added weighted token level metric

evaluation_metrics.py

@@ -83,9 +83,30 @@ class TokenMacroMetric(EvaluationMetric):
         )
 
 
+class TokenWeightedMetric(EvaluationMetric):
+    def __init__(self) -> None:
+        super().__init__()
+
+        self.name = "Token Based Evaluation with Weighted Average"
+        self.description = ""
+
+    @staticmethod
+    def get_evaluation_metric(gt_ner_span, pred_ner_span, text, tags) -> float:
+        return round(
+            classification_report(
+                get_token_output_labels(gt_ner_span, text),
+                get_token_output_labels(pred_ner_span, text),
+                labels=tags,
+                output_dict=True,
+            )["weighted avg"]["f1-score"],
+            2,
+        )
+
+
 EVALUATION_METRICS = [
     PartialSpanOverlapMetric(),
     ExactSpanOverlapMetric(),
     TokenMicroMetric(),
     TokenMacroMetric(),
+    TokenWeightedMetric(),
 ]