import requests import random import time import pandas as pd import gradio as gr import numpy as np import torch from transformers import AutoTokenizer, AutoModelForSequenceClassification from transformers import pipeline import shap def read1(lang, num_selected_former): if lang in ['en']: fname = 'data1_en.txt' else: fname = 'data1_nl_10.txt' with open(fname, encoding='utf-8') as f: content = f.readlines() index_selected = random.randint(0,len(content)/2-1) while index_selected == num_selected_former: index_selected = random.randint(0,len(content)/2-1) text = eval(content[index_selected*2]) interpretation = eval(content[int(index_selected*2+1)]) if lang == 'en': min_len = 4 else: min_len = 2 while len(text['text'].split(' ')) <= min_len or '\\' in text['text'] or '//' in text['text']: index_selected = random.randint(0,len(content)/2-1) text = eval(content[int(index_selected*2)]) res_tmp = [(i, 0) for i in text['text'].split(' ')] res = {"original": text['text'], "interpretation": res_tmp} return res, lang, index_selected def read1_written(lang): if lang in ['en']: fname = 'data1_en.txt' else: fname = 'data1_nl_10.txt' with open(fname, encoding='utf-8') as f: content = f.readlines() index_selected = random.randint(0,len(content)/2-1) text = eval(content[index_selected*2]) if lang == 'en': min_len = 4 else: min_len = 2 while (len(text['text'].split(' '))) <= min_len or '\\' in text['text'] or '//' in text['text']: # while (len(text['text'].split(' '))) <= min_len: index_selected = random.randint(0,len(content)/2-1) text = eval(content[int(index_selected*2)]) # interpretation = [(i, 0) for i in text['text'].split(' ')] # res = {"original": text['text'], "interpretation": interpretation} # print(res) return text['text'] def func1(lang_selected, num_selected, human_predict, num1, num2, user_important): chatbot = [] # num1: Human score; num2: AI score if lang_selected in ['en']: fname = 'data1_en.txt' else: fname = 'data1_nl_10.txt' with open(fname) as f: content = f.readlines() text = eval(content[int(num_selected*2)]) interpretation = eval(content[int(num_selected*2+1)]) if lang_selected in ['en']: golden_label = text['label'] * 2.5 else: golden_label = text['label'] * 10 ''' # (START) API version -- quick API_URL = "https://api-inference.huggingface.co/models/nlptown/bert-base-multilingual-uncased-sentiment" # API_URL = "https://api-inference.huggingface.co/models/cmarkea/distilcamembert-base-sentiment" headers = {"Authorization": "Bearer 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"} response = requests.post(API_URL, headers=headers, json=text['text']) output = response.json() # result = dict() star2num = { "5 stars": 100, "4 stars": 75, "3 stars": 50, "2 stars": 25, "1 star": 0, } print(output) out = output[0][0] # (END) API version ''' # (START) off-the-shelf version -- slow at the beginning # Load model directly tokenizer = AutoTokenizer.from_pretrained("nlptown/bert-base-multilingual-uncased-sentiment") model = AutoModelForSequenceClassification.from_pretrained("nlptown/bert-base-multilingual-uncased-sentiment") # Use a pipeline as a high-level helper device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") print(device) classifier = pipeline("sentiment-analysis", model=model, tokenizer=tokenizer, device=device) output = classifier([text['text']]) star2num = { "5 stars": 10, "4 stars": 7.5, "3 stars": 5, "2 stars": 2.5, "1 star": 0, } print(output) out = output[0] # (END) off-the-shelf version ai_predict = star2num[out['label']] # result[label] = out['score'] user_select = "You focused on " flag_select = False if user_important == "": user_select += "nothing. Interesting! " else: user_select += "'" + user_important + "'. " # for i in range(len(user_marks)): # if user_marks[i][1] != None and h1[i][0] not in ["P", "N"]: # flag_select = True # user_select += "'" + h1[i][0] + "'" # if i == len(h1) - 1: # user_select += ". " # else: # user_select += ", " # if not flag_select: # user_select += "nothing. Interesting! " user_select += "Wanna see how the AI made the guess? Click here. ⬅️" if lang_selected in ['en']: # 0 1 2 3 4 5 6 7 8 9 10 if ai_predict == golden_label: if abs(human_predict - golden_label) <= 2: # Both correct golden_label = int((human_predict + ai_predict) / 2) ai_predict = golden_label chatbot.append(("The correct answer is " + str(golden_label) + ". Congratulations! πŸŽ‰ Both of you get the correct answer!", user_select)) num1 += 1 num2 += 1 else: golden_label += random.randint(-1, 1) * 0.5 while golden_label > 10 or golden_label < 0: golden_label += random.randint(-1, 1) * 0.5 golden_label = int(golden_label) ai_predict = golden_label chatbot.append(("The correct answer is " + str(golden_label) + ". Sorry.. AI wins in this round.", user_select)) num2 += 1 else: if abs(human_predict - golden_label) < abs(ai_predict - golden_label): if abs(human_predict - golden_label) < 2: golden_label = int((golden_label + human_predict) / 2) ai_predict += random.randint(-1, 1) * 0.5 ai_predict = int(ai_predict) chatbot.append(("The correct answer is " + str(golden_label) + ". Great! πŸŽ‰ You are closer to the answer and better than AI!", user_select)) num1 += 1 else: golden_label = int(golden_label) ai_predict = int(ai_predict) chatbot.append(("The correct answer is " + str(golden_label) + ". Both wrong... Maybe next time you'll win!", user_select)) else: golden_label = int(golden_label) ai_predict = int(ai_predict) chatbot.append(("The correct answer is " + str(golden_label) + ". Sorry.. No one gets the correct answer. But nice try! πŸ˜‰", user_select)) else: if golden_label == 10: if ai_predict > 5 and human_predict > 5: golden_label = int((human_predict + ai_predict)/2) + random.randint(-1, 1) while golden_label > 10: golden_label = int((human_predict + ai_predict)/2) + random.randint(-1, 1) ai_predict = int((golden_label + ai_predict) / 2) chatbot.append(("The correct answer is " + str(golden_label) + ". Congratulations! πŸŽ‰ Both of you get the correct answer!", user_select)) num1 += 1 num2 += 1 elif ai_predict > 5 and human_predict <= 5: golden_label -= random.randint(0, 3) ai_predict = 7 + random.randint(-1, 2) chatbot.append(("The correct answer is " + str(golden_label) + ". Sorry.. AI wins in this round.", user_select)) num2 += 1 elif ai_predict <= 5 and human_predict > 5: golden_label = human_predict + random.randint(-1, 1) while golden_label > 10: golden_label = human_predict + random.randint(-1, 1) ai_predict = int(ai_predict) golden_label = int(golden_label) chatbot.append(("The correct answer is " + str(golden_label) + ". Great! πŸŽ‰ You are close to the answer and better than AI!", user_select)) num1 += 1 else: golden_label = int(golden_label) ai_predict = int(ai_predict) chatbot.append(("The correct answer is " + str(golden_label) + ". Sorry... No one gets the correct answer. But nice try! πŸ˜‰", user_select)) else: if ai_predict < 5 and human_predict < 5: golden_label = int((human_predict + ai_predict)/2) + random.randint(-1, 1) while golden_label < 0: golden_label = int((human_predict + ai_predict)/2) + random.randint(-1, 1) ai_predict = int((golden_label + ai_predict) / 2) chatbot.append(("The correct answer is " + str(golden_label) + ". Congratulations! πŸŽ‰ Both of you get the correct answer!", user_select)) num1 += 1 num2 += 1 elif ai_predict < 5 and human_predict >= 5: golden_label += random.randint(0, 3) ai_predict = 3 + random.randint(-2, 1) chatbot.append(("The correct answer is " + str(golden_label) + ". Sorry.. AI wins in this round.", user_select)) num2 += 1 elif ai_predict >= 5 and human_predict < 5: golden_label = human_predict + random.randint(-1, 1) while golden_label < 0: golden_label = human_predict + random.randint(-1, 1) ai_predict = int(ai_predict) chatbot.append(("The correct answer is " + str(golden_label) + ". Great! πŸŽ‰ You are close to the answer and better than AI!", user_select)) num1 += 1 else: golden_label = int(golden_label) ai_predict = int(ai_predict) chatbot.append(("The correct answer is " + str(golden_label) + ". Sorry... No one gets the correct answer. But nice try! πŸ˜‰", user_select)) # data = pd.DataFrame( # { # "Role": ["AI πŸ€–", "HUMAN πŸ‘¨πŸ‘©"], # "Scores": [num2, num1], # } # ) # scroe_human = ''' # Human: ''' + str(int(num1)) # scroe_robot = ''' # Robot: ''' + str(int(num2)) # tot_scores = ''' ###

πŸ€– Machine   ''' + str(int(num2)) + '''   VS   ''' + str(int(num1)) + '''   Human πŸ‘¨πŸ‘©

''' # tot_scores = ''' ####

Today's Scores:

# ####

πŸ€– Machine   ''' + str(int(num2)) + '''   VS   ''' + str(int(num1)) + '''   Human πŸ™‹

''' tot_scores = ''' ####

Today's Scores:     πŸ€– Machine   ''' + str(int(num2)) + '''   VS   ''' + str(int(num1)) + '''   Human πŸ™‹

''' # num_tmp = max(num1, num2) # y_lim_upper = (int((num_tmp + 3)/10)+1) * 10 # figure = gr.BarPlot.update( # data, # x="Role", # y="Scores", # color="Role", # vertical=False, # y_lim=[0,y_lim_upper], # color_legend_position='none', # height=250, # width=500, # show_label=False, # container=False, # ) # tooltip=["Role", "Scores"], return ai_predict, chatbot, num1, num2, tot_scores def interpre1(lang_selected, num_selected): if lang_selected in ['en']: fname = 'data1_en.txt' tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased-finetuned-sst-2-english") else: fname = 'data1_nl_10.txt' tokenizer = AutoTokenizer.from_pretrained("DTAI-KULeuven/robbert-v2-dutch-sentiment") with open(fname) as f: content = f.readlines() text = eval(content[int(num_selected*2)]) interpretation = eval(content[int(num_selected*2+1)]) encodings = tokenizer(text['text'], return_offsets_mapping=True) print(encodings['offset_mapping']) is_subword = [False, False] for i in range(2, len(encodings['offset_mapping'])): if encodings['offset_mapping'][i][0] == encodings['offset_mapping'][i-1][1]: is_subword.append(True) else: is_subword.append(False) print(is_subword) interpretation_combined = [] index_tmp = 0 while index_tmp < (len(interpretation) - 1): if not is_subword[index_tmp+1]: interpretation_combined.append((interpretation[index_tmp][0], -interpretation[index_tmp][1])) index_tmp += 1 else: text_combined = interpretation[index_tmp][0] score_combinded = interpretation[index_tmp][1] length = 1 while is_subword[index_tmp+length]: text_combined += interpretation[index_tmp+length][0] score_combinded += interpretation[index_tmp+length][1] length += 1 interpretation_combined.append((text_combined, -score_combinded/length)) index_tmp += length interpretation_combined.append(('', 0.0)) print(interpretation_combined) res = {"original": text['text'], "interpretation": interpretation_combined} # pos = [] # neg = [] # res = [] # for i in interpretation: # if i[1] > 0: # pos.append(i[1]) # elif i[1] < 0: # neg.append(i[1]) # else: # continue # median_pos = np.median(pos) # median_neg = np.median(neg) # res.append(("P", "+")) # res.append(("/", None)) # res.append(("N", "-")) # res.append(("Review:", None)) # for i in interpretation: # if i[1] > median_pos: # res.append((i[0], "+")) # elif i[1] < median_neg: # res.append((i[0], "-")) # else: # res.append((i[0], None)) return res def func1_written(text_written, human_predict, lang_written): chatbot = [] # num1: Human score; num2: AI score ''' # (START) API version API_URL = "https://api-inference.huggingface.co/models/nlptown/bert-base-multilingual-uncased-sentiment" # API_URL = "https://api-inference.huggingface.co/models/cmarkea/distilcamembert-base-sentiment" headers = {"Authorization": "Bearer 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"} response = requests.post(API_URL, headers=headers, json=text_written) output = response.json() # result = dict() star2num = { "5 stars": 100, "4 stars": 75, "3 stars": 50, "2 stars": 25, "1 star": 0, } out = output[0][0] # (END) API version ''' # (START) off-the-shelf version # tokenizer = AutoTokenizer.from_pretrained("nlptown/bert-base-multilingual-uncased-sentiment") # model = AutoModelForSequenceClassification.from_pretrained("nlptown/bert-base-multilingual-uncased-sentiment") device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") classifier = pipeline("sentiment-analysis", model="nlptown/bert-base-multilingual-uncased-sentiment", device=device) output = classifier([text_written]) star2num = { "5 stars": 10, "4 stars": 7.5, "3 stars": 5, "2 stars": 2.5, "1 star": 0, } print(output) out = output[0] # (END) off-the-shelf version ai_predict = star2num[out['label']] # result[label] = out['score'] if abs(ai_predict - human_predict) <= 2: ai_predict = int(ai_predict) chatbot.append(("AI gives it a close score! πŸŽ‰", "⬅️ Feel free to try another one! This time let’s see if you can trick the AI into giving a wrong rating. ⬅️")) else: ai_predict += int(random.randint(-1, 1)) while ai_predict > 10 or ai_predict < 0: ai_predict += int(random.randint(-1, 1)) chatbot.append(("AI thinks in a different way from human. πŸ˜‰", "⬅️ Feel free to try another one! ⬅️")) # sentiment_classifier = pipeline("text-classification", return_all_scores=True) device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") if lang_written == "Dutch": sentiment_classifier = pipeline("text-classification", model='DTAI-KULeuven/robbert-v2-dutch-sentiment', return_all_scores=True, device=device) tokenizer = AutoTokenizer.from_pretrained("DTAI-KULeuven/robbert-v2-dutch-sentiment") else: sentiment_classifier = pipeline("text-classification", model='distilbert-base-uncased-finetuned-sst-2-english', return_all_scores=True, device=device) tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased-finetuned-sst-2-english") explainer = shap.Explainer(sentiment_classifier) shap_values = explainer([text_written]) interpretation = list(zip(shap_values.data[0], shap_values.values[0, :, 1])) encodings = tokenizer(text_written, return_offsets_mapping=True) print(encodings['offset_mapping']) is_subword = [False, False] for i in range(2, len(encodings['offset_mapping'])): if encodings['offset_mapping'][i][0] == encodings['offset_mapping'][i-1][1]: is_subword.append(True) else: is_subword.append(False) print(is_subword) interpretation_combined = [] index_tmp = 0 while index_tmp < (len(interpretation) - 1): if not is_subword[index_tmp+1]: interpretation_combined.append((interpretation[index_tmp][0], -interpretation[index_tmp][1])) index_tmp += 1 else: text_combined = interpretation[index_tmp][0] score_combinded = interpretation[index_tmp][1] length = 1 while is_subword[index_tmp+length]: text_combined += interpretation[index_tmp+length][0] score_combinded += interpretation[index_tmp+length][1] length += 1 interpretation_combined.append((text_combined, -score_combinded/length)) index_tmp += length interpretation_combined.append(('', 0.0)) print(interpretation_combined) res = {"original": text_written, "interpretation": interpretation_combined} print(res) return res, ai_predict, chatbot