Fix get_loss for values outside of labels in senter (#5730)

* Fix get_loss for None alignments in senter

When converting the `sent_start` values back to `SentenceRecognizer`
labels, handle `None` alignments.

* Handle SENT_START as -1

Handle SENT_START as -1 (or -1 converted to uint64) by treating any
values other than 1 the same as 0 in `SentenceRecognizer.get_loss`.

spacy/pipeline/pipes.pyx

@@ -523,7 +523,18 @@ class SentenceRecognizer(Tagger):
     def get_loss(self, examples, scores):
         labels = self.labels
         loss_func = SequenceCategoricalCrossentropy(names=labels, normalize=False)
-        truths = [[labels[x] for x in eg.get_aligned("sent_start")] for eg in examples]
+        truths = []
+        for eg in examples:
+            eg_truth = []
+            for x in eg.get_aligned("sent_start"):
+                if x == None:
+                    eg_truth.append(None)
+                elif x == 1:
+                    eg_truth.append(labels[1])
+                else:
+                    # anything other than 1: 0, -1, -1 as uint64
+                    eg_truth.append(labels[0])
+            truths.append(eg_truth)
         d_scores, loss = loss_func(scores, truths)
         if self.model.ops.xp.isnan(loss):
             raise ValueError("nan value when computing loss")

spacy/tests/pipeline/test_senter.py

@@ -38,6 +38,11 @@ def test_overfitting_IO():
     train_examples = []
     for t in TRAIN_DATA:
         train_examples.append(Example.from_dict(nlp.make_doc(t[0]), t[1]))
+    # add some cases where SENT_START == -1
+    train_examples[0].reference[10].is_sent_start = False
+    train_examples[1].reference[1].is_sent_start = False
+    train_examples[1].reference[11].is_sent_start = False
+
     nlp.add_pipe(senter)
     optimizer = nlp.begin_training()