app.py

import tensorflow as tf
import gradio as gr
from webScraping import parseURL
import spacy


'''
Preprocessing Abstract data from given url 
'''
def preprocess_abstract_data(url):
    abstract_paragraph = parseURL(url)
    nlp =spacy.load('en_core_web_lg')
    nlp.add_pipe("sentencizer")  #will add sentencizer to NLP pipeline,responsible for splitting a paragraph into individual sentence
    doc= nlp(abstract_paragraph)
    abstract_sentences = [str(sentence) for sentence in list(doc.sents)]
    abstract_sent_dictList=[]

    #Total Line parameter
    total_line_in_abstract = len(abstract_sentences)-1  #as indexing will start from 0 in for loop

    for index,value in enumerate(abstract_sentences):
        abstract_sent_dict={}
        abstract_sent_dict['line_number'] = index #line no parameter
        abstract_sent_dict['text'] =value #text
        abstract_sent_dict['total_lines'] = total_line_in_abstract
        abstract_sent_dict['line_position'] = str(index) +"_of_" + str(total_line_in_abstract)
        abstract_sent_dictList.append(abstract_sent_dict)

    # Get all line_number values from sample abstract
    test_abstract_line_numbers = [sent_dictionary['line_number'] for sent_dictionary in abstract_sent_dictList]
    # One-hot encode to same depth as training data, so model accepts right input shape
    abstract_line_numbers_one_hot = tf.one_hot(test_abstract_line_numbers,depth=15)

    test_abstract_total_lines = [sent_dictionary['total_lines'] for sent_dictionary in abstract_sent_dictList]
    abstract_total_lines_one_hot = tf.one_hot(test_abstract_total_lines,20)
    
    #list(sample_sentence) will convert into list of characters
    abstract_characters = [' '.join(list(sentence)) for sentence in abstract_sentences]
    return abstract_sentences,abstract_characters,abstract_line_numbers_one_hot,abstract_total_lines_one_hot

'''
Making Prediction 
'''
    
def make_prediction(url):
    abstract_sentences,abstract_characters,abstract_line_numbers_one_hot,abstract_total_lines_one_hot=preprocess_abstract_data(url)
    class_names= ['BACKGROUND','CONCLUSIONS','METHODS','OBJECTIVE','RESULTS'] #our model has encode in this format that's why
    skimlit_model = tf.keras.models.load_model('tribid_token_char_lineNo_totalLine_embedded_model')
    skimlit_model_pred_prob = skimlit_model.predict(x=((tf.constant(abstract_sentences),
                                                        tf.constant(abstract_characters),
                                                        abstract_line_numbers_one_hot,
                                                        abstract_total_lines_one_hot
                                                        )))
    skimlit_model_preds = tf.argmax(skimlit_model_pred_prob,axis=1)
    skimlit_model_prediction_label = [class_names[prediction] for prediction in skimlit_model_preds]
    # for text,target in zip(abstract_sentences,skimlit_model_prediction_label):
    #     #print(f"for Line Position:\n{linePosition}\nText:\n{text}\nPredicted Target Label:\n{target}\n")
    #     print(f"{target} : {text}")    
    
    
    background_text,conclusion_text,methods_text,objective_text,results_text ="", "", "", "", ""
    for text,target in zip(abstract_sentences,skimlit_model_prediction_label):
        if target=="BACKGROUND":
            background_text +=text
        elif target == "CONCLUSIONS":
            conclusion_text +=text
        elif target == "METHODS":
            methods_text += text
        elif target == "OBJECTIVE":
            objective_text += text
        elif target == "RESULTS":
            results_text += text
    
    #gradio output component
    abstract_output = [objective_text,background_text,methods_text,results_text,conclusion_text]
    return abstract_output

inputs = gr.Textbox(placeholder="Paste your PubMed article URL here",interactive=True,label="URL")
outputs=[gr.Textbox(label="OBJECTIVE"),
         gr.Textbox(label="BACKGROUND"),
         gr.Textbox(label="METHODS"),
         gr.Textbox(label="RESULTS"),
         gr.Textbox(label="CONCLUSIONS")]

examples = [
    ["https://pubmed.ncbi.nlm.nih.gov/20232240/"],
    ["https://pubmed.ncbi.nlm.nih.gov/22244707/"]
]

app = gr.Interface(fn=make_prediction,
                   inputs=inputs,
                   outputs=outputs,
                   title="PubMed Article Abstract Skimming Tool",
                   description="Classifies abstract sentences of your PubMed article into the role they play (e.g. objective, methods, results, etc)  to get the overview of literature within seconds.",
                   theme="soft",
                   examples=examples,
                   )
app.launch()