* Refactor sense tagger to get rid of intermediary layers

This commit is contained in:
Matthew Honnibal 2015-07-03 13:31:11 +02:00
parent 6735439abf
commit 2fbcdd0ea8

View File

@ -1,13 +1,17 @@
from thinc.api cimport Example
from thinc.typedefs cimport atom_t
from .typedefs cimport flags_t from .typedefs cimport flags_t
from .structs cimport TokenC from .structs cimport TokenC
from .strings cimport StringStore from .strings cimport StringStore
from .tokens cimport Tokens from .tokens cimport Tokens
from ._ml cimport Model
from .senses cimport POS_SENSES, N_SENSES, encode_sense_strs from .senses cimport POS_SENSES, N_SENSES, encode_sense_strs
from .gold cimport GoldParse from .gold cimport GoldParse
from .parts_of_speech cimport NOUN, VERB
from thinc.learner cimport LinearModel
from thinc.features cimport Extractor
from thinc.typedefs cimport atom_t, weight_t, feat_t
from os import path
@ -173,6 +177,8 @@ cdef int fill_token(atom_t* ctxt, const TokenC* token) except -1:
cdef int fill_context(atom_t* ctxt, const TokenC* token) except -1: cdef int fill_context(atom_t* ctxt, const TokenC* token) except -1:
# NB: we have padding to keep us safe here
# See tokens.pyx
fill_token(&ctxt[P2W], token - 2) fill_token(&ctxt[P2W], token - 2)
fill_token(&ctxt[P1W], token - 1) fill_token(&ctxt[P1W], token - 1)
@ -185,62 +191,79 @@ cdef int fill_context(atom_t* ctxt, const TokenC* token) except -1:
cdef class SenseTagger: cdef class SenseTagger:
cdef readonly StringStore strings cdef readonly StringStore strings
cdef readonly Model model cdef readonly LinearModel model
cdef readonly Extractor extractor
cdef readonly model_dir
def __init__(self, StringStore strings, model_dir): def __init__(self, StringStore strings, model_dir):
self.strings = strings if model_dir is not None and path.isdir(model_dir):
model_dir = path.join(model_dir, 'model')
templates = unigrams + bigrams + trigrams templates = unigrams + bigrams + trigrams
self.model = Model(N_SENSES, templates, model_dir) self.extractor = Extractor(templates)
self.model = LinearModel(N_SENSES, self.extractor.n_templ)
self.model_dir = model_dir
if self.model_dir and path.exists(self.model_dir):
self.model.load(self.model_dir, freq_thresh=0)
self.strings = strings
def __call__(self, Tokens tokens): def __call__(self, Tokens tokens):
eg = Example(self.model.n_classes, CONTEXT_SIZE, self.model.n_feats, cdef atom_t[CONTEXT_SIZE] context
self.model.n_feats) cdef int i, guess, n_feats
cdef int i cdef const TokenC* token
for i in range(tokens.length): for i in range(tokens.length):
n_valid = self._set_valid(<bint*>eg.c.is_valid, pos_senses(&tokens.data[i])) token = &tokens.data[i]
if n_valid >= 1: if token.pos in (NOUN, VERB):
fill_context(eg.c.atoms, &tokens.data[i]) fill_context(context, token)
self.model.predict(eg) feats = self.extractor.get_feats(context, &n_feats)
tokens.data[i].sense = eg.c.guess scores = self.model.get_scores(feats, n_feats)
tokens.data[i].sense = self.best_in_set(scores, POS_SENSES[<int>token.pos])
def train(self, Tokens tokens, GoldParse gold): def train(self, Tokens tokens, GoldParse gold):
eg = Example(self.model.n_classes, CONTEXT_SIZE, self.model.n_feats+1, cdef int i, j
self.model.n_feats+1)
cdef int i
for i, ssenses in enumerate(gold.ssenses): for i, ssenses in enumerate(gold.ssenses):
if ssenses: if ssenses:
gold.c.ssenses[i] = encode_sense_strs(ssenses) gold.c.ssenses[i] = encode_sense_strs(ssenses)
else: else:
gold.c.ssenses[i] = pos_senses(&tokens.data[i]) gold.c.ssenses[i] = pos_senses(&tokens.data[i])
cdef atom_t[CONTEXT_SIZE] context
cdef int n_feats
cdef feat_t f_key
cdef int f_i
cdef int cost = 0 cdef int cost = 0
for i in range(tokens.length): for i in range(tokens.length):
if tokens.data[i].lex.senses == 0 or tokens.data[i].lex.senses == 1: token = &tokens.data[i]
continue if token.pos in (NOUN, VERB) \
self._set_costs(<bint*>eg.c.is_valid, eg.c.costs, gold.c.ssenses[i]) and token.lex.senses >= 2 \
fill_context(eg.c.atoms, &tokens.data[i]) and gold.c.ssenses[i] >= 2:
fill_context(context, token)
self.model.train(eg) feats = self.extractor.get_feats(context, &n_feats)
scores = self.model.get_scores(feats, n_feats)
tokens.data[i].sense = eg.c.guess token.sense = self.best_in_set(scores, POS_SENSES[<int>token.pos])
cost += eg.c.cost best = self.best_in_set(scores, gold.c.ssenses[i])
guess_counts = {}
gold_counts = {}
if token.sense != best:
for j in range(n_feats):
f_key = feats[j].key
f_i = feats[j].i
feat = (f_i, f_key)
gold_counts[feat] = gold_counts.get(feat, 0) + 1.0
guess_counts[feat] = guess_counts.get(feat, 0) - 1.0
#self.model.update({token.sense: guess_counts, best: gold_counts})
return cost return cost
cdef int _set_valid(self, bint* is_valid, flags_t senses) except -1: cdef int best_in_set(self, const weight_t* scores, flags_t senses) except -1:
cdef int n_valid cdef weight_t max_ = 0
cdef flags_t bit cdef int argmax = -1
is_valid[0] = False cdef flags_t i
for bit in range(1, N_SENSES): for i in range(N_SENSES):
is_valid[bit] = senses & (1 << bit) if (senses & (1 << i)) and (argmax == -1 or scores[i] > max_):
n_valid += is_valid[bit] max_ = scores[i]
return n_valid argmax = i
assert argmax >= 0
cdef int _set_costs(self, bint* is_valid, int* costs, flags_t senses): return argmax
cdef flags_t bit
is_valid[0] = False
costs[0] = 1
for bit in range(1, N_SENSES):
is_valid[bit] = True
costs[bit] = 0 if (senses & (1 << bit)) else 1
cdef flags_t pos_senses(const TokenC* token) nogil: cdef flags_t pos_senses(const TokenC* token) nogil: