app.py

import os

os.system("pip install transformers")

import transformers

from transformers.pipelines.token_classification import TokenClassificationPipeline

class MyPipeline(TokenClassificationPipeline):
  def preprocess(self, sentence, offset_mapping=None):
      truncation = True if self.tokenizer.model_max_length and self.tokenizer.model_max_length > 0 else False
      model_inputs = self.tokenizer(
          sentence,
          return_tensors=self.framework,
          truncation=truncation,
          return_special_tokens_mask=True,
          return_offsets_mapping=self.tokenizer.is_fast,
      )


      length = len(model_inputs['input_ids'][0]) - 2
      tokens = tokenizer.tokenize(sentence)
      seek = 0
      offset_mapping_list = [[(0, 0)]]
      for i in range(length):
        if tokens[i][-2:] == '@@':
          offset_mapping_list[0].append((seek, seek + len(tokens[i]) - 2))
          seek += len(tokens[i]) - 2            
        else:
          offset_mapping_list[0].append((seek, seek + len(tokens[i])))
          seek += len(tokens[i]) + 1
      offset_mapping_list[0].append((0, 0))

      # if offset_mapping:
      #     model_inputs["offset_mapping"] = offset_mapping

      model_inputs['offset_mapping'] = offset_mapping_list
      model_inputs["sentence"] = sentence

      return model_inputs

model_checkpoint = "DD0101/disfluency-base"

my_classifier = pipeline(
  "token-classification", model=model_checkpoint, aggregation_strategy="simple", pipeline_class=MyPipeline)


import gradio as gr

def ner(text):
  output = my_classifier(text)
  for entity in output:
    entity['entity'] = entity.pop('entity_group')
  return {'text': text, 'entities': output}

examples = dataset['test'][:10]['text']

demo = gr.Interface(ner, 
                    gr.Textbox(label='Text', placeholder="Enter sentence here..."),
                    gr.HighlightedText(label='Highlighted Output'),
                    examples=examples,
                    title="Disfluency Detection",
                    description="This is an easy-to-use built in Gradio for desmontrating a NER System that identifies disfluency-entities in \
                    Vietnamese utterances",
                    theme=gr.themes.Soft())

demo.launch()