import pandas as pd import os import json from sklearn.preprocessing import StandardScaler import argparse # this function put on the good format the date from the nlp dataset def to_date(my_date): set_month= {'Jan': '01', 'Feb': '02', 'Mar': '03', 'Apr': '04', 'May': '05', 'Jun': '06', 'Jul': '07', 'Aug': '08', 'Sep': '09', 'Oct': '10', 'Nov': '11', 'Dec': '12'} month = set_month[my_date[4:7]] day = my_date[8:10] year = my_date[26:30] final_date = year+'-'+month+'-'+day return final_date # open the knowledge base of the corresponding state with the corresponding category def open_state_file(dataset_path: str, STATE: str, CATEGORY: str) -> list[str]: tags = set() with open('./utils/Keywords/'+STATE+'_keywords_'+CATEGORY+'_no_dupe.txt', 'r') as fp: for line in fp: # remove linebreak from a current name # linebreak is the last character of each line x = line[:-1] # add current item to the list tags.add(x.lower()) return tags # the function take a time series dataframe (df) and return another dataset with all the data put in one parameter # (Window) except the date and the label which are in different columns # window size is equal to the size of the window def window_creation(df: object, window_size: int, timestamp_colum: str) -> object: """ Create a TS window for each date. Distance of sliding window is 1. df: Dataframe of a whole TS record in one state. Each row represents the record on one date. window_size: Size of TS window. label_column: Name of the column of label timestamp_colum: Name of the column of time stamp """ df = df.reset_index() df = df.drop(columns=['index']) good_columns = list(df.columns) good_columns.remove(timestamp_colum) window_df = df[good_columns] scaler = StandardScaler() window_df = pd.DataFrame(scaler.fit_transform(window_df), columns=window_df.columns) date_df = df[timestamp_colum] window_data = [] label_data = [] date_data = [] for i in range(len(df)-window_size): my_window = window_df.loc[i:i+window_size-1] window_data.append(my_window.to_numpy()) date_data.append((date_df.loc[i+window_size-1])) return_value = pd.DataFrame({'Date': date_data, 'Window': window_data}) return return_value # the function that link the NLP data and the time series data for English dataset # crisis_to_link is one row from crisis knowledge # directory is the path of NLP data # dataset_path is the path of time series data # window_size is the size of the window for time series data # fillNA is the value to fill the missing value def linker( crisis_to_link: object, corpus_path: str, # path of corpus directory ts_dataset_path: str, # path of ts data directory window_size: int, date_column: str, fillNA : int) -> object: # list of states associated with the crisis type_of_crisis = crisis_to_link['Crisis'] states = crisis_to_link['Places'].replace('[', '').replace(']', '').split(',') # features to keep features = ['date', 'mean_speed_wind', 'precipitation', 'snow_fall', 'snow_depth', 'temp_max', 'temp_min', 'temp_mean', 'wind_dir_angle_2min', 'wind_dir_angle_5sec', 'wind_max_speed_2min', 'wind_max_speed_5sec'] df_return = pd.DataFrame({'Date': [], 'Text': [],'Crisis_Type' :[], 'Window': [], 'label_humanitarian': [], 'label_useful' : [], 'label_urgent' : [], 'label_sudden' : []}) for state in states: # Open the time series file corresponding to the state ts_df = pd.read_csv(ts_dataset_path + '/' + state + '_Time_Series_Data.csv') ts_df = ts_df[features].sort_values(by='date') # Frequency of TS is per day. ts_df = ts_df.fillna(fillNA) # Window Creation window_df = window_creation(ts_df, window_size, date_column) #since New Zealand has no keyword, we skip this part if state != 'New_Zealand' : # we get the keywords to link the location mention to the state tags_city = open_state_file(ts_dataset_path, state, 'city') tags_county = open_state_file(ts_dataset_path, state, 'county') tags_state = open_state_file(ts_dataset_path, state, 'state') date = [] text = [] label_humanitarian = [] label_useful = [] label_urgent = [] label_sudden = [] window_data = [] crisis_label = [] # Process NLP data and link to time series data for root, dirs, files in os.walk(corpus_path + '/' + crisis_to_link['Path_name']): for fil in files: if fil.endswith('.jsonl'): with open(os.path.join(root,fil), 'r') as json_file: json_list = list(json_file) # a list of '{}', each {} is a tweet and its features. for json_str in json_list: # for each tweet loaded result = json.loads(json_str) place = result['location_mentions'] # if the location mention is empty (the tweet does not refer to particular place), # thanks to the crisis_to_link, we know which crisis this tweet make reference # (the tweet speak about this crisis) so we assume that if location mention is empty # we assume that the tweet make a reference to the current state since this state is the localisation of the crisis #we still have the New Zealand special case if place == [] or state == 'New_Zealand': # Put NLP date on the same format as time series date date_NLP = to_date(result['created_at']) # Check if there is matching date between time series and tweets. if list(window_df['Window'][ window_df['Date'] == date_NLP]) != [] : date.append(date_NLP) text.append(result['text']) linked_data = window_df[window_df['Date'] == date[-1]] my_window = list(linked_data['Window'])[0] window_data.append(my_window) # for the label, we take reference from the time series label label_humanitarian.append(result['humAID_class']) crisis_label.append(type_of_crisis) if result['humAID_class'] == 'not_humanitarian' : label_useful.append('not_humanitarian') label_urgent.append('not_humanitarian') else : label_useful.append('useful') if result['humAID_class'] == 'rescue_volunteering_or_donation_effort' : label_urgent.append('not_urgent') else : label_urgent.append('urgent') if result['humAID_class'] == 'not_humanitarian' : label_sudden.append('Not_Crisis_period') else : if type_of_crisis == 'Earthquake' or type_of_crisis == 'WildFire' : label_sudden.append('Sudden_Crisis') else : label_sudden.append('Non_Sudden_Crisis') else: for zone in place: # if the location mention refer to a state and this reference is in our knowledge database bool1 = zone['type'] == 'State' and zone['text'].lower() in tags_state # if the location mention refer to a county and this reference is in our knowledge database bool2 = zone['type'] == 'County' and zone['text'].lower() in tags_county # if the location mention refer to a city and this reference is in our knowledge database bool3 = zone['type'] == 'City/town' and zone['text'].lower() in tags_city if bool1 or bool2 or bool3: date.append(to_date(result['created_at'])) text.append(result['text']) linked_data = window_df[window_df['Date'] == date[-1]] my_window = list(linked_data['Window'])[0] window_data.append(my_window) label_humanitarian.append(result['humAID_class']) crisis_label.append(type_of_crisis) if result['humAID_class'] == 'not_humanitarian' : label_useful.append('not_humanitarian') label_urgent.append('not_humanitarian') else : label_useful.append('useful') if result['humAID_class'] == 'rescue_volunteering_or_donation_effort' : label_urgent.append('not_urgent') else : label_urgent.append('urgent') if result['humAID_class'] == 'not_humanitarian' : label_sudden.append('Not_Crisis_period') else : if type_of_crisis == 'Earthquake' or type_of_crisis == 'WildFire' : label_sudden.append('Sudden_Crisis') else : label_sudden.append('Non_Sudden_Crisis') df = pd.DataFrame({'Date': date, 'Text': text,'Crisis_Type' :crisis_label, 'Window': window_data, 'label_humanitarian': label_humanitarian, 'label_useful' : label_useful, 'label_urgent' : label_urgent, 'label_sudden' : label_sudden}) df_return = pd.concat([df_return, df]) return df_return # loading file parser = argparse.ArgumentParser() parser.add_argument("-w", "--window_size",type = int, default = 5, help="size (in number of days) of our daily time series data") parser.add_argument("-o", "--output_file",type = str, default = './Multi_modal_dataset/CrisisTS_Eng_Personnalized.csv', help="name of your output file") args = parser.parse_args() directory_nlp = './Textual_Data/English_Corpus' directory_time_series = './Time_Series/EnglishTS' path_knowledge = './utils/crisis_knowledge_eng.csv' knowledge = pd.read_csv(path_knowledge, sep='\t') # (9, 3), 3 columns represent Crisis, Places, Path_name (tweets) # for all test crisis in english All_possible_crisis = ['Hurricane', 'WildFire', 'Flood','Earthquake'] Test_crisis = 'Hurricane' Train_knowledge = knowledge[knowledge['Crisis'] != Test_crisis] Test_knowledge = knowledge[knowledge['Crisis'] == Test_crisis] # link multi_modal_test_data = pd.DataFrame({'Date': [], 'Text': [],'Crisis_Type' :[], 'Window': [], 'label_humanitarian': [], 'label_useful' : [], 'label_urgent' : [], 'label_sudden' : []}) for index, crisis in knowledge.iterrows(): print('Linking textual data and meteorological data of '+crisis['Path_name']+ ' ...') multi_modal_test_data = pd.concat([multi_modal_test_data, linker(crisis, directory_nlp, directory_time_series, args.window_size, 'date', 0)]) print('Done !') multi_modal_test_data.to_csv(args.output_file,sep = '\t')