app.py

# https://huggingface.co/spaces/micknikolic/pdf-abstract-summarizer

# Here are the imports

import pdfplumber
import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
from bert_score import score as bert_score
from io import BytesIO
from scipy.io.wavfile import write as write_wav
import gradio as gr
import numpy as np
from gtts import gTTS

# Here is the code

##Instantiating model and tokenizer.

pegasus_research_model = AutoModelForSeq2SeqLM.from_pretrained("UNIST-Eunchan/Research-Paper-Summarization-Pegasus-x-ArXiv")
pegasus_research_model = pegasus_research_model.to("cuda")
pegasus_research_tokenizer = AutoTokenizer.from_pretrained("UNIST-Eunchan/Research-Paper-Summarization-Pegasus-x-ArXiv")

##Defining functions.

def extract_abstract(uploaded_file):
    with pdfplumber.open(uploaded_file) as pdf:
        abstract = ""
        for page in pdf.pages:
            text = page.extract_text(x_tolerance=1, y_tolerance=1)
            if text:
                text_lower = text.lower()
                if "abstract" in text_lower:
                    start_index = text_lower.find("abstract")
                    end_index = text_lower.find("introduction", start_index)
                    if end_index == -1:
                        end_index = len(text)
                    abstract = text[start_index:end_index]
                    break
    return abstract

def text_chunker(text, tokenizer, max_tokens):
    tokens = tokenizer.encode(text)
    num_chunks = len(tokens) // max_tokens + (len(tokens) % max_tokens > 0)
    chunked_tokens = [
        tokens[i * max_tokens : (i + 1) * max_tokens] for i in range(num_chunks)
    ]
    chunked_text = [tokenizer.decode(chunk, skip_special_tokens=True) for chunk in chunked_tokens]
    return chunked_text

def pegasus_research_summarize(text):
    inputs = pegasus_research_tokenizer.encode("summarize: " + text,
                                               return_tensors="pt",
                                               max_length=800,
                                               truncation=True)
    summary_ids = pegasus_research_model.generate(inputs.to("cuda"),
                                                  max_length=150,
                                                  min_length=40,
                                                  length_penalty=0.5,
                                                  num_beams=4,
                                                  early_stopping=True
                                                  )
    summary = pegasus_research_tokenizer.decode(summary_ids[0],
                                                skip_special_tokens=True)
    return summary

def select_best_sentence(summary, reference_text):
    sentences = summary.split('.')
    sentences = [sentence.strip() for sentence in sentences if sentence.strip()]

    if not sentences:
        return "", "0.00 (Very Low Similarity)"

    _, _, f1_scores = bert_score(sentences, [reference_text] * len(sentences), lang="en", rescale_with_baseline=True)
    best_sentence_index = np.argmax(f1_scores)
    best_sentence = sentences[best_sentence_index]
    best_f1_score = round(f1_scores[best_sentence_index].item(), 2)

    score_label = ""
    if best_f1_score <= 0.20:
        score_label = " (Very Low Similarity)"
    elif best_f1_score <= 0.40:
        score_label = " (Low Similarity)"
    elif best_f1_score <= 0.60:
        score_label = " (Moderate Similarity)"
    elif best_f1_score <= 0.80:
        score_label = " (High Similarity)"
    else:
        score_label = " (Very High Similarity)"

    best_f1_score_with_label = f"{best_f1_score}{score_label}"

    return best_sentence, best_f1_score_with_label


def convert_to_audio(text):
    tts = gTTS(text, lang='en')
    buffer = BytesIO()
    tts.write_to_fp(buffer)
    buffer.seek(0)
    audio_bytes = buffer.read()

    return audio_bytes

def pr_recursive_summarize(text, reference_text, recursion_l=0):
    recursion_level = recursion_l + 1
    print(f"Pegasus Research level: {recursion_level}\n")
    tokens = pegasus_research_tokenizer.tokenize(text)
    expectedCountOfChunks = max(len(tokens) / 150, 1)
    max_length = int(len(tokens) / expectedCountOfChunks) + 2

    chunks = text_chunker(text, pegasus_research_tokenizer, max_tokens=800)
    print(f"Number of chunks: {len(chunks)}")

    summaries = []
    for i, chunk in enumerate(chunks, 1):
        print(f"Chunk no.{i}:")
        print(chunk, "\n")
        summary = pegasus_research_summarize(chunk)
        print("Summary:", summary)
        summaries.append(summary)
        print("_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_")
        torch.cuda.empty_cache()

    concatenated_summary = ' '.join(summaries)
    tokens = pegasus_research_tokenizer.tokenize(concatenated_summary)

    if len(tokens) > 50 and recursion_level <= 10:
        print("Recursive")
        return pr_recursive_summarize(concatenated_summary, reference_text, recursion_level)
    else:
        final_summary = concatenated_summary
        if len(chunks) > 1:
            final_summary = pegasus_research_summarize(concatenated_summary)

        sentences = final_summary.split(".")
        sentences = [s.strip() for s in sentences if s.strip()]

        if not sentences:
            return None, 0.0

        p, r, f1_scores = bert_score(sentences, [reference_text]*len(sentences), lang="en")
        best_sentence_index = np.argmax(f1_scores)
        best_sentence = sentences[best_sentence_index]
        best_f1_score = f1_scores[best_sentence_index].item()

        return best_sentence, best_f1_score

def summarize_and_convert_to_audio(pdf_file):
    abstract_text = extract_abstract(pdf_file)
    if not abstract_text:
        return "No 'abstract' section found in the uploaded PDF. Please upload a different PDF.", None, "0.00 (Very Low Similarity)"

    best_sentence, best_f1_score = pr_recursive_summarize(abstract_text, abstract_text)
    audio_bytes = convert_to_audio(best_sentence)

    return audio_bytes, best_sentence, f"{best_f1_score:.2f} (Very High Similarity)"

##Building the Gradio UI.

iface = gr.Interface(
    fn=summarize_and_convert_to_audio,
    inputs=gr.File(label="Upload PDF"),
    outputs=[
        gr.Audio(label="Audio"),
        gr.Textbox(label="Summary sentence"),
        gr.Textbox(label="Bert F1-Score")
    ],
    title="PDF Abstract Summarizer and Audio Converter",
    description="Upload a PDF file to extract and summarize its 'abstract' section. The best summary sentence based on its Bert F1-score will be converted into speech and the score's interpretation will be displayed. (A PDF file needs to contain the Abstract section.)",
    examples=[
        "Hidden Technical Debt in Machine Learning Systems.pdf",
        "On the Cross-Entropy Method.pdf"
    ]
)

iface.launch()