import csv import sys # Increase CSV field size limit csv.field_size_limit(sys.maxsize) import gradio as gr import pandas as pd def data_pre_processing(file_responses): console_messages.append("Starting data pre-processing...") # Financial Weights can be anything (ultimately the row-wise weights are aggregated and the corresponding fractions are obtained from that rows' total tax payed) try: # Define the columns to be processed # Developing Numeric Columns # Convert columns to numeric and fill NaN values with 0 file_responses['Personal_TaxDirection_1_TaxWeightageAllocated'] = pd.to_numeric(file_responses['Personal_TaxDirection_1_TaxWeightageAllocated'], errors='coerce').fillna(0) file_responses['Personal_TaxDirection_2_TaxWeightageAllocated'] = pd.to_numeric(file_responses['Personal_TaxDirection_2_TaxWeightageAllocated'], errors='coerce').fillna(0) file_responses['Personal_TaxDirection_3_TaxWeightageAllocated'] = pd.to_numeric(file_responses['Personal_TaxDirection_3_TaxWeightageAllocated'], errors='coerce').fillna(0) file_responses['Latest estimated Tax payment?'] = pd.to_numeric(file_responses['Latest estimated Tax payment?'], errors='coerce').fillna(0) # Adding a new column 'TotalWeightageAllocated' by summing specific columns by their names file_responses['TotalWeightageAllocated'] = file_responses['Personal_TaxDirection_1_TaxWeightageAllocated'] + file_responses['Personal_TaxDirection_2_TaxWeightageAllocated'] + file_responses['Personal_TaxDirection_3_TaxWeightageAllocated'] # Creating Datasets (we assume everything has been provided to us in English, or the translations have been done already) # Renaming the datasets into similar column headings initial_dataset_1 = file_responses.rename(columns={ 'Personal_TaxDirection_1_Wish': 'Problem_Description', 'Personal_TaxDirection_1_GeographicalLocation': 'Geographical_Location', 'Personal_TaxDirection_1_TaxWeightageAllocated': 'Financial_Weight' })[['Problem_Description', 'Geographical_Location', 'Financial_Weight']] initial_dataset_2 = file_responses.rename(columns={ 'Personal_TaxDirection_2_Wish': 'Problem_Description', 'Personal_TaxDirection_2_GeographicalLocation': 'Geographical_Location', 'Personal_TaxDirection_2_TaxWeightageAllocated': 'Financial_Weight' })[['Problem_Description', 'Geographical_Location', 'Financial_Weight']] initial_dataset_3 = file_responses.rename(columns={ 'Personal_TaxDirection_3_Wish': 'Problem_Description', 'Personal_TaxDirection_3_GeographicalLocation': 'Geographical_Location', 'Personal_TaxDirection_3_TaxWeightageAllocated': 'Financial_Weight' })[['Problem_Description', 'Geographical_Location', 'Financial_Weight']] # Calculating the actual TaxAmount to be allocated against each WISH (by overwriting the newly created columns) initial_dataset_1['Financial_Weight'] = file_responses['Personal_TaxDirection_1_TaxWeightageAllocated'] * file_responses['Latest estimated Tax payment?'] / file_responses['TotalWeightageAllocated'] initial_dataset_2['Financial_Weight'] = file_responses['Personal_TaxDirection_2_TaxWeightageAllocated'] * file_responses['Latest estimated Tax payment?'] / file_responses['TotalWeightageAllocated'] initial_dataset_3['Financial_Weight'] = file_responses['Personal_TaxDirection_3_TaxWeightageAllocated'] * file_responses['Latest estimated Tax payment?'] / file_responses['TotalWeightageAllocated'] # Removing useless rows # Drop rows where Problem_Description is NaN or an empty string initial_dataset_1 = initial_dataset_1.dropna(subset=['Problem_Description'], axis=0) initial_dataset_2 = initial_dataset_2.dropna(subset=['Problem_Description'], axis=0) initial_dataset_3 = initial_dataset_3.dropna(subset=['Problem_Description'], axis=0) # Convert 'Problem_Description' column to string type initial_dataset_1['Problem_Description'] = initial_dataset_1['Problem_Description'].astype(str) initial_dataset_2['Problem_Description'] = initial_dataset_2['Problem_Description'].astype(str) initial_dataset_3['Problem_Description'] = initial_dataset_3['Problem_Description'].astype(str) # Merging the Datasets # Vertically concatenating (merging) the 3 DataFrames merged_dataset = pd.concat([initial_dataset_1, initial_dataset_2, initial_dataset_3], ignore_index=True) # Different return can be used to check the processing console_messages.append("Data pre-processing completed.") # return file_responses return merged_dataset except Exception as e: console_messages.append(f"Error during data pre-processing: {str(e)}") # return str(e), console_messages return None import spacy from transformers import AutoTokenizer, AutoModel import torch # Load SpaCy model # Install the 'en_core_web_sm' model if it isn't already installed try: nlp = spacy.load('en_core_web_sm') except OSError: # Instead of this try~catch, we could also include this < https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.5.0/en_core_web_sm-3.5.0.tar.gz > in the requirements.txt to directly load it from spacy.cli import download download('en_core_web_sm') nlp = spacy.load('en_core_web_sm') # Load Hugging Face Transformers model tokenizer = AutoTokenizer.from_pretrained("sentence-transformers/all-mpnet-base-v2") model = AutoModel.from_pretrained("sentence-transformers/all-mpnet-base-v2") # def combined_text_processing(text): # # Basic NLP processing using SpaCy # doc = nlp(text) # lemmatized_text = ' '.join([token.lemma_ for token in doc]) # # Advanced text representation using Hugging Face Transformers # inputs = tokenizer(lemmatized_text, return_tensors="pt", truncation=False, padding=True) # with torch.no_grad(): # outputs = model(**inputs) # return outputs.last_hidden_state.mean(dim=1).squeeze().numpy() import re import nltk from nltk.corpus import stopwords from nltk.tokenize import word_tokenize # Download necessary NLTK data nltk.download('punkt') nltk.download('stopwords') nltk.download('averaged_perceptron_tagger') # def combined_text_processing(text): # # Remove punctuation, numbers, URLs, and special characters # text = re.sub(r'[^\w\s]', '', text) # Remove punctuation and special characters # text = re.sub(r'\d+', '', text) # Remove numbers # text = re.sub(r'http\S+', '', text) # Remove URLs # # Tokenize and remove stopwords # tokens = word_tokenize(text.lower()) # Convert to lowercase # stop_words = set(stopwords.words('english')) # tokens = [word for word in tokens if word not in stop_words] # # Lemmatize tokens using SpaCy # doc = nlp(' '.join(tokens)) # lemmatized_text = ' '.join([token.lemma_ for token in doc]) # # Apply Hugging Face Transformers # inputs = tokenizer(lemmatized_text, return_tensors="pt", truncation=False, padding=True) # with torch.no_grad(): # outputs = model(**inputs) # return outputs.last_hidden_state.mean(dim=1).squeeze().numpy() import numpy as np import sentencepiece as sp from transformers import pipeline # Load a summarization model summarizer = pipeline("summarization") def Summarized_text(passed_text): try: # Summarization summarize_text = summarizer(passed_text, max_length=70, min_length=30, do_sample=False)[0]['summary_text'] return summarize_text except Exception as e: print(f"Summarization failed: {e}") return passed_text ###### Will uncomment Summarization during final deployment... as it takes a lot of time def Lemmatize_text(text): # Text Cleaning text = re.sub(r'[^\w\s]', '', text) text = re.sub(r'\d+', '', text) text = re.sub(r'http\S+', '', text) # Remove https URLs text = re.sub(r'www\.\S+', '', text) # Remove www URLs # Tokenize and remove stopwords tokens = word_tokenize(text.lower()) stop_words = set(stopwords.words('english')) custom_stopwords = {'example', 'another'} # Add custom stopwords tokens = [word for word in tokens if word not in stop_words and word not in custom_stopwords] # NER - Remove named entities doc = nlp(' '.join(tokens)) tokens = [token.text for token in doc if not token.ent_type_] # POS Tagging (optional) pos_tags = nltk.pos_tag(tokens) tokens = [word for word, pos in pos_tags if pos in ['NN', 'NNS']] # Filter nouns # Lemmatize tokens using SpaCy doc = nlp(' '.join(tokens)) lemmatized_text = ' '.join([token.lemma_ for token in doc]) return lemmatized_text # Return the cleaned and lemmatized text from random import random def text_processing_for_domain(text): # First, get the summarized text summarized_text = "" # summarized_text = Summarized_text(text) # Then, lemmatize the original text lemmatized_text = "" lemmatized_text = Lemmatize_text(text) if lemmatized_text and summarized_text: # Join both the summarized and lemmatized text if random() > 0.5: combined_text = summarized_text + " " + lemmatized_text else: combined_text = lemmatized_text + " " + summarized_text return combined_text elif summarized_text: return summarized_text elif lemmatized_text: return lemmatized_text else: return "Sustainability and Longevity" # Default FailSafe # # 2. Clustering from ChatGPT # # Libraries: scikit-learn, sentence-transformers # # Use sentence embeddings and clustering algorithms to group similar project proposals. # from bertopic import BERTopic # def perform_clustering(texts, n_clusters): # topic_model = BERTopic(n_topics=n_clusters) # topics, _ = topic_model.fit_transform(texts) # return topics, topic_model # # Clustering function call # clustered_df, cluster_centers = clustering(processed_df) # Method 1: Sentence Transformers + KMeans # # 2. Clustering: from Claude # # Use BERTopic for advanced topic modeling and clustering. # from bertopic import BERTopic # def perform_clustering(texts, n_clusters): # topic_model = BERTopic(n_topics=n_clusters) # topics, _ = topic_model.fit_transform(texts) # return topics, topic_model # # Clustering function call # problem_clusters, problem_model = perform_clustering(processed_df['Problem_Description'], n_clusters=10) # location_clusters, location_model = perform_clustering(processed_df['Geographical_Location'], n_clusters=5) # After this Method 2: BERTopic function, the following need to be done: # processed_df['Problem_Cluster'] = problem_clusters # 2. Meta AI Function: Sentence Transformers + Hierarchical Clustering + Silhouette Analysis # Now this also includes: # Topic Modeling using BERTopic: Integrated BERTopic to extract representative words for each cluster. # Cluster Visualization: Added a simple visualization to display the top words in each cluster. # Hyperparameter Tuning: Include a parameter to adjust the number of top words to display for each cluster. from sentence_transformers import SentenceTransformer from sklearn.cluster import AgglomerativeClustering, KMeans from sklearn.feature_extraction.text import TfidfVectorizer from sklearn.metrics import silhouette_score from bertopic import BERTopic from collections import Counter def extract_problem_domains(df, text_column='Processed_ProblemDescription_forDomainExtraction', # text_column='Problem_Description', cluster_range=(5, 15), top_words=10, # method='sentence_transformers' method='tfidf_kmeans' ): console_messages.append("Extracting Problem Domains...") if method == 'sentence_transformers': # Sentence Transformers approach model = SentenceTransformer('all-mpnet-base-v2') embeddings = model.encode(df[text_column].tolist()) # Perform hierarchical clustering with Silhouette Analysis silhouette_scores = [] for n_clusters in range(cluster_range[0], cluster_range[1] + 1): clustering = AgglomerativeClustering(n_clusters=n_clusters) cluster_labels = clustering.fit_predict(embeddings) silhouette_avg = silhouette_score(embeddings, cluster_labels) silhouette_scores.append(silhouette_avg) # Determine the optimal number of clusters optimal_n_clusters = cluster_range[0] + silhouette_scores.index(max(silhouette_scores)) # Perform clustering with the optimal number of clusters clustering = AgglomerativeClustering(n_clusters=optimal_n_clusters) cluster_labels = clustering.fit_predict(embeddings) elif method == 'tfidf_kmeans': # TF-IDF Vectorization and K-Means approach vectorizer = TfidfVectorizer(stop_words='english', max_features=3000) X = vectorizer.fit_transform(df[text_column]) # Perform K-Means clustering with Silhouette Analysis silhouette_scores = [] for n_clusters in range(cluster_range[0], cluster_range[1] + 1): kmeans = KMeans(n_clusters=n_clusters)#, random_state=42) cluster_labels = kmeans.fit_predict(X) silhouette_avg = silhouette_score(X, cluster_labels) silhouette_scores.append(silhouette_avg) # Determine the optimal number of clusters optimal_n_clusters = cluster_range[0] + silhouette_scores.index(max(silhouette_scores)) # Perform final clustering with optimal number of clusters kmeans = KMeans(n_clusters=optimal_n_clusters) #, random_state=42) cluster_labels = kmeans.fit_predict(X) # # BERTopic approach (commented out) console_messages.append("BERT is currently commented...") # topic_model = BERTopic() # topics, _ = topic_model.fit_transform(df[text_column].tolist()) # topic_model.reduce_topics(df[text_column].tolist(), nr_topics=optimal_n_clusters) # cluster_labels = topics # Get representative words for each cluster if method == 'sentence_transformers': cluster_representations = {} for i in range(optimal_n_clusters): cluster_words = df.loc[cluster_labels == i, text_column].str.cat(sep=' ').split() cluster_representations[i] = [word for word, _ in Counter(cluster_words).most_common(top_words)] elif method == 'tfidf_kmeans': feature_names = vectorizer.get_feature_names_out() cluster_representations = {} for i in range(optimal_n_clusters): # center = kmeans.cluster_centers_[i] # # print(f"top_words: {top_words}, type: {type(top_words)}") # # print(f"center.argsort(): {center.argsort()}, type: {type(center.argsort())}") # console_messages.append(f"top_words: {top_words}, type: {type(top_words)}") # console_messages.append(f"center.argsort(): {center.argsort()}, type: {type(center.argsort())}") # # top_word_indices = center.argsort()[-top_words:][::-1] # top_word_indices = center.argsort()[-top_words:][::-1].tolist() # Indexes of top words # top_words = [feature_names[index] for index in top_word_indices] # cluster_representations[i] = top_words try: center = kmeans.cluster_centers_[i] console_messages.append(f"Processing cluster {i}") console_messages.append(f"Center shape: {center.shape}, type: {type(center)}") if isinstance(center, np.ndarray) and center.ndim == 1: top_word_indices = center.argsort()[-top_words:][::-1] # top_word_indices = center.argsort()[-top_words:][::-1].tolist() console_messages.append(f"Top word indices for cluster {i}: {top_word_indices}") top_words = [feature_names[index] for index in top_word_indices] console_messages.append(f"Top words: {top_words}") cluster_representations[i] = top_words else: console_messages.append(f"Error: Cluster center is not a 1D array for cluster {i}") except Exception as e: console_messages.append(f"Error processing cluster {i}: {str(e)}") console_messages.append(f"Center: {center}") console_messages.append(f"Number of clusters: {optimal_n_clusters}") console_messages.append(f"Sample cluster words: {cluster_representations[0][:5]}...") # Map cluster labels to representative words df["Problem_Cluster"] = cluster_labels df['Problem_Category_Words'] = [cluster_representations[label] for label in cluster_labels] # console_messages.append("Returning from Problem Domain Extraction function.") console_messages.append("Problem Domain Extraction completed.") return df, optimal_n_clusters # Usage # clustered_df, optimal_n_clusters = optimal_Problem_clustering(processed_df) # print(f'Optimal number of clusters: {optimal_n_clusters}') # def nlp_pipeline(original_df): def nlp_pipeline(original_df): console_messages.append("Starting NLP pipeline...") # Data Preprocessing processed_df = data_pre_processing(original_df) # merged_dataset # Starting the Pipeline for Domain Extraction console_messages.append("Executing Text processing function for Domain identification") # Apply the text_processing_for_domain function to the DataFrame processed_df['Processed_ProblemDescription_forDomainExtraction'] = processed_df['Problem_Description'].apply(text_processing_for_domain) console_messages.append("Removing entries which could not be allocated to any Problem Domain") # processed_df = processed_df.dropna(subset=['Processed_ProblemDescription_forDomainExtraction'], axis=0) # Drop rows where 'Processed_ProblemDescription_forDomainExtraction' contains empty arrays processed_df = processed_df[processed_df['Processed_ProblemDescription_forDomainExtraction'].apply(lambda x: len(x) > 0)] # Domain Clustering try: domain_df, optimal_n_clusters = extract_problem_domains(processed_df) # print(f"Optimal clusters: {optimal_clusters}") # print(result_df.head()) # console_messages.append(f"Optimal clusters: {optimal_n_clusters}") console_messages.append("NLP pipeline completed.") return domain_df except Exception as e: # print(f"Error in extract_problem_domains: {e}") console_messages.append(f"Error in extract_problem_domains: {str(e)}") return processed_df # return domain_df # problem_clusters, problem_model = perform_clustering(processed_df['Problem_Description'], n_clusters=10) # location_clusters, location_model = perform_clustering(processed_df['Geographical_Location'], n_clusters=5) console_messages = [] def process_excel(file): console_messages.append("Processing starts. Reading the uploaded Excel file...") # Ensure the file path is correct file_path = file.name if hasattr(file, 'name') else file # Read the Excel file df = pd.read_excel(file_path) try: # Process the DataFrame console_messages.append("Processing the DataFrame...") result_df = nlp_pipeline(df) # output_file = "Output_ProjectProposals.xlsx" output_file = "Output_Proposals.xlsx" result_df.to_excel(output_file, index=False) console_messages.append("Processing completed. Ready for download.") return output_file, "\n".join(console_messages) # Return the processed DataFrame as Excel file except Exception as e: # return str(e) # Return the error message # error_message = f"Error processing file: {str(e)}" # print(error_message) # Log the error console_messages.append(f"Error during processing: {str(e)}") # return error_message, "Santanu Banerjee" # Return the error message to the user return None, "\n".join(console_messages) # example_files = ['#TaxDirection (Responses)_BasicExample.xlsx', # '#TaxDirection (Responses)_IntermediateExample.xlsx', # '#TaxDirection (Responses)_UltimateExample.xlsx' # ] example_files = ['a.xlsx',] import random a_random_object = random.choice(["⇒", "↣", "↠", "→"]) # Define the Gradio interface interface = gr.Interface( fn=process_excel, # The function to process the uploaded file inputs=gr.File(type="filepath", label="Upload Excel File here. \t Be sure to check that the column headings in your upload are the same as in the Example files below. \t (Otherwise there will be Error during the processing)"), # File upload input examples=example_files, # Add the example files outputs=[ gr.File(label="Download the processed Excel File containing the ** Project Proposals ** for each Location~Problem paired combination"), # File download output gr.Textbox(label="Console Messages", lines=15, interactive=False) # Console messages output ], # title="Excel File Uploader", # title="Upload Excel file containing #TaxDirections → Download HyperLocal Project Proposals\n", title = ( "

" "Upload Excel file containing #TaxDirections " # "→ " # "⇒ ↣ ↠ " " " +a_random_object +" " "Download HyperLocal Project Proposals" "

\n" ), description=( "

This tool allows for the systematic evaluation and proposal of solutions tailored to specific location-problem pairs, ensuring efficient resource allocation and project planning. For more information, visit #TaxDirection weblink.

" "

Upload an Excel file to process and download the result or use the Example files:

" "

(click on any of them to directly process the file and Download the result)

" "

Processed output contains a Project Proposal for each Location~Problem paired combination (i.e. each cell).

" "

Corresponding Budget Allocation and estimated Project Completion Time are provided in different sheets.

" "

Note: The example files provided above are for demonstration purposes. Feel free to upload your own Excel files to see the results. If you have any questions, refer to the documentation-links or contact support.

" ) # Solid description with right-aligned second sentence ) # Launch the interface if __name__ == "__main__": interface.launch()