diff --git a/ref-test/common/data_tools.py b/ref-test/common/data_tools.py
index 0dfadf1..d406e7b 100644
--- a/ref-test/common/data_tools.py
+++ b/ref-test/common/data_tools.py
@@ -128,7 +128,9 @@ def evaluate_answers(dataset: dict, answers: dict):
             if str(q_no) in answers:
                 correct = block['correct']
                 correct_answer = block['options'][correct]
-                if answers[str(q_no)] == correct_answer:
+                submitted_answer = answers[str(q_no)]
+                submitted_answer = submitted_answer.replace('‘', '‘').replace('’', '’').replace('—', '—')
+                if submitted_answer == correct_answer:
                     score += 1
                     for tag in block['tags']:
                         if tag not in tags: