2014-12-16 11:44:43 +00:00
|
|
|
"""
|
|
|
|
MALT-style dependency parser
|
|
|
|
"""
|
|
|
|
from __future__ import unicode_literals
|
|
|
|
cimport cython
|
2014-12-18 22:30:50 +00:00
|
|
|
from libc.stdint cimport uint32_t, uint64_t
|
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
|
|
|
|
|
|
|
|
from cymem.cymem cimport Pool, Address
|
2014-12-18 22:30:50 +00:00
|
|
|
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
|
|
|
|
|
2014-12-18 00:33:25 +00:00
|
|
|
from .arc_eager cimport TransitionSystem, Transition
|
2014-12-16 11:44:43 +00:00
|
|
|
|
2014-12-18 22:30:50 +00:00
|
|
|
from ._state cimport init_state, State, is_final, get_idx, get_s0, get_s1, get_n0, get_n1
|
2014-12-16 11:44:43 +00:00
|
|
|
|
|
|
|
from . import _parse_features
|
|
|
|
from ._parse_features cimport fill_context, CONTEXT_SIZE
|
|
|
|
|
|
|
|
|
|
|
|
DEBUG = False
|
|
|
|
def set_debug(val):
|
|
|
|
global DEBUG
|
|
|
|
DEBUG = val
|
|
|
|
|
|
|
|
|
|
|
|
cdef unicode print_state(State* s, list words):
|
|
|
|
words = list(words) + ['EOL']
|
2014-12-16 16:19:43 +00:00
|
|
|
top = words[s.stack[0]]
|
|
|
|
second = words[s.stack[-1]]
|
2014-12-16 11:44:43 +00:00
|
|
|
n0 = words[s.i]
|
|
|
|
n1 = words[s.i + 1]
|
|
|
|
return ' '.join((second, top, '|', n0, n1))
|
|
|
|
|
|
|
|
|
|
|
|
def get_templates(name):
|
2014-12-17 10:09:29 +00:00
|
|
|
pf = _parse_features
|
2014-12-17 22:05:31 +00:00
|
|
|
if name == 'zhang':
|
2014-12-31 08:40:59 +00:00
|
|
|
return pf.unigrams, pf.arc_eager
|
2014-12-17 22:05:31 +00:00
|
|
|
else:
|
2015-01-05 06:54:13 +00:00
|
|
|
return pf.unigrams, (pf.unigrams + pf.s0_n0 + pf.s1_n0 + pf.s0_n1 + pf.n0_n1 + \
|
2014-12-31 08:40:59 +00:00
|
|
|
pf.tree_shape + pf.trigrams)
|
2014-12-16 11:44:43 +00:00
|
|
|
|
|
|
|
|
|
|
|
cdef class GreedyParser:
|
|
|
|
def __init__(self, model_dir):
|
|
|
|
assert os.path.exists(model_dir) and os.path.isdir(model_dir)
|
|
|
|
self.cfg = Config.read(model_dir, 'config')
|
|
|
|
self.moves = TransitionSystem(self.cfg.left_labels, self.cfg.right_labels)
|
2014-12-31 08:40:59 +00:00
|
|
|
hasty_templ, full_templ = get_templates(self.cfg.features)
|
2015-01-07 14:19:23 +00:00
|
|
|
self.model = Model(self.moves.n_moves, full_templ, model_dir)
|
2014-12-16 11:44:43 +00:00
|
|
|
|
2015-01-17 05:21:17 +00:00
|
|
|
def __call__(self, Tokens tokens):
|
2014-12-16 11:44:43 +00:00
|
|
|
cdef:
|
2014-12-18 00:33:25 +00:00
|
|
|
Transition guess
|
2014-12-18 22:30:50 +00:00
|
|
|
uint64_t state_key
|
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 = init_state(mem, tokens.data, tokens.length)
|
|
|
|
while not is_final(state):
|
2014-12-19 16:48:51 +00:00
|
|
|
fill_context(context, state)
|
2014-12-31 08:40:59 +00:00
|
|
|
scores = self.model.score(context)
|
2014-12-19 16:48:51 +00:00
|
|
|
guess = self.moves.best_valid(scores, state)
|
2014-12-18 00:33:25 +00:00
|
|
|
self.moves.transition(state, &guess)
|
2015-01-24 06:29:04 +00:00
|
|
|
# Messily tell Tokens object the string names of the dependency labels
|
|
|
|
tokens._dep_strings = [None] * len(self.moves.label_ids)
|
|
|
|
for label, id_ in self.moves.label_ids.items():
|
|
|
|
tokens._dep_strings[id_] = label
|
2015-01-25 04:33:54 +00:00
|
|
|
tokens.is_parsed = True
|
2014-12-17 10:09:29 +00:00
|
|
|
return 0
|
2014-12-16 11:44:43 +00:00
|
|
|
|
|
|
|
def train_sent(self, Tokens tokens, list gold_heads, list gold_labels):
|
|
|
|
cdef:
|
2014-12-18 09:41:32 +00:00
|
|
|
const Feature* feats
|
|
|
|
const weight_t* scores
|
2014-12-18 00:33:25 +00:00
|
|
|
Transition guess
|
|
|
|
Transition gold
|
2014-12-16 11:44:43 +00:00
|
|
|
|
|
|
|
cdef int n_feats
|
|
|
|
cdef atom_t[CONTEXT_SIZE] context
|
|
|
|
cdef Pool mem = Pool()
|
2014-12-17 10:09:29 +00:00
|
|
|
cdef int* heads_array = <int*>mem.alloc(tokens.length, sizeof(int))
|
|
|
|
cdef int* labels_array = <int*>mem.alloc(tokens.length, sizeof(int))
|
|
|
|
cdef int i
|
|
|
|
for i in range(tokens.length):
|
|
|
|
heads_array[i] = gold_heads[i]
|
|
|
|
labels_array[i] = self.moves.label_ids[gold_labels[i]]
|
|
|
|
|
2014-12-16 11:44:43 +00:00
|
|
|
cdef State* state = init_state(mem, tokens.data, tokens.length)
|
|
|
|
while not is_final(state):
|
2014-12-31 08:40:59 +00:00
|
|
|
fill_context(context, state)
|
|
|
|
scores = self.model.score(context)
|
2014-12-16 11:44:43 +00:00
|
|
|
guess = self.moves.best_valid(scores, state)
|
2014-12-18 00:33:25 +00:00
|
|
|
best = self.moves.best_gold(&guess, scores, state, heads_array, labels_array)
|
2014-12-31 08:40:59 +00:00
|
|
|
self.model.update(context, guess.clas, best.clas, guess.cost)
|
2014-12-18 00:33:25 +00:00
|
|
|
self.moves.transition(state, &guess)
|
2014-12-17 10:09:29 +00:00
|
|
|
cdef int n_corr = 0
|
2014-12-16 11:44:43 +00:00
|
|
|
for i in range(tokens.length):
|
2014-12-17 22:05:31 +00:00
|
|
|
n_corr += (i + state.sent[i].head) == gold_heads[i]
|
2014-12-16 11:44:43 +00:00
|
|
|
return n_corr
|