Iterate over lr_edges until sents are correct (#4702)

Iterate over lr_edges until all heads are within the current sentence.
Instead of iterating over them for a fixed number of iterations, check
whether the sentence boundaries are correct for the heads and stop when
all are correct. Stop after a maximum of 10 iterations, providing a
warning in this case since the sentence boundaries may not be correct.

spacy/errors.py

@@ -101,6 +101,7 @@ class Warnings(object):
             "the Knowledge Base.")
     W025 = ("'{name}' requires '{attr}' to be assigned, but none of the "
             "previous components in the pipeline declare that they assign it.")
+    W026 = ("Unable to set all sentence boundaries from dependency parses.")
 
 
 @add_codes

spacy/tests/parser/test_parse.py

@@ -148,3 +148,20 @@ def test_parser_arc_eager_finalize_state(en_tokenizer, en_parser):
     assert tokens[4].left_edge.i == 0
     assert tokens[4].right_edge.i == 4
     assert tokens[4].head.i == 4
+
+
+def test_parser_set_sent_starts(en_vocab):
+    words = ['Ein', 'Satz', '.', 'Außerdem', 'ist', 'Zimmer', 'davon', 'überzeugt', ',', 'dass', 'auch', 'epige-', '\n', 'netische', 'Mechanismen', 'eine', 'Rolle', 'spielen', ',', 'also', 'Vorgänge', ',', 'die', '\n', 'sich', 'darauf', 'auswirken', ',', 'welche', 'Gene', 'abgelesen', 'werden', 'und', '\n', 'welche', 'nicht', '.', '\n']
+    heads = [1, 0, -1, 27, 0, -1, 1, -3, -1, 8, 4, 3, -1, 1, 3, 1, 1, -11, -1, 1, -9, -1, 4, -1, 2, 1, -6, -1, 1, 2, 1, -6, -1, -1, -17, -31, -32, -1]
+    deps = ['nk', 'ROOT', 'punct', 'mo', 'ROOT', 'sb', 'op', 'pd', 'punct', 'cp', 'mo', 'nk', '', 'nk', 'sb', 'nk', 'oa', 're', 'punct', 'mo', 'app', 'punct', 'sb', '', 'oa', 'op', 'rc', 'punct', 'nk', 'sb', 'oc', 're', 'cd', '', 'oa', 'ng', 'punct', '']
+    doc = get_doc(
+        en_vocab, words=words, deps=deps, heads=heads
+    )
+    for i in range(len(words)):
+        if i == 0 or i == 3:
+            assert doc[i].is_sent_start == True
+        else:
+            assert doc[i].is_sent_start == None
+    for sent in doc.sents:
+        for token in sent:
+            assert token.head in sent

spacy/tokens/doc.pxd

@@ -21,6 +21,9 @@ ctypedef fused LexemeOrToken:
 cdef int set_children_from_heads(TokenC* tokens, int length) except -1
 
 
+cdef int _set_lr_kids_and_edges(TokenC* tokens, int length, int loop_count) except -1
+
+
 cdef int token_by_start(const TokenC* tokens, int length, int start_char) except -2
 
 

spacy/tokens/doc.pyx

@@ -1153,35 +1153,69 @@ cdef int set_children_from_heads(TokenC* tokens, int length) except -1:
         tokens[i].r_kids = 0
         tokens[i].l_edge = i
         tokens[i].r_edge = i
-    # Three times, for non-projectivity. See issue #3170. This isn't a very
-    # satisfying fix, but I think it's sufficient.
-    for loop_count in range(3):
-        # Set left edges
-        for i in range(length):
-            child = &tokens[i]
-            head = &tokens[i + child.head]
-            if child < head and loop_count == 0:
-                head.l_kids += 1
-            if child.l_edge < head.l_edge:
-                head.l_edge = child.l_edge
-            if child.r_edge > head.r_edge:
-                head.r_edge = child.r_edge
-        # Set right edges - same as above, but iterate in reverse
-        for i in range(length-1, -1, -1):
-            child = &tokens[i]
-            head = &tokens[i + child.head]
-            if child > head and loop_count == 0:
-                head.r_kids += 1
-            if child.r_edge > head.r_edge:
-                head.r_edge = child.r_edge
-            if child.l_edge < head.l_edge:
-                head.l_edge = child.l_edge
+    cdef int loop_count = 0
+    cdef bint heads_within_sents = False
+    # Try up to 10 iterations of adjusting lr_kids and lr_edges in order to
+    # handle non-projective dependency parses, stopping when all heads are
+    # within their respective sentence boundaries. We have documented cases
+    # that need at least 4 iterations, so this is to be on the safe side
+    # without risking getting stuck in an infinite loop if something is
+    # terribly malformed.
+    while not heads_within_sents:
+        heads_within_sents = _set_lr_kids_and_edges(tokens, length, loop_count)
+        if loop_count > 10:
+            user_warning(Warnings.W026)
+        loop_count += 1
     # Set sentence starts
     for i in range(length):
         if tokens[i].head == 0 and tokens[i].dep != 0:
             tokens[tokens[i].l_edge].sent_start = True
 
 
+cdef int _set_lr_kids_and_edges(TokenC* tokens, int length, int loop_count) except -1:
+    # May be called multiple times due to non-projectivity. See issues #3170
+    # and #4688.
+    # Set left edges
+    cdef TokenC* head
+    cdef TokenC* child
+    cdef int i, j
+    for i in range(length):
+        child = &tokens[i]
+        head = &tokens[i + child.head]
+        if child < head and loop_count == 0:
+            head.l_kids += 1
+        if child.l_edge < head.l_edge:
+            head.l_edge = child.l_edge
+        if child.r_edge > head.r_edge:
+            head.r_edge = child.r_edge
+    # Set right edges - same as above, but iterate in reverse
+    for i in range(length-1, -1, -1):
+        child = &tokens[i]
+        head = &tokens[i + child.head]
+        if child > head and loop_count == 0:
+            head.r_kids += 1
+        if child.r_edge > head.r_edge:
+            head.r_edge = child.r_edge
+        if child.l_edge < head.l_edge:
+            head.l_edge = child.l_edge
+    # Get sentence start positions according to current state
+    sent_starts = set()
+    for i in range(length):
+        if tokens[i].head == 0 and tokens[i].dep != 0:
+            sent_starts.add(tokens[i].l_edge)
+    cdef int curr_sent_start = 0
+    cdef int curr_sent_end = 0
+    # Check whether any heads are not within the current sentence
+    for i in range(length):
+        if (i > 0 and i in sent_starts) or i == length - 1:
+            curr_sent_end = i
+            for j in range(curr_sent_start, curr_sent_end):
+                if tokens[j].head + j < curr_sent_start or tokens[j].head + j >= curr_sent_end + 1:
+                    return False
+            curr_sent_start = i
+    return True
+
+
 cdef int _get_tokens_lca(Token token_j, Token token_k):
     """Given two tokens, returns the index of the lowest common ancestor
     (LCA) among the two. If they have no common ancestor, -1 is returned.