spaCy/spacy/syntax/arc_eager.pyx
2015-06-02 18:38:07 +02:00

457 lines
15 KiB
Cython

from __future__ import unicode_literals
from ._state cimport State
from ._state cimport has_head, get_idx, get_s0, get_n0, get_left, get_right
from ._state cimport is_final, at_eol, pop_stack, push_stack, add_dep
from ._state cimport head_in_buffer, children_in_buffer
from ._state cimport head_in_stack, children_in_stack
from ._state cimport count_left_kids
from ..structs cimport TokenC
from .transition_system cimport do_func_t, get_cost_func_t
from ..gold cimport GoldParse
DEF NON_MONOTONIC = True
DEF USE_BREAK = True
cdef weight_t MIN_SCORE = -90000
# Break transition from here
# http://www.aclweb.org/anthology/P13-1074
cdef enum:
SHIFT
REDUCE
LEFT
RIGHT
BREAK
CONSTITUENT
ADJUST
N_MOVES
MOVE_NAMES = [None] * N_MOVES
MOVE_NAMES[SHIFT] = 'S'
MOVE_NAMES[REDUCE] = 'D'
MOVE_NAMES[LEFT] = 'L'
MOVE_NAMES[RIGHT] = 'R'
MOVE_NAMES[BREAK] = 'B'
MOVE_NAMES[CONSTITUENT] = 'C'
MOVE_NAMES[ADJUST] = 'A'
cdef class ArcEager(TransitionSystem):
@classmethod
def get_labels(cls, gold_parses):
move_labels = {SHIFT: {'': True}, REDUCE: {'': True}, RIGHT: {},
LEFT: {'ROOT': True}, BREAK: {'ROOT': True},
CONSTITUENT: {}, ADJUST: {'': True}}
for raw_text, sents in gold_parses:
for (ids, words, tags, heads, labels, iob), ctnts in sents:
for child, head, label in zip(ids, heads, labels):
if label != 'ROOT':
if head < child:
move_labels[RIGHT][label] = True
elif head > child:
move_labels[LEFT][label] = True
for start, end, label in ctnts:
move_labels[CONSTITUENT][label] = True
return move_labels
cdef int preprocess_gold(self, GoldParse gold) except -1:
for i in range(gold.length):
if gold.heads[i] is None: # Missing values
gold.c_heads[i] = i
gold.c_labels[i] = -1
else:
gold.c_heads[i] = gold.heads[i]
gold.c_labels[i] = self.strings[gold.labels[i]]
for end, brackets in gold.brackets.items():
for start, label_strs in brackets.items():
gold.c_brackets[start][end] = 1
for label_str in label_strs:
# Add the encoded label to the set
gold.brackets[end][start].add(self.strings[label_str])
cdef Transition lookup_transition(self, object name) except *:
if '-' in name:
move_str, label_str = name.split('-', 1)
label = self.label_ids[label_str]
else:
label = 0
move = MOVE_NAMES.index(move_str)
for i in range(self.n_moves):
if self.c[i].move == move and self.c[i].label == label:
return self.c[i]
def move_name(self, int move, int label):
label_str = self.strings[label]
if label_str:
return MOVE_NAMES[move] + '-' + label_str
else:
return MOVE_NAMES[move]
cdef Transition init_transition(self, int clas, int move, int label) except *:
# TODO: Apparent Cython bug here when we try to use the Transition()
# constructor with the function pointers
cdef Transition t
t.score = 0
t.clas = clas
t.move = move
t.label = label
if move == SHIFT:
t.do = _do_shift
t.get_cost = _shift_cost
elif move == REDUCE:
t.do = _do_reduce
t.get_cost = _reduce_cost
elif move == LEFT:
t.do = _do_left
t.get_cost = _left_cost
elif move == RIGHT:
t.do = _do_right
t.get_cost = _right_cost
elif move == BREAK:
t.get_cost = _break_cost
elif move == CONSTITUENT:
t.get_cost = _constituent_cost
elif move == ADJUST:
t.do = _do_adjust
t.get_cost = _adjust_cost
else:
raise Exception(move)
return t
cdef int initialize_state(self, State* state) except -1:
push_stack(state)
cdef int finalize_state(self, State* state) except -1:
cdef int root_label = self.strings['ROOT']
for i in range(state.sent_len):
if state.sent[i].head == 0 and state.sent[i].dep == 0:
state.sent[i].dep = root_label
cdef int set_valid(self, bint* output, const State* s) except -1:
cdef bint[N_MOVES] is_valid
is_valid[SHIFT] = _can_shift(s)
is_valid[REDUCE] = _can_reduce(s)
is_valid[LEFT] = _can_left(s)
is_valid[RIGHT] = _can_right(s)
is_valid[BREAK] = _can_break(s)
is_valid[CONSTITUENT] = _can_constituent(s)
is_valid[ADJUST] = _can_adjust(s)
cdef int i
for i in range(self.n_moves):
output[i] = is_valid[self.c[i].move]
cdef Transition best_valid(self, const weight_t* scores, const State* s) except *:
cdef bint[N_MOVES] is_valid
is_valid[SHIFT] = _can_shift(s)
is_valid[REDUCE] = _can_reduce(s)
is_valid[LEFT] = _can_left(s)
is_valid[RIGHT] = _can_right(s)
is_valid[BREAK] = _can_break(s)
is_valid[CONSTITUENT] = _can_constituent(s)
is_valid[ADJUST] = _can_adjust(s)
cdef Transition best
cdef weight_t score = MIN_SCORE
cdef int i
for i in range(self.n_moves):
if scores[i] > score and is_valid[self.c[i].move]:
best = self.c[i]
score = scores[i]
assert best.clas < self.n_moves
assert score > MIN_SCORE
# Label Shift moves with the best Right-Arc label, for non-monotonic
# actions
if best.move == SHIFT:
score = MIN_SCORE
for i in range(self.n_moves):
if self.c[i].move == RIGHT and scores[i] > score:
best.label = self.c[i].label
score = scores[i]
return best
cdef int _do_shift(const Transition* self, State* state) except -1:
# Set the dep label, in case we need it after we reduce
if NON_MONOTONIC:
state.sent[state.i].dep = self.label
push_stack(state)
cdef int _do_left(const Transition* self, State* state) except -1:
# Interpret left-arcs from EOL as attachment to root
if at_eol(state):
add_dep(state, state.stack[0], state.stack[0], self.label)
else:
add_dep(state, state.i, state.stack[0], self.label)
pop_stack(state)
cdef int _do_right(const Transition* self, State* state) except -1:
add_dep(state, state.stack[0], state.i, self.label)
push_stack(state)
cdef int _do_reduce(const Transition* self, State* state) except -1:
if NON_MONOTONIC and not has_head(get_s0(state)):
add_dep(state, state.stack[-1], state.stack[0], get_s0(state).dep)
pop_stack(state)
cdef int _do_break(const Transition* self, State* state) except -1:
state.sent[state.i-1].sent_end = True
while state.stack_len != 0:
if get_s0(state).head == 0:
get_s0(state).dep = self.label
state.stack -= 1
state.stack_len -= 1
if not at_eol(state):
push_stack(state)
cdef int _shift_cost(const Transition* self, const State* s, GoldParse gold) except -1:
if not _can_shift(s):
return 9000
cost = 0
cost += head_in_stack(s, s.i, gold.c_heads)
cost += children_in_stack(s, s.i, gold.c_heads)
# If we can break, and there's no cost to doing so, we should
if _can_break(s) and _break_cost(self, s, gold) == 0:
cost += 1
return cost
cdef int _right_cost(const Transition* self, const State* s, GoldParse gold) except -1:
if not _can_right(s):
return 9000
cost = 0
if gold.c_heads[s.i] == s.stack[0]:
cost += self.label != gold.c_labels[s.i]
return cost
# This indicates missing head
if gold.c_labels[s.i] != -1:
cost += head_in_buffer(s, s.i, gold.c_heads)
cost += children_in_stack(s, s.i, gold.c_heads)
cost += head_in_stack(s, s.i, gold.c_heads)
return cost
cdef int _left_cost(const Transition* self, const State* s, GoldParse gold) except -1:
if not _can_left(s):
return 9000
cost = 0
if gold.c_heads[s.stack[0]] == s.i:
cost += self.label != gold.c_labels[s.stack[0]]
return cost
# If we're at EOL, then the left arc will add an arc to ROOT.
elif at_eol(s):
# Are we root?
if gold.c_labels[s.stack[0]] != -1:
# If we're at EOL, prefer to reduce or break over left-arc
if _can_reduce(s) or _can_break(s):
cost += gold.c_heads[s.stack[0]] != s.stack[0]
# Are we labelling correctly?
cost += self.label != gold.c_labels[s.stack[0]]
return cost
cost += head_in_buffer(s, s.stack[0], gold.c_heads)
cost += children_in_buffer(s, s.stack[0], gold.c_heads)
if NON_MONOTONIC and s.stack_len >= 2:
cost += gold.c_heads[s.stack[0]] == s.stack[-1]
if gold.c_labels[s.stack[0]] != -1:
cost += gold.c_heads[s.stack[0]] == s.stack[0]
return cost
cdef int _reduce_cost(const Transition* self, const State* s, GoldParse gold) except -1:
if not _can_reduce(s):
return 9000
cdef int cost = 0
cost += children_in_buffer(s, s.stack[0], gold.c_heads)
if NON_MONOTONIC:
cost += head_in_buffer(s, s.stack[0], gold.c_heads)
return cost
cdef int _break_cost(const Transition* self, const State* s, GoldParse gold) except -1:
if not _can_break(s):
return 9000
# When we break, we Reduce all of the words on the stack.
cdef int cost = 0
# Number of deps between S0...Sn and N0...Nn
for i in range(s.i, s.sent_len):
cost += children_in_stack(s, i, gold.c_heads)
cost += head_in_stack(s, i, gold.c_heads)
return cost
cdef inline bint _can_shift(const State* s) nogil:
return not at_eol(s)
cdef inline bint _can_right(const State* s) nogil:
return s.stack_len >= 1 and not at_eol(s)
cdef inline bint _can_left(const State* s) nogil:
if NON_MONOTONIC:
return s.stack_len >= 1 #and not missing_brackets(s)
else:
return s.stack_len >= 1 and not has_head(get_s0(s))
cdef inline bint _can_reduce(const State* s) nogil:
if NON_MONOTONIC:
return s.stack_len >= 2 #and not missing_brackets(s)
else:
return s.stack_len >= 2 and has_head(get_s0(s))
cdef inline bint _can_break(const State* s) nogil:
cdef int i
if not USE_BREAK:
return False
elif at_eol(s):
return False
#elif NON_MONOTONIC:
# return True
else:
# In the Break transition paper, they have this constraint that prevents
# Break if stack is disconnected. But, if we're doing non-monotonic parsing,
# we prefer to relax this constraint. This is helpful in parsing whole
# documents, because then we don't get stuck with words on the stack.
seen_headless = False
for i in range(s.stack_len):
if s.sent[s.stack[-i]].head == 0:
if seen_headless:
return False
else:
seen_headless = True
# TODO: Constituency constraints
return True
cdef inline bint _can_constituent(const State* s) nogil:
if s.stack_len < 1:
return False
return False
#else:
# # If all stack elements are popped, can't constituent
# for i in range(s.ctnts.stack_len):
# if not s.ctnts.is_popped[-i]:
# return True
# else:
# return False
cdef inline bint _can_adjust(const State* s) nogil:
return False
#if s.ctnts.stack_len < 2:
# return False
#cdef const Constituent* b1 = s.ctnts.stack[-1]
#cdef const Constituent* b0 = s.ctnts.stack[0]
#if (b1.head + b1.head.head) != b0.head:
# return False
#elif b0.head >= b1.head:
# return False
#elif b0 >= b1:
# return False
cdef int _constituent_cost(const Transition* self, const State* s, GoldParse gold) except -1:
if not _can_constituent(s):
return 9000
raise Exception("Constituent move should be disabled currently")
# The gold standard is indexed by end, then by start, then a set of labels
#brackets = gold.brackets(get_s0(s).r_edge, {})
#if not brackets:
# return 2 # 2 loss for bad bracket, only 1 for good bracket bad label
# Index the current brackets in the state
#existing = set()
#for i in range(s.ctnt_len):
# if ctnt.end == s.r_edge and ctnt.label == self.label:
# existing.add(ctnt.start)
#cdef int loss = 2
#cdef const TokenC* child
#cdef const TokenC* s0 = get_s0(s)
#cdef int n_left = count_left_kids(s0)
# Iterate over the possible start positions, and check whether we have a
# (start, end, label) match to the gold tree
#for i in range(1, n_left):
# child = get_left(s, s0, i)
# if child.l_edge in brackets and child.l_edge not in existing:
# if self.label in brackets[child.l_edge]
# return 0
# else:
# loss = 1 # If we see the start position, set loss to 1
#return loss
cdef int _adjust_cost(const Transition* self, const State* s, GoldParse gold) except -1:
if not _can_adjust(s):
return 9000
raise Exception("Adjust move should be disabled currently")
# The gold standard is indexed by end, then by start, then a set of labels
#gold_starts = gold.brackets(get_s0(s).r_edge, {})
# Case 1: There are 0 brackets ending at this word.
# --> Cost is sunk, but must allow brackets to begin
#if not gold_starts:
# return 0
# Is the top bracket correct?
#gold_labels = gold_starts.get(s.ctnt.start, set())
# TODO: Case where we have a unary rule
# TODO: Case where two brackets end on this word, with top bracket starting
# before
#cdef const TokenC* child
#cdef const TokenC* s0 = get_s0(s)
#cdef int n_left = count_left_kids(s0)
#cdef int i
# Iterate over the possible start positions, and check whether we have a
# (start, end, label) match to the gold tree
#for i in range(1, n_left):
# child = get_left(s, s0, i)
# if child.l_edge in brackets:
# if self.label in brackets[child.l_edge]:
# return 0
# else:
# loss = 1 # If we see the start position, set loss to 1
#return loss
cdef int _do_constituent(const Transition* self, State* state) except -1:
return False
#cdef Constituent* bracket = new_bracket(state.ctnts)
#bracket.parent = NULL
#bracket.label = self.label
#bracket.head = get_s0(state)
#bracket.length = 0
#attach(bracket, state.ctnts.stack)
# Attach rightward children. They're in the brackets array somewhere
# between here and B0.
#cdef Constituent* node
#cdef const TokenC* node_gov
#for i in range(1, bracket - state.ctnts.stack):
# node = bracket - i
# node_gov = node.head + node.head.head
# if node_gov == bracket.head:
# attach(bracket, node)
cdef int _do_adjust(const Transition* self, State* state) except -1:
return False
#cdef Constituent* b0 = state.ctnts.stack[0]
#cdef Constituent* b1 = state.ctnts.stack[1]
#assert (b1.head + b1.head.head) == b0.head
#assert b0.head < b1.head
#assert b0 < b1
#attach(b0, b1)
## Pop B1 from stack, but keep B0 on top
#state.ctnts.stack -= 1
#state.ctnts.stack[0] = b0