chinhon's picture
Update app.py
93853e6
import gradio as gr
import numpy as np
import pandas as pd
import re
import shap
from transformers import (
AutoTokenizer,
AutoModelForSequenceClassification,
TextClassificationPipeline,
)
tokenizer = AutoTokenizer.from_pretrained("chinhon/fake_tweet_detect")
model = AutoModelForSequenceClassification.from_pretrained("chinhon/fake_tweet_detect")
tweet_detector = TextClassificationPipeline(model=model, tokenizer=tokenizer)
# tweak the extent of text cleaning as you wish
def clean_text(text):
text = re.sub(r"http\S+", "", text)
text = re.sub(r"\n", " ", text)
text = re.sub(r"\'t", " not", text) # Change 't to 'not'
text = re.sub(r"(@.*?)[\s]", " ", text) # Remove @name
text = re.sub(r"$\d+\W+|\b\d+\b|\W+\d+$", " ", text) # remove digits
text = re.sub(r"[^\w\s\#]", "", text) # remove special characters except hashtags
text = text.strip(" ")
text = re.sub(
" +", " ", text
).strip() # get rid of multiple spaces and replace with a single
return text
def tweet_detect(text):
data = [clean_text(text)]
prediction = tweet_detector(data)
pred_label = [x.get("label") for x in prediction]
if pred_label == ["LABEL_1"]:
return "Fake Tweet"
elif pred_label == ["LABEL_0"]:
return "Real Tweet"
#Define Gradio interface
gradio_ui = gr.Interface(
fn=tweet_detect,
title="Detect Fake Tweets",
description="Enter a tweet and see if a Distilbert model can identify if it was written by state-backed trolls. DISCLAIMER: While the model was fine tuned on 100k real and troll tweets, and achieved high accuracy in my tests, its performance drops significantly against the day-to-day barrage of content on Twitter. As such, this app is intended as an example for understanding the limits of AI/ML in highly complex problems like fake media detection, and not as a final arbiter of whether someone's tweet is real or not.",
inputs=gr.Textbox(lines=10, label="Paste tweet text here [English Only]"),
outputs=gr.Label(type="auto", label="Prediction"),
interpretation="shap",
article="Details of the fine tuning and tests are in this Medium post: https://bit.ly/3tueP36",
)
gradio_ui.launch(enable_queue=True)