diff --git a/qurator/sbb_ner/ground_truth/data_processor.py b/qurator/sbb_ner/ground_truth/data_processor.py
index 577448b..85245f8 100644
--- a/qurator/sbb_ner/ground_truth/data_processor.py
+++ b/qurator/sbb_ner/ground_truth/data_processor.py
@@ -388,6 +388,9 @@ def convert_examples_to_features(example, label_map, max_seq_len, tokenizer):
     for i, word in enumerate(example.text_a):  # example.text_a is a sequence of words
 
         token = tokenizer.tokenize(word)
+
+        # import ipdb;ipdb.set_trace()
+
         tokens.extend(token)
 
         label_1 = example.label[i] if i < len(example.label) else 'O'
diff --git a/qurator/sbb_ner/webapp/app.py b/qurator/sbb_ner/webapp/app.py
index b1d566f..cbeb25b 100644
--- a/qurator/sbb_ner/webapp/app.py
+++ b/qurator/sbb_ner/webapp/app.py
@@ -219,57 +219,88 @@ def ner(model_id=None):
 
     output = []
 
-    for (tokens, word_predictions),  (input_sentence, _) in zip(prediction, sentences):
+    for (tokens, token_predictions),  (input_sentence, _) in zip(prediction, sentences):
 
+        output_text = ""
         original_text = "".join(input_sentence)
         original_word_positions = \
             [pos for positions in [[idx] * len(word) for idx, word in enumerate(input_sentence)] for pos in positions]
 
         word = ''
-        last_prediction = 'O'
+        word_prediction = 'O'
         output_sentence = []
 
-        for pos, (token, word_pred) in enumerate(zip(tokens, word_predictions)):
+        for pos, (token, token_prediction) in enumerate(zip(tokens, token_predictions)):
 
-            if word_pred == '[SEP]':
-                word_pred = 'O'
+            if not token.startswith('##') and token_prediction == 'X' or token_prediction == '[SEP]':
+                token_prediction = 'O'
 
-            if not token.startswith('##') and token != '[UNK]':
-                if len(word) > 0:
-                    output_sentence.append({'word': word, 'prediction': last_prediction})
+            orig_pos = len(output_text + word)
 
+            # if the current word length is greater than 0
+            # and its either a word start token (does not start with ##) and not an unknown token or the original text
+            # positions indicate a word break
+            if len(word) > 0 and ((not token.startswith('##') and token != '[UNK]') or
+                                  (orig_pos > 0 and
+                                   original_word_positions[orig_pos-1] != original_word_positions[orig_pos])):
+                output_sentence.append({'word': word, 'prediction': word_prediction})
+                output_text += word
                 word = ''
+                word_prediction = 'O'
 
             if token == '[UNK]':
 
-                orig_pos = len("".join([pred['word'] for pred in output_sentence]) + word)
+                orig_pos = len(output_text + word)
 
+                # are we on a word boundary?
                 if orig_pos > 0 and original_word_positions[orig_pos-1] != original_word_positions[orig_pos]:
-                    output_sentence.append({'word': word, 'prediction': last_prediction})
+
+                    # we are on a word boundary - start a new word ...
+                    output_sentence.append({'word': word, 'prediction': word_prediction})
+                    output_text += word
                     word = ''
+                    word_prediction = 'O'
 
-                word += original_text[orig_pos]
+                # get character that corresponds to [UNK] token from original text
+                token = original_text[orig_pos]
 
-                if word_pred != 'X':
-                    last_prediction = word_pred
+            else:
+                token = token[2:] if token.startswith('##') else token
 
-                continue
+            # if the output_text plus the current word and token is not a prefix of the original text, it means,
+            # we would miss characters. Therefore we take the missing characters from the original text at the current
+            # word position
+            while not original_text.startswith(output_text + word + token) \
+                    and len(output_text + word) < len(original_text):
 
-            if not token.startswith('##') and word_pred == 'X':
-                word_pred = 'O'
+                word += original_text[len(output_text + word)]
 
-            token = token[2:] if token.startswith('##') else token
+                orig_pos = len(output_text + word)
+
+                # are we on a word boundary?
+                if orig_pos > 0 and original_word_positions[orig_pos - 1] != original_word_positions[orig_pos]:
+                    # we are on a word boundary - start a new word ...
+                    output_sentence.append({'word': word, 'prediction': word_prediction})
+                    output_text += word
+                    word = ''
+                    word_prediction = 'O'
 
             word += token
 
-            if word_pred != 'X':
-                last_prediction = word_pred
+            if token_prediction != 'X':
+                word_prediction = token_prediction
 
         if len(word) > 0:
-            output_sentence.append({'word': word, 'prediction': last_prediction})
+            output_text += word
+            output_sentence.append({'word': word, 'prediction': word_prediction})
 
         output.append(output_sentence)
 
+        try:
+            assert output_text == original_text
+        except AssertionError:
+            import ipdb;ipdb.set_trace()
+
     for output_sentence, (input_sentence, _) in zip(output, sentences):
 
         try:
@@ -278,6 +309,7 @@ def ner(model_id=None):
             logger.warning('Input and output different!!! \n\n\nInput: {}\n\nOutput: {}\n'.
                            format("".join(input_sentence).replace(" ", ""),
                                   "".join([pred['word'] for pred in output_sentence])))
+
     torch.cuda.empty_cache()
 
     return jsonify(output)