-16 11:44:43 +00:00
"""
MALT-style dependency parser
"""
from __future__ import unicode_literals
cimport cython
from libc.stdint cimport uint32_t, uint64_t
2014-12-16 11:44:43 +00:00
import random
import os.path
from os import path
2014-12-16 11:44:43 +00:00
import shutil
import json
from cymem.cymem cimport Pool, Address
from murmurhash.mrmr cimport hash64
2014-12-16 11:44:43 +00:00
from thinc.typedefs cimport weight_t, class_t, feat_t, atom_t
from util import Config
from thinc.features cimport Extractor
from thinc.features cimport Feature
from thinc.features cimport count_feats
from thinc.learner cimport LinearModel
from ..tokens cimport Tokens, TokenC
from ..strings cimport StringStore
2014-12-16 11:44:43 +00:00
from .arc_eager cimport TransitionSystem, Transition
from .transition_system import OracleError
2014-12-16 11:44:43 +00:00
from ._state cimport new_state, State, is_final, get_idx, get_s0, get_s1, get_n0, get_n1
from ..gold cimport GoldParse
2014-12-16 11:44:43 +00:00
from . import _parse_features
from ._parse_features cimport fill_context, CONTEXT_SIZE
2015-04-19 08:31:31 +00:00
DEBUG = False
2014-12-16 11:44:43 +00:00
def set_debug(val):
global DEBUG
DEBUG = val
cdef unicode print_state(State* s, list words):
words = list(words) + ['EOL']
2015-01-28 16:18:29 +00:00
top = words[s.stack[0]] + '_%d' % s.sent[s.stack[0]].head
second = words[s.stack[-1]] + '_%d' % s.sent[s.stack[-1]].head
third = words[s.stack[-2]] + '_%d' % s.sent[s.stack[-2]].head
2015-03-27 16:29:58 +00:00
n0 = words[s.i] if s.i < len(words) else 'EOL'
n1 = words[s.i + 1] if s.i+1 < len(words) else 'EOL'
if s.ents_len:
ent = '%s %d-%d' % (s.ent.label, s.ent.start, s.ent.end)
else:
ent = '-'
return ' '.join((ent, str(s.stack_len), third, second, top, '|', n0, n1))
2014-12-16 11:44:43 +00:00
def get_templates(name):
2014-12-17 10:09:29 +00:00
pf = _parse_features
2015-03-24 04:08:35 +00:00
if name == 'ner':
return pf.ner
elif name == 'debug':
return pf.unigrams
else:
return (pf.unigrams + pf.s0_n0 + pf.s1_n0 + pf.s0_n1 + pf.n0_n1 + \
pf.tree_shape + pf.trigrams)
2014-12-16 11:44:43 +00:00
cdef class GreedyParser:
def __init__(self, StringStore strings, model_dir, transition_system):
2014-12-16 11:44:43 +00:00
assert os.path.exists(model_dir) and os.path.isdir(model_dir)
self.cfg = Config.read(model_dir, 'config')
self.moves = transition_system(strings, self.cfg.labels)
templates = get_templates(self.cfg.features)
self.model = Model(self.moves.n_moves, templates, model_dir)
2014-12-16 11:44:43 +00:00
def __call__(self, Tokens tokens):
if tokens.length == 0:
return 0
2014-12-16 11:44:43 +00:00
cdef atom_t[CONTEXT_SIZE] context
cdef int n_feats
cdef Pool mem = Pool()
cdef State* state = new_state(mem, tokens.data, tokens.length)
self.moves.initialize_state(state)
cdef Transition guess
2014-12-16 11:44:43 +00:00
while not is_final(state):
fill_context(context, state)
scores = self.model.score(context, False)
guess = self.moves.best_valid(scores, state)
2015-02-18 09:41:06 +00:00
guess.do(&guess, state)
self.moves.finalize_state(state)
tokens.set_parse(state.sent)
2014-12-17 10:09:29 +00:00
return 0
2014-12-16 11:44:43 +00:00
def train(self, Tokens tokens, GoldParse gold):
2015-05-11 14:12:03 +00:00
py_words = [w.orth_ for w in tokens]
self.moves.preprocess_gold(gold)
2014-12-16 11:44:43 +00:00
cdef Pool mem = Pool()
cdef State* state = new_state(mem, tokens.data, tokens.length)
self.moves.initialize_state(state)
cdef int cost
cdef const Feature* feats
cdef const weight_t* scores
cdef Transition guess
cdef Transition best
cdef atom_t[CONTEXT_SIZE] context
loss = 0
2014-12-16 11:44:43 +00:00
while not is_final(state):
fill_context(context, state)
scores = self.model.score(context, True)
2014-12-16 11:44:43 +00:00
guess = self.moves.best_valid(scores, state)
best = self.moves.best_gold(scores, state, gold)
2015-04-19 08:31:31 +00:00
cost = guess.get_cost(&guess, state, gold)
self.model.update(context, guess.clas, best.clas, cost)
guess.do(&guess, state)
loss += cost
self.moves.finalize_state(state)
return loss