Add support for the null answer in `QuestionAnsweringPipeline` (#3441)

* Add support for the null answer in `QuestionAnsweringPipeline`

* black

* Fix min null score computation

* Fix a PR comment

src/transformers/pipelines.py

@@ -944,6 +944,7 @@ class QuestionAnsweringPipeline(Pipeline):
         kwargs.setdefault("max_answer_len", 15)
         kwargs.setdefault("max_seq_len", 384)
         kwargs.setdefault("max_question_len", 64)
+        kwargs.setdefault("handle_impossible_answer", False)
 
         if kwargs["topk"] < 1:
             raise ValueError("topk parameter should be >= 1 (got {})".format(kwargs["topk"]))
@@ -982,6 +983,7 @@ class QuestionAnsweringPipeline(Pipeline):
                         start, end = self.model(**fw_args)
                         start, end = start.cpu().numpy(), end.cpu().numpy()
 
+            min_null_score = 1000000  # large and positive
             answers = []
             for (feature, start_, end_) in zip(features, start, end):
                 # Normalize logits and spans to retrieve the answer
@@ -994,8 +996,9 @@ class QuestionAnsweringPipeline(Pipeline):
                     end_ * np.abs(np.array(feature.p_mask) - 1),
                 )
 
-                # TODO : What happens if not possible
-                # Mask CLS
+                if kwargs["handle_impossible_answer"]:
+                    min_null_score = min(min_null_score, (start_[0] * end_[0]).item())
+
                 start_[0] = end_[0] = 0
 
                 starts, ends, scores = self.decode(start_, end_, kwargs["topk"], kwargs["max_answer_len"])
@@ -1013,6 +1016,10 @@ class QuestionAnsweringPipeline(Pipeline):
                     }
                     for s, e, score in zip(starts, ends, scores)
                 ]
+
+            if kwargs["handle_impossible_answer"]:
+                answers.append({"score": min_null_score, "start": 0, "end": 0, "answer": ""})
+
             answers = sorted(answers, key=lambda x: x["score"], reverse=True)[: kwargs["topk"]]
             all_answers += answers