import gradio as gr import requests import random import time import pandas as pd from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline from game1 import read1, func1, interpre1, func1_written from game2 import func2 from game3 import read3, func3, interpre3, func3_written def ret_en(): return 'en' def ret_nl(): return 'nl' def reset_scores(): data = pd.DataFrame( { "Role": ["AI 🤖", "HUMAN 👨👩"], "Scores": [0, 0], } ) tot_scores = ''' ###

Machine   ''' + str(int(0)) + '''   VS   ''' + str(int(0)) + '''   Human

''' # scroe_human = ''' # Human: ''' + str(int(0)) # scroe_robot = ''' # Robot: ''' + str(int(0)) # tooltip=["Role", "Scores"], return 0, 0, tot_scores def reset_modules(): res_empty = {"original": "", "interpretation": []} return res_empty, 0, 0, [], "" with gr.Blocks(theme=gr.themes.Default(text_size=gr.themes.sizes.text_md)) as demo: pre_load_1 = pipeline("sentiment-analysis", model="nlptown/bert-base-multilingual-uncased-sentiment") pre_load_2 = pipeline("text-classification", model='DTAI-KULeuven/robbert-v2-dutch-sentiment') pre_load_3 = pipeline("text-classification", model='distilbert-base-uncased-finetuned-sst-2-english') pre_load_4 = pipeline("text-classification", model="padmajabfrl/Gender-Classification") with gr.Row(): num1 = gr.Number(value=0, container=False, show_label=False, visible=False) num2 = gr.Number(value=0, container=False, show_label=False, visible=False) with gr.Column(scale=2): placeholder = gr.Markdown( ''' ## Welcome to the Language Model Explanation Challenge! Language Models (LMs) are powerful AI tools to understand and generate human language.
However, they sometimes make mistakes... and it's hard to know why!

Are *humans* or *machines* better at understanding language?
→ Play a game against AI to find out!
Does AI think like you or not at all?
→ Check out the color highlighting to see which parts of the sentence are more important for the machine.
Can you outsmart the AI?
→ Try to write a text that will trick it into the wrong decision

Choose one of the three tasks below ... and start to play! ''' #* **Like or Dislike** provides a movie/food/book review. You (and AI) are required to guess its score. #The one with the correct or close answer win the score. #* **Human or Machine** provides a paragraph. You (and AI) need to judge if it is written by humans or machines. #The one with the correct or close answer win the score. #* **Man or Woman** allows you to write a text. #If you could successfully trick the AI into guessing the wrong gender, you get the score. ) # with gr.Column(scale=1): # logo = gr.Image('logo.png', height=230, width=600, min_width=80, show_label=False, show_share_button=False, interactive=False, container=False) with gr.Tab("Like or Dislike"): text_en = gr.Textbox(label="", value="en", visible=False) text_nl = gr.Textbox(label="", value="nl", visible=False) lang_selected = gr.Textbox(label="", value="", visible=False) num_selected_1 = gr.Number(value=0, container=False, show_label=False, visible=False) with gr.Row(): with gr.Column(scale=2): with gr.Row(): sample_button_en = gr.Button("Click to get a review in English.", size='sm') # gr.Markdown('''

or

''') sample_button_nl = gr.Button("Click to get a review in Dutch.", size='sm') input_text = gr.Textbox(label="Review:", value="HELLO! Hallo!", visible=False, container=False) interpretation1 = gr.components.Interpretation(input_text) slider_1_1 = gr.Slider(label="Your rating: Dislike(0) —> Like(100)", container=True, min_width=200, height=80, show_label=True, interactive=True) user_important = gr.Textbox(label="Which words are your guesses based on?", placeholder="Enter words that you think are important for the task") with gr.Column(scale=1): gr.Markdown( ''' ## Today's Scores ''' ) with gr.Column(scale=1): gr.Image('icon_robot.png', min_width=5, show_label=False, show_share_button=False, interactive=False, container=False) with gr.Column(scale=4): tot_scores_1 = gr.Markdown( ''' ###

Machine   ''' + str(int(0)) + '''   VS   ''' + str(int(0)) + '''   Human

''' ) with gr.Column(scale=1): gr.Image('icon_user.png', min_width=5, show_label=False, show_share_button=False, interactive=False, container=False) gr.Markdown( ''' ## Like or Dislike You're given a short review of a movie, book or restaurant. The goal of this game is to guess how *positive* the review is, from 0 (=extremely bad) to 100 (=fantastic). * Step 1. Get an English or Dutch review and guess the corresponding score. * Step 2. Check the score guessed by AI. Who gets the most correct answer wins. * Step 3. Check the word highlighting to understand how AI made its decision. ''' ) with gr.Row(): with gr.Column(scale=2): chat_button_1 = gr.Button("Click to see AI's rating", size='sm') slider_1_2 = gr.Slider(label="AI rating: Dislike(0) —> Like(100)", container=True, min_width=200, height=80, show_label=True, interactive=True) interpre_button = gr.Button("See how AI got its rating", size='sm') placeholder_text = gr.Textbox(label="Review: (Red: Positive; Blue: Negative)", value="HELLO! Hallo!", visible=False) interpretation2 = gr.components.Interpretation(placeholder_text) with gr.Column(scale=1): chatbot1 = gr.Chatbot(height=230, min_width=50, container=False) # height=300 #################################################################################################### gr.Markdown(''' *** ''') gr.Markdown( ''' # Now try with your own review! ''' ) with gr.Row(): with gr.Column(scale=2): text_written = gr.Textbox(label="Review: ", placeholder="Enter your own review about a movie/restaurant/book.", visible=True) # image_1_3 = gr.Image('icon_user.png', height=80, width=80, min_width=80, show_label=False, show_share_button=False, interactive=False) slider_1_3 = gr.Slider(label="Your rating: Dislike(0) —> Like(100)", container=True, min_width=200, height=80, show_label=True, interactive=True) lang_written = gr.Radio(["English", "Dutch"], label="Language:", info="In which language is the review written?") chat_button_2 = gr.Button("Click to see AI's rating", size='sm') placeholder_written_text = gr.Textbox(label="Review: (Red: Positive; Blue: Negative)", value="HELLO! Hallo!", visible=False) interpretation4 = gr.components.Interpretation(placeholder_written_text) slider_1_4 = gr.Slider(label="AI rating: Dislike(0) —> Like(100)", container=True, min_width=200, height=80, show_label=True, interactive=True) with gr.Column(scale=1): chatbot2 = gr.Chatbot(height=350, min_width=50, container=False) # height=300 sample_button_en.click(read1, inputs=[text_en, num_selected_1], outputs=[interpretation1, lang_selected, num_selected_1]) sample_button_nl.click(read1, inputs=[text_nl, num_selected_1], outputs=[interpretation1, lang_selected, num_selected_1]) num_selected_1.change(reset_modules, outputs=[interpretation2, slider_1_1, slider_1_2, chatbot1, user_important]) chat_button_1.click(func1, inputs=[lang_selected, num_selected_1, slider_1_1, num1, num2, user_important], outputs=[slider_1_2, chatbot1, num1, num2, tot_scores_1]) interpre_button.click(interpre1, inputs=[lang_selected, num_selected_1], outputs=[interpretation2]) chat_button_2.click(func1_written, inputs=[text_written, slider_1_3, lang_written], outputs=[interpretation4, slider_1_4, chatbot2]) # with gr.Tab("Human or Machine"): # with gr.Row(): # text_input_2 = gr.Textbox() # text_output_2 = gr.Label() # text_button_2 = gr.Button("Check") with gr.Tab("Male or Female"): num_selected_3 = gr.Number(value=0, container=False, show_label=False, visible=False) with gr.Row(): with gr.Column(scale=2): with gr.Row(): # gr.Markdown('''

or

''') sample_button_en_3 = gr.Button("Click to get a sentence", size='sm') input_text_mf = gr.Textbox(label="Sentence:", value="HELLO! Hallo!", visible=False, container=False) interpretation_mf_1 = gr.components.Interpretation(input_text_mf) slider_3_1 = gr.Slider(label="Human: Male ——> Female", container=True, min_width=200, height=80, show_label=True, interactive=True) user_important_mf = gr.Textbox(label="Which words are your guesses based on?", placeholder="Enter words that you think are important for the task") with gr.Column(scale=1): gr.Markdown( ''' ## Today's Scores ''' ) tot_scores_2 = gr.Markdown( ''' ###

Machine   ''' + str(int(0)) + '''   VS   ''' + str(int(0)) + '''   Human

''' ) gr.Markdown( ''' ## Male or Female You're given a sentence spoken by a speaker. The goal of this game is to guess the gender of the speaker, from 0 (=Male) to 100 (=Female). * Step 1. Get a sentence and guess the gender of the speaker. * Step 2. Check the gender guessed by AI. Who gets the most correct answer wins. * Step 3. Check the word highlighting to understand how AI made its decision. ''' ) with gr.Row(): with gr.Column(scale=2): chat_button_mf = gr.Button("Click to see AI's answer.", size='sm') slider_3_2 = gr.Slider(label="AI: Male ——> Female", container=True, min_width=200, height=80, show_label=True, interactive=True) interpre_button_mf = gr.Button("See how AI gets the answer.", size='sm') placeholder_text_mf = gr.Textbox(label="Sentence: (Red: Female; Blue: Male)", value="HELLO! Hallo!", visible=False) interpretation_mf_2 = gr.components.Interpretation(placeholder_text_mf) with gr.Column(scale=1): chatbot_mf_1 = gr.Chatbot(height=230, min_width=50, container=False) #################################################################################################### gr.Markdown(''' *** ''') gr.Markdown( ''' # Now try your own sentence! ''' ) with gr.Row(): with gr.Column(scale=2): text_written_mf = gr.Textbox(label="Sentence: ", placeholder="Enter your sentence.", visible=True) slider_3_3 = gr.Slider(label="Human: Male ——> Female", container=True, min_width=200, height=80, show_label=True, interactive=True) chat_button_mf_2 = gr.Button("Click to see AI's answer.", size='sm') placeholder_written_text_mf = gr.Textbox(label="Sentence: (Red: Female; Blue: Male)", value="HELLO! Hallo!", visible=False) interpretation_mf_4 = gr.components.Interpretation(placeholder_written_text_mf) slider_3_4 = gr.Slider(label="AI: Male ——> Female", container=True, min_width=200, height=80, show_label=True, interactive=True) with gr.Column(scale=1): chatbot_mf_2 = gr.Chatbot(height=350, min_width=50, container=False) # height=300 sample_button_en_3.click(read3, inputs=[num_selected_3], outputs=[interpretation_mf_1, num_selected_3]) num_selected_3.change(reset_modules, outputs=[interpretation_mf_2, slider_3_1, slider_3_2, chatbot_mf_1, user_important_mf]) chat_button_mf.click(func3, inputs=[num_selected_3, slider_3_1, num1, num2, user_important_mf], outputs=[slider_3_2, chatbot_mf_1, num1, num2, tot_scores_2]) interpre_button_mf.click(interpre3, inputs=[num_selected_3], outputs=[interpretation_mf_2]) chat_button_mf_2.click(func3_written, inputs=[text_written_mf, slider_3_3], outputs=[interpretation_mf_4, slider_3_4, chatbot_mf_2]) if __name__ == "__main__": demo.launch()