2014-12-16 11:44:43 +00:00
|
|
|
"""
|
|
|
|
MALT-style dependency parser
|
|
|
|
"""
|
|
|
|
from __future__ import unicode_literals
|
|
|
|
cimport cython
|
2015-06-10 02:20:23 +00:00
|
|
|
|
|
|
|
from cpython.ref cimport PyObject, Py_INCREF, Py_XDECREF
|
|
|
|
|
2014-12-18 22:30:50 +00:00
|
|
|
from libc.stdint cimport uint32_t, uint64_t
|
2015-06-02 16:38:41 +00:00
|
|
|
from libc.string cimport memset, memcpy
|
2014-12-16 11:44:43 +00:00
|
|
|
import random
|
|
|
|
import os.path
|
2014-12-31 08:40:59 +00:00
|
|
|
from os import path
|
2014-12-16 11:44:43 +00:00
|
|
|
import shutil
|
|
|
|
import json
|
2015-07-08 17:52:30 +00:00
|
|
|
import sys
|
2014-12-16 11:44:43 +00:00
|
|
|
|
|
|
|
from cymem.cymem cimport Pool, Address
|
2014-12-18 22:30:50 +00:00
|
|
|
from murmurhash.mrmr cimport hash64
|
2015-06-08 12:49:04 +00:00
|
|
|
from thinc.typedefs cimport weight_t, class_t, feat_t, atom_t, hash_t
|
2014-12-16 11:44:43 +00:00
|
|
|
|
|
|
|
from util import Config
|
|
|
|
|
2015-07-14 21:47:03 +00:00
|
|
|
from thinc.api cimport Example, ExampleC
|
2014-12-16 11:44:43 +00:00
|
|
|
|
2015-06-01 22:28:02 +00:00
|
|
|
|
2015-07-13 18:20:58 +00:00
|
|
|
from ..structs cimport TokenC
|
|
|
|
|
|
|
|
from ..tokens.doc cimport Doc
|
2015-03-14 15:06:35 +00:00
|
|
|
from ..strings cimport StringStore
|
2014-12-16 11:44:43 +00:00
|
|
|
|
2015-06-11 23:50:23 +00:00
|
|
|
|
2015-02-22 05:32:07 +00:00
|
|
|
from .transition_system import OracleError
|
2015-06-14 17:01:26 +00:00
|
|
|
from .transition_system cimport TransitionSystem, Transition
|
2014-12-16 11:44:43 +00:00
|
|
|
|
2015-05-24 19:35:02 +00:00
|
|
|
from ..gold cimport GoldParse
|
2014-12-16 11:44:43 +00:00
|
|
|
|
|
|
|
from . import _parse_features
|
2015-06-09 19:20:14 +00:00
|
|
|
from ._parse_features cimport CONTEXT_SIZE
|
2015-06-09 21:23:28 +00:00
|
|
|
from ._parse_features cimport fill_context
|
|
|
|
from .stateclass cimport StateClass
|
2014-12-16 11:44:43 +00:00
|
|
|
|
2015-07-14 21:47:03 +00:00
|
|
|
from .._ml cimport arg_max_if_true
|
|
|
|
|
2014-12-16 11:44:43 +00:00
|
|
|
|
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
|
|
|
|
|
|
|
|
|
|
|
|
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':
|
2015-03-10 17:00:23 +00:00
|
|
|
return pf.ner
|
2015-03-24 03:29:01 +00:00
|
|
|
elif name == 'debug':
|
|
|
|
return pf.unigrams
|
2015-06-28 09:36:11 +00:00
|
|
|
elif name.startswith('embed'):
|
2015-06-27 02:18:47 +00:00
|
|
|
return (pf.words, pf.tags, pf.labels)
|
2014-12-17 22:05:31 +00:00
|
|
|
else:
|
2015-06-14 19:17:39 +00:00
|
|
|
return (pf.unigrams + pf.s0_n0 + pf.s1_n0 + pf.s1_s0 + pf.s0_n1 + pf.n0_n1 + \
|
2015-02-21 04:30:31 +00:00
|
|
|
pf.tree_shape + pf.trigrams)
|
2014-12-16 11:44:43 +00:00
|
|
|
|
|
|
|
|
2015-07-08 10:35:46 +00:00
|
|
|
def ParserFactory(transition_system):
|
|
|
|
return lambda strings, dir_: Parser(strings, dir_, transition_system)
|
|
|
|
|
|
|
|
|
2015-06-01 22:28:02 +00:00
|
|
|
cdef class Parser:
|
2015-03-14 15:06:35 +00:00
|
|
|
def __init__(self, StringStore strings, model_dir, transition_system):
|
2015-07-08 17:52:30 +00:00
|
|
|
if not os.path.exists(model_dir):
|
|
|
|
print >> sys.stderr, "Warning: No model found at", model_dir
|
|
|
|
elif not os.path.isdir(model_dir):
|
|
|
|
print >> sys.stderr, "Warning: model path:", model_dir, "is not a directory"
|
2015-07-08 17:59:16 +00:00
|
|
|
else:
|
|
|
|
self.cfg = Config.read(model_dir, 'config')
|
2015-07-08 18:04:55 +00:00
|
|
|
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
|
|
|
|
2015-07-08 16:53:00 +00:00
|
|
|
def __call__(self, Doc tokens):
|
2015-06-10 00:03:38 +00:00
|
|
|
cdef StateClass stcls = StateClass.init(tokens.data, tokens.length)
|
|
|
|
self.moves.initialize_state(stcls)
|
2015-06-26 04:25:36 +00:00
|
|
|
|
2015-06-28 20:36:03 +00:00
|
|
|
cdef Example eg = Example(self.model.n_classes, CONTEXT_SIZE,
|
|
|
|
self.model.n_feats, self.model.n_feats)
|
2015-07-24 02:53:27 +00:00
|
|
|
with nogil:
|
|
|
|
self.parse(stcls, eg.c)
|
2015-07-14 21:47:03 +00:00
|
|
|
tokens.set_parse(stcls._sent)
|
|
|
|
|
2015-08-08 23:31:54 +00:00
|
|
|
def get_state(self, Doc tokens, initial_actions, continue_for=0):
|
2015-08-08 21:32:15 +00:00
|
|
|
cdef StateClass stcls = StateClass.init(tokens.data, tokens.length)
|
|
|
|
self.moves.initialize_state(stcls)
|
|
|
|
cdef object action_name
|
|
|
|
cdef Transition action
|
2015-08-08 23:31:54 +00:00
|
|
|
cdef Example eg
|
2015-08-08 21:32:15 +00:00
|
|
|
for action_name in initial_actions:
|
2015-08-08 23:31:54 +00:00
|
|
|
try:
|
|
|
|
action = self.moves.lookup_transition(action_name)
|
|
|
|
except IndexError:
|
|
|
|
break
|
2015-08-08 21:32:15 +00:00
|
|
|
action.do(stcls, action.label)
|
2015-08-08 23:31:54 +00:00
|
|
|
else:
|
|
|
|
eg = Example(self.model.n_classes, CONTEXT_SIZE,
|
|
|
|
self.model.n_feats, self.model.n_feats)
|
|
|
|
while not stcls.is_final() and continue_for != 0:
|
|
|
|
memset(eg.c.scores, 0, eg.c.nr_class * sizeof(weight_t))
|
|
|
|
self.moves.set_valid(eg.c.is_valid, stcls)
|
|
|
|
fill_context(eg.c.atoms, stcls)
|
|
|
|
self.model.set_scores(eg.c.scores, eg.c.atoms)
|
|
|
|
eg.guess = arg_max_if_true(eg.c.scores, eg.c.is_valid, self.model.n_classes)
|
2015-08-08 23:40:13 +00:00
|
|
|
self.moves.c[eg.guess].do(stcls, self.moves.c[eg.guess].label)
|
2015-08-08 23:31:54 +00:00
|
|
|
if stcls.is_final():
|
|
|
|
self.moves.finalize_state(stcls)
|
2015-08-08 21:32:15 +00:00
|
|
|
tokens.set_parse(stcls._sent)
|
|
|
|
return stcls
|
|
|
|
|
2015-07-14 21:47:03 +00:00
|
|
|
cdef void parse(self, StateClass stcls, ExampleC eg) nogil:
|
2015-06-09 23:35:28 +00:00
|
|
|
while not stcls.is_final():
|
2015-07-14 21:47:03 +00:00
|
|
|
memset(eg.scores, 0, eg.nr_class * sizeof(weight_t))
|
2015-07-17 20:37:24 +00:00
|
|
|
self.moves.set_valid(eg.is_valid, stcls)
|
2015-07-14 21:47:03 +00:00
|
|
|
fill_context(eg.atoms, stcls)
|
|
|
|
self.model.set_scores(eg.scores, eg.atoms)
|
|
|
|
eg.guess = arg_max_if_true(eg.scores, eg.is_valid, self.model.n_classes)
|
|
|
|
self.moves.c[eg.guess].do(stcls, self.moves.c[eg.guess].label)
|
2015-06-09 23:35:28 +00:00
|
|
|
self.moves.finalize_state(stcls)
|
2015-06-01 22:28:02 +00:00
|
|
|
|
2015-07-14 12:11:23 +00:00
|
|
|
def train(self, Doc tokens, GoldParse gold):
|
2015-06-01 23:34:19 +00:00
|
|
|
self.moves.preprocess_gold(gold)
|
2015-06-10 00:03:38 +00:00
|
|
|
cdef StateClass stcls = StateClass.init(tokens.data, tokens.length)
|
|
|
|
self.moves.initialize_state(stcls)
|
2015-06-28 20:36:03 +00:00
|
|
|
cdef Example eg = Example(self.model.n_classes, CONTEXT_SIZE,
|
|
|
|
self.model.n_feats, self.model.n_feats)
|
2015-06-30 12:26:32 +00:00
|
|
|
cdef weight_t loss = 0
|
2015-06-09 23:35:28 +00:00
|
|
|
words = [w.orth_ for w in tokens]
|
2015-06-30 12:26:32 +00:00
|
|
|
cdef Transition G
|
2015-06-09 23:35:28 +00:00
|
|
|
while not stcls.is_final():
|
2015-06-28 20:36:03 +00:00
|
|
|
memset(eg.c.scores, 0, eg.c.nr_class * sizeof(weight_t))
|
2015-08-08 21:32:15 +00:00
|
|
|
|
2015-07-17 20:37:24 +00:00
|
|
|
self.moves.set_costs(eg.c.is_valid, eg.c.costs, stcls, gold)
|
2015-08-08 21:32:15 +00:00
|
|
|
|
2015-06-28 20:36:03 +00:00
|
|
|
fill_context(eg.c.atoms, stcls)
|
2015-06-01 22:28:02 +00:00
|
|
|
|
2015-06-26 04:25:36 +00:00
|
|
|
self.model.train(eg)
|
2015-06-01 22:28:02 +00:00
|
|
|
|
2015-06-30 12:26:32 +00:00
|
|
|
G = self.moves.c[eg.c.guess]
|
2015-06-01 22:28:02 +00:00
|
|
|
|
2015-06-28 20:36:03 +00:00
|
|
|
self.moves.c[eg.c.guess].do(stcls, self.moves.c[eg.c.guess].label)
|
2015-06-30 12:26:32 +00:00
|
|
|
loss += eg.c.loss
|
|
|
|
return loss
|