Use complete snippets as suggested by BLURB paper

bioasq_task_b.py

@@ -643,15 +643,19 @@ class BioasqTaskBDataset(datasets.GeneratorBasedBuilder):
                         continue
                     if self.config.type_subset is not None and self.config.type_subset != record["type"]:
                         continue
-                    for i, snippet in enumerate(record["snippets"]):
-                        yield uid, {
-                            "id": f'{record["id"]}_{i}',
-                            "document_id": snippet["document"],
-                            "question_id": record["id"],
-                            "question": record["body"],
-                            "type": record["type"],
-                            "choices": [],
-                            "context": snippet["text"],
-                            "answer": self._get_exact_answer(record),
-                        }
-                        uid += 1
+                    snippets = record['snippets']
+                    snippets_text = [snippet['text'].replace(
+                        '\n', ' ') for snippet in snippets]
+                    passage = " ".join(snippets_text)
+
+                    yield uid, {
+                        "id": record["id"],
+                        "document_id": record["id"],
+                        "question_id": record["id"],
+                        "question": record["body"],
+                        "type": record["type"],
+                        "choices": [],
+                        "context": passage,
+                        "answer": self._get_exact_answer(record),
+                    }
+                    uid += 1