Merge pull request #346 from wbwseeker/sentbnd_bug

introduce sentence boundaries for additional root tokens

spacy/syntax/arc_eager.pyx

@@ -399,31 +399,34 @@ cdef class ArcEager(TransitionSystem):
         cdef TokenC* orig_head
         cdef int new_edge
         cdef int child_i
-        cdef TokenC* head_i
+        cdef int head_i
         for i in range(st.length):
             if st._sent[i].head == 0 and st._sent[i].dep == 0:
                 st._sent[i].dep = self.root_label
             # If we're not using the Break transition, we segment via root-labelled
             # arcs between the root words.
             elif USE_ROOT_ARC_SEGMENT and st._sent[i].dep == self.root_label:
-                orig_head_id = st._sent[i].head
+                orig_head_id = i + st._sent[i].head
                 orig_head = &st._sent[orig_head_id]
                 if i < orig_head_id: # i is left dependent
                     orig_head.l_kids -= 1
                     if i == orig_head.l_edge: # i is left-most child
                         # find the second left-most child and make it the new l_edge
                         new_edge = orig_head_id
-                        child_i = i
+                        child_i = i+1
                         while child_i < orig_head_id:
-                            if st._sent[child_i].head == orig_head_id:
+                            if child_i + st._sent[child_i].head == orig_head_id:
                                 new_edge = child_i
+                                break
                             child_i += 1
                         # then walk up the path to root and update the l_edges of all ancestors
                         # the logic here works because the tree is guaranteed to be projective
-                        head_i = &st._sent[orig_head.head]
-                        while head_i.l_edge == orig_head.l_edge:
-                            head_i.l_edge = new_edge
-                            head_i = &st._sent[head_i.head]
+                        head_i = orig_head_id + orig_head.head
+                        while st._sent[head_i].l_edge == orig_head.l_edge:
+                            st._sent[head_i].l_edge = new_edge
+                            if st._sent[head_i].head == 0:
+                                break
+                            head_i += st._sent[head_i].head
                         orig_head.l_edge = new_edge
 
                 elif i > orig_head_id: # i is right dependent
@@ -431,22 +434,27 @@ cdef class ArcEager(TransitionSystem):
                     if i == orig_head.r_edge:
                         # find the second right-most child and make it the new r_edge
                         new_edge = orig_head_id
-                        child_i = i
+                        child_i = i-1
                         while child_i > orig_head_id:
-                            if st._sent[child_i].head == orig_head_id:
+                            if child_i + st._sent[child_i].head == orig_head_id:
                                 new_edge = child_i
+                                break
                             child_i -= 1
-                        # then walk up the path to root and update the l_edges of all ancestors
+                        # then walk up the path to root and update the r_edges of all ancestors
                         # the logic here works because the tree is guaranteed to be projective
-                        head_i = &st._sent[orig_head.head]
-                        while head_i.r_edge == orig_head.r_edge:
-                            head_i.r_edge = new_edge
-                            head_i = &st._sent[head_i.head]
+                        head_i = orig_head_id + orig_head.head
+                        while st._sent[head_i].r_edge == orig_head.r_edge:
+                            st._sent[head_i].r_edge = new_edge
+                            if st._sent[head_i].head == 0:
+                                break
+                            head_i += st._sent[head_i].head
                         orig_head.r_edge = new_edge
 
-                # note that this can create non-projective trees if there are arcs
+                # note that this may create non-projective trees if there are arcs
                 # between nodes on both sides of the new root node
                 st._sent[i].head = 0
+                st._sent[st._sent[i].l_edge].sent_start = True
+
 
     cdef int set_valid(self, int* output, const StateC* st) nogil:
         cdef bint[N_MOVES] is_valid

spacy/tests/parser/test_parse.py

@@ -19,3 +19,80 @@ def test_one_word_sentence(EN):
 	with EN.parser.step_through(doc) as _:
 		pass
 	assert doc[0].dep != 0
+
+
+def apply_transition_sequence(model, doc, sequence):
+    with model.parser.step_through(doc) as stepwise:
+        for transition in sequence:
+            stepwise.transition(transition)
+
+
+@pytest.mark.models
+def test_arc_eager_finalize_state(EN):
+	# right branching
+	example = EN.tokenizer.tokens_from_list(u"a b c d e".split(' '))
+	apply_transition_sequence(EN, example, ['R-nsubj','D','R-nsubj','R-nsubj','D','R-ROOT'])
+
+	assert example[0].n_lefts == 0
+	assert example[0].n_rights == 2
+	assert example[0].left_edge.i == 0
+	assert example[0].right_edge.i == 3
+	assert example[0].head.i == 0
+
+	assert example[1].n_lefts == 0
+	assert example[1].n_rights == 0
+	assert example[1].left_edge.i == 1
+	assert example[1].right_edge.i == 1
+	assert example[1].head.i == 0
+
+	assert example[2].n_lefts == 0
+	assert example[2].n_rights == 1
+	assert example[2].left_edge.i == 2
+	assert example[2].right_edge.i == 3
+	assert example[2].head.i == 0
+
+	assert example[3].n_lefts == 0
+	assert example[3].n_rights == 0
+	assert example[3].left_edge.i == 3
+	assert example[3].right_edge.i == 3
+	assert example[3].head.i == 2
+
+	assert example[4].n_lefts == 0
+	assert example[4].n_rights == 0
+	assert example[4].left_edge.i == 4
+	assert example[4].right_edge.i == 4
+	assert example[4].head.i == 4
+
+	# left branching
+	example = EN.tokenizer.tokens_from_list(u"a b c d e".split(' '))
+	apply_transition_sequence(EN, example, ['S','L-nsubj','L-ROOT','S','L-nsubj','L-nsubj'])
+
+	assert example[0].n_lefts == 0
+	assert example[0].n_rights == 0
+	assert example[0].left_edge.i == 0
+	assert example[0].right_edge.i == 0
+	assert example[0].head.i == 0
+
+	assert example[1].n_lefts == 0
+	assert example[1].n_rights == 0
+	assert example[1].left_edge.i == 1
+	assert example[1].right_edge.i == 1
+	assert example[1].head.i == 2
+
+	assert example[2].n_lefts == 1
+	assert example[2].n_rights == 0
+	assert example[2].left_edge.i == 1
+	assert example[2].right_edge.i == 2
+	assert example[2].head.i == 4
+
+	assert example[3].n_lefts == 0
+	assert example[3].n_rights == 0
+	assert example[3].left_edge.i == 3
+	assert example[3].right_edge.i == 3
+	assert example[3].head.i == 4
+
+	assert example[4].n_lefts == 2
+	assert example[4].n_rights == 0
+	assert example[4].left_edge.i == 1
+	assert example[4].right_edge.i == 4
+	assert example[4].head.i == 4

spacy/tests/parser/test_sbd.py

@@ -1,7 +1,7 @@
 from __future__ import unicode_literals
 
 import pytest
-
+from spacy.tokens import Doc
 
 
 @pytest.mark.models
@@ -42,7 +42,7 @@ def test_single_question(EN):
 
 @pytest.mark.models
 def test_sentence_breaks_no_space(EN):
-    doc = EN.tokenizer.tokens_from_list('This is a sentence . This is another one .'.split(' '))
+    doc = EN.tokenizer.tokens_from_list(u'This is a sentence . This is another one .'.split(' '))
     EN.tagger(doc)
     with EN.parser.step_through(doc) as stepwise:
         # stack empty, automatic Shift (This)
@@ -83,7 +83,7 @@ def test_sentence_breaks_no_space(EN):
 
 @pytest.mark.models
 def test_sentence_breaks_with_space(EN):
-    doc = EN.tokenizer.tokens_from_list('\t This is \n a sentence \n \n . \n \t \n This is another \t one .'.split(' '))
+    doc = EN.tokenizer.tokens_from_list(u'\t This is \n a sentence \n \n . \n \t \n This is another \t one .'.split(' '))
     EN.tagger(doc)
     with EN.parser.step_through(doc) as stepwise:
         # stack empty, automatic Shift (This)
@@ -120,3 +120,56 @@ def test_sentence_breaks_with_space(EN):
     for tok in doc:
         assert tok.dep != 0 or tok.is_space
     assert [ tok.head.i for tok in doc ] == [1,2,2,2,5,2,5,5,2,8,8,8,13,13,16,14,13,13]
+
+
+
+def apply_transition_sequence(model, doc, sequence):
+    with model.parser.step_through(doc) as stepwise:
+        for transition in sequence:
+            stepwise.transition(transition)
+
+
+@pytest.mark.models
+def test_sbd_for_root_label_dependents(EN):
+    """
+    make sure that the parser properly introduces a sentence boundary without
+    the break transition by checking for dependents with the root label
+    """
+    example = EN.tokenizer.tokens_from_list(u"I saw a firefly It glowed".split(' '))
+    EN.tagger(example)
+    apply_transition_sequence(EN, example, ['L-nsubj','S','L-det','R-dobj','D','S','L-nsubj','R-ROOT'])
+
+    assert example[1].head.i == 1
+    assert example[5].head.i == 5
+
+    sents = list(example.sents)
+    assert len(sents) == 2
+    assert sents[1][0].orth_ == u'It'
+
+
+
+@pytest.mark.models
+def test_sbd_serialization(EN):
+    """
+    test that before and after serialization, the sentence boundaries are the same even
+    if the parser predicted two roots for the sentence that were made into two sentences
+    after parsing by arc_eager.finalize()
+
+    This is actually an interaction between the sentence boundary prediction and doc.from_array
+    The process is the following: if the parser doesn't predict a sentence boundary but attaches
+    a word with the ROOT label, the second root node is made root of its own sentence after parsing.
+    During serialization, sentence boundary information is lost and reintroduced when the code
+    is deserialized by introducing sentence starts at every left-edge of every root node.
+
+    BUG that is tested here: So far, the parser wasn't introducing a sentence start when 
+    it introduced the second root node.
+    """
+
+    example = EN.tokenizer.tokens_from_list(u"I bought a couch from IKEA. It was n't very comfortable .".split(' '))
+    EN.tagger(example)
+    apply_transition_sequence(EN, example, ['L-nsubj','S','L-det','R-dobj','D','R-prep','R-pobj','D','D','S','L-nsubj','R-ROOT','R-neg','D','S','L-advmod','R-acomp','D','R-punct'])
+
+    example_serialized = Doc(EN.vocab).from_bytes(example.to_bytes())
+
+    assert example.to_bytes() == example_serialized.to_bytes()
+    assert [s.text for s in example.sents] == [s.text for s in example_serialized.sents]