more correct interpretation of BERT tokens

5 years ago · 5b218430dc
parent 29eb154fb8
commit 5b218430dc
1 changed files with 3 additions and 0 deletions
--- a/qurator/sbb_ner/webapp/app.py
+++ b/qurator/sbb_ner/webapp/app.py
@ -236,6 +236,9 @@ def ner(model_id=None):

                word += original_text[orig_pos]

+                if word_pred != 'X':
+                    last_prediction = word_pred
+
                continue

            if not token.startswith('##') and word_pred == 'X':