diff --git a/src/transformers/pipelines/token_classification.py b/src/transformers/pipelines/token_classification.py
index 5dc243d1acf2..f3c78d3498d5 100644
--- a/src/transformers/pipelines/token_classification.py
+++ b/src/transformers/pipelines/token_classification.py
@@ -304,7 +304,9 @@ def gather_pre_entities(
                         start_ind = start_ind.item()
                         end_ind = end_ind.item()
                 word_ref = sentence[start_ind:end_ind]
-                if getattr(self.tokenizer._tokenizer.model, "continuing_subword_prefix", None):
+                if getattr(self.tokenizer, "_tokenizer", None) and getattr(
+                    self.tokenizer._tokenizer.model, "continuing_subword_prefix", None
+                ):
                     # This is a BPE, word aware tokenizer, there is a correct way
                     # to fuse tokens
                     is_subword = len(word) != len(word_ref)