Transforming PDFs: Summarizing Information with Transformers in Python

Akshit Behera 12 Jul, 2023 • 12 min read


Transformers are revolutionizing natural language processing, providing accurate text representations by capturing word relationships. Extracting critical information from PDFs is vital today, and transformers offer an efficient solution for automating PDF summarization. The adaptability of transformers makes these models invaluable for handling various document formats. Applications span industries like law, finance, and academia. This article presents a Python project showcasing PDF summarization using transformers. By following the guide, readers can unlock the transformative potential of these models and gain insights from extensive PDFs. Embrace the power of transformers for automated document analysis and embark on a journey of efficiency.

Transforming PDFs: Summarizing Information with Transformers in Python

Learning Objectives

In this project, readers will gain critical skills that align with the outlined learning objectives. These objectives include:

  1. Gain a deep understanding of the complex operations of transformers, which can revolutionize how natural language processing tasks, such as text summarization, are tackled.
  2. Learn how to perform PDF parsing and text extraction using advanced Python libraries like PyPDF2, allowing text extraction from PDF files. Address the complexities that come with dealing with a variety of formats and layouts.
  3. Get acquainted with essential preprocessing techniques that enhance text summarization quality. Accomplish tasks such as tokenization, stop word removal, and handling unique characters or formatting complexities.
  4. Tap into the power of transformers by applying advanced text summarization techniques. Gain practical experience in using pre-trained transformer models like T5 for the extractive summarization of PDF documents.

This article was published as a part of the Data Science Blogathon.

Project Description

Within this project, our objective revolves around harnessing the potential of Python transformers to accomplish automated PDF file summarization. We aim to optimize the extraction of vital details from PDFs, mitigating the laborious nature of the manual analysis. By employing transformers for text summarization, we endeavor to expedite document analysis, thereby heightening efficiency and productivity. By implementing pre-trained transformer models, we seek to generate succinct summaries that encapsulate crucial information within PDF documents. Empowering readers with the expertise to deploy transformers for streamlined PDF summarization in their projects constitutes the core of our project’s intent.

PDF Summarization | Text Extraction | Python | Transformers Model

Problem Statement

Minimizing the time and human effort required to extract critical information from PDF documents constitutes a significant hurdle. Manually summarizing lengthy PDFs is characterized by its labor-intensive nature, rendering it prone to human errors and limited in its capacity to handle extensive volumes of textual data. These obstacles significantly impede efficiency and productivity in document analysis, particularly when confronted with an overwhelming number of PDFs.

The importance of automating this process using transformers cannot be overstated. By harnessing the transformative capabilities of transformers, we can seamlessly extract pertinent details, encompassing essential insights, noteworthy discoveries, and pivotal arguments, from PDF documents autonomously. The deployment of transformers optimizes the summarization workflow, alleviates human involvement, and expedites the retrieval of critical information. This automation empowers professionals across diverse domains to make swift, well-informed decisions, remain abreast of cutting-edge research, and effectively navigate PDF documents’ copious amounts of information.

PDF Summarization | Text Extraction | Python | Transformers Model


Our innovative approach for this project entails harnessing transformers to summarize PDF documents. We will emphasize extractive text summarization, which involves extracting salient information from the original text instead of generating entirely new sentences. This aligns seamlessly with our project’s objectives of producing concise and informative summaries that encapsulate the pivotal details gleaned from PDFs.

To materialize this approach, we shall proceed as follows:

  1. PDF Parsing and Text Extraction: We will use the PyPDF2 library to navigate the PDF file, extracting the textual content from each page. The extracted text will be meticulously compiled for subsequent processing.
  2. Text Encoding and Summarization: Employing the transformers library, we will harness the power of the T5ForConditionalGeneration model. With its pre-trained capabilities, this model proves instrumental for text generation tasks. Initializing the model and tokenizer, we shall encode the extracted text using the T5 tokenizer, ensuring proper representation for subsequent steps.
  3. Summary Generation: Facilitated by the encoded input, we shall summon the power of the T5ForConditionalGeneration model to generate the summary. By exercising fine-grained control over the summary’s length, quality, and other pertinent attributes, we can tailor the summarization process to meet our specific requirements. The outcome of this step shall manifest as an encoded representation of the generated summary.
  4. Decoding the Summary: The culminating step involves decoding the generated summary into a human-readable format. This critical process is executed using the tokenizer, which expertly translates the encoded summary into coherent, comprehensible text. The decoded summary is a concise, information-rich summary derived from the original PDF document.


In this context, let’s consider a hypothetical scenario that revolves around the human resources function of a multinational corporation, XYZ Enterprises. XYZ Enterprises receives a substantial volume of PDF resumes and job applications from candidates across the globe for various job positions. Reviewing each application manually and extracting relevant information poses a significant challenge for the HR team due to time constraints and potential inconsistencies.

XYZ Enterprises can streamline its candidate evaluation process by employing transformers for PDF summarizations. With the transformative power of transformers, the HR team can automate the extraction of vital details from resumes and applications. By generating concise summaries, transformers can highlight critical information such as qualifications, experience, skills, and achievements, enabling quick and efficient evaluation.

By leveraging transformers for PDF summarization in this scenario, XYZ Enterprises can expedite the candidate screening process, ensuring that only the most relevant and qualified candidates proceed to subsequent selection rounds. The utilization of transformers demonstrates their practical application in enhancing efficiency and accuracy in the human resources function, facilitating a more streamlined and effective hiring process for the organization.

Setting Up the Environment

We must meticulously establish a Python environment infused with the requisite libraries and dependencies to embark upon the PDF summarization project with transformers. Below, we outline the process step by step:

  1. Python Installation: Verify that Python is installed on your system. Access the official Python website ( to download the latest version suitable for your operating system. Refer to the provided installation instructions to set up Python successfully.
  2. Library Installation: Initiate a terminal, command prompt session, or any IDE and employ pip, the Python package manager, to install the indispensable libraries. Execute the subsequent commands:
pip install PyPDF2
pip install transformers

These commands will install the PyPDF2 library for PDF parsing and the transformers library for leveraging transformer models.

3. Additional Requirements: Tailor your environment to accommodate specific project needs by considering potential supplementary libraries or dependencies. For instance, if your project demands the utilization of a particular pre-trained transformer model such as BERT, installing the corresponding Hugging Face transformers model is imperative:

pip install transformers==4.12.0

4. Text Summarization Model: Certain transformer models employed for text summarization may entail supplementary downloads or installations. Comply with the instructions provided in the model’s documentation to download and configure the essential files, should the need arise.

Data Preparation

A meticulous approach to collecting and organizing PDF documents is essential to lay the foundation for the project and ensure seamless data handling. Moreover, addressing PDF format variations and performing OCR on scanned PDFs requires careful consideration. Here, we outline the recommended steps:

Data Collection

Gather the PDF documents necessary for the project and ensure they are accessible within the AI environment. For our purpose, let us assume that HR is hiring for a data science role and has received the resumes from four candidates. Upload the resumes in PDF format to the designated directory, in this case, the ‘/content/pdf_files’ directory. Verify that the PDF files are readily available for subsequent processing steps.

import os
import PyPDF2
from PIL import Image
import pytesseract

# Directory for storing PDF resumes and job applications
pdf_directory = '/content/pdf_files'

# Directory for storing extracted text from PDFs
text_directory = '/content/extracted_text'

# OCR output directory for scanned PDFs
ocr_directory = '/content/ocr_output'

# Create directories if they don't exist
os.makedirs(pdf_directory, exist_ok=True)
os.makedirs(text_directory, exist_ok=True)
os.makedirs(ocr_directory, exist_ok=True)

Organizing the PDFs

Create a coherent folder structure to organize the PDF files systematically. Utilize appropriate categorization methods such as job positions, application dates, or candidate names to ensure a logical arrangement of the files. This organizational framework facilitates easy retrieval and enhances data handling efficiency throughout the project.

PDF Format Handling and Text Extraction

PDF files often exhibit diverse formats, layouts, and encodings. Account for these variations by employing appropriate preprocessing techniques. In the provided code snippet, the PyPDF2 library is utilized to open each PDF file, extract text from each page, and save the extracted text as individual text files. The extracted text is stored in the ‘/content/extracted_text.’ directory. This step standardizes the data and ensures that the text content is readily accessible for further processing stages.

for file_name in os.listdir(pdf_directory):
    if file_name.endswith('.pdf'):
        # Open the PDF file
        with open(os.path.join(pdf_directory, file_name), 'rb') as file:
            # Create a PDF reader object
            reader = PyPDF2.PdfReader(file)

            # Extract text from each page
            text = ''
            for page in reader.pages:
                text += page.extract_text()

            # Save the extracted text as a text file
            text_file_name = file_name.replace('.pdf', '.txt')
            text_file_path = os.path.join(text_directory, text_file_name)
            with open(text_file_path, 'w') as text_file:

OCR for Scanned PDFs (Optional)

Scanned PDFs or PDFs containing images require Optical Character Recognition (OCR) techniques to convert the embedded images into machine-readable text. The code snippet showcases the utilization of the pytesseract library to perform OCR on scanned PDFs. The OCR text is saved as separate files in the ‘/content/ocr_output’ directory. This optional step unlocks the text content embedded within scanned PDFs, broadening the scope of data processing.

# Optional Step
for file_name in os.listdir(pdf_directory):
    if file_name.endswith('.pdf'):
        # Open the PDF file
        with, file_name)) as img:
            # Perform OCR using pytesseract
            ocr_text = pytesseract.image_to_string(img, lang='eng')

            # Save the OCR output as a text file
            ocr_file_name = file_name.replace('.pdf', '.txt')
            ocr_file_path = os.path.join(ocr_directory, ocr_file_name)
            with open(ocr_file_path, 'w') as ocr_file:

PDF Parsing and Text Extraction

To access the valuable information within PDF resumes and job applications, it is crucial to parse the
PDF files and extract the text content. This process involves addressing various formats, layouts, and challenges that may arise. Let’s delve into the steps required for parsing and extracting text from PDF files:

  1. Set up the directory where the PDF resumes and job applications are stored. In this case, we utilize the ‘/content/pdf_files’ directory as the designated location.
  2. Obtain a list of files present in the specified PDF directory. Filter out only the PDF files by examining their extensions and considering those ending with ‘.pdf.’
  3. Employ a loop to iterate over each resume file. For every file, follow the subsequent procedures:

A. Opening the File: Open the resume file in ‘rb’ (read binary) mode using the open() function and a context manager. This ensures secure file handling and automatic closure upon completion.

B. Creating a PDF Reader Object:  To establish a PDF reader object, use the PyPDF2 library’s PdfReader() functiont. This object enables access to the content within the PDF file.

C. Extracting Text from Pages: Extract the text content from each PDF file page. Employ a loop to
iterate through the pages using the pages attribute of the PDF reader object. Extract the text from each page using the extract_text() method and concatenate it with the existing text.

D. The extracted text within the text variable is accumulated throughout the extraction process. This variable holds the combined text content derived from all pages within the PDF file.

# Directory for storing PDF resumes and job applications
pdf_directory = '/content/pdf_files'

resume_files = []
for file_name in os.listdir(pdf_directory):
    if file_name.endswith('.pdf'):
        resume_files.append(os.path.join(pdf_directory, file_name))

resume_summaries = []  # To store the generated summaries

# Loop through each resume file
for resume_file in resume_files:
    with open(resume_file, 'rb') as file:
        # Create a PDF reader object
        reader = PyPDF2.PdfReader(file)

        # Extract text from each page
        text = ''
        for page in reader.pages:
            text += page.extract_text()

Implementing Text Summarization with Transformers

In the pursuit of text summarization, transformers have emerged as cutting-edge deep learning architecture. They exhibit exceptional capabilities in condensing information while retaining the essence of the original text. Let’s dive into the implementation steps, highlighting the utilization of pre-trained models like T5 for text summarization.

  1. Model and Tokenizer Initialization: Begin by initializing the T5 model and tokenizer. These components serve as the backbone of our text summarization pipeline. In this instance, we instantiate the T5 model with the “t5-base” architecture.
  2. Text Encoding: Prepare the input text for summarization by encoding it using the tokenizer. This step converts the text into a numerical representation that the model can understand. To guide the model towards summarization, we prepend the text with the instruction “summarize.”
  3. Summary Generation: Leverage the power of the model to generate a summary. Employing a beam search algorithm, the model explores various paths to produce the most fitting summary. Fine-tune the summary length, quality, and other parameters, such as length penalty and the number of beams, to achieve desired outcomes.
  4. Summary Decoding: Using the tokenizer, decode the summary’s numerical representation into human-readable text. This decoding step allows us to obtain a comprehensive overview that encapsulates the crucial details from the original text.
  5. Storing the Summaries: Capture the generated summaries by keeping them in the resume_summaries list, providing a centralized repository for future utilization.
  6. Printing the Summaries: Iterate through the resume_summaries list and present the generated summaries for each resume, accompanied by an appropriate identifier.
# Continuing the loop from the previous step
        from transformers 
        import T5ForConditionalGeneration,T5Tokenizer

        # Initialize the model and tokenizer
        model = T5ForConditionalGeneration.
        tokenizer = T5Tokenizer.

        # Encode the text
        inputs = tokenizer.encode("summarize: " + text, 
        return_tensors="pt", max_length=1000, 

        # Generate the summary
        outputs = model.generate(inputs, 
        max_length=1000, min_length=100, 
        length_penalty=2.0, num_beams=4, 

        # Decode the summary
        summary = tokenizer.decode(outputs[0])


# Print the generated summaries for each resume
for i, summary in enumerate(resume_summaries):
    print(f"Summary for Resume {i+1}:")


For the four resumes we processed, we get the following output.

Summary for Resume 1:

<pad> 8+ years of IT experience with 5+ years in the big data domain, currently working as a Lead Data Engineer with AirisData with expertise in Pyspark, Spark SQL, PySpark, Data Frame, RDD. Credit Suisse: Rave excellence award in December 2020 • Brillio Technologies: Employee of the quarter in December 2020 • Centurylink Technologies: Spot award in Nov 2016 • Centirylink Technologies: Outstanding team award in September 2015.</s>

Summary for Resume 2:

<pad> Designed and implemented a Hadoop cluster to store and process large amounts of data. Developed Spark applications for data processing, data cleansing, and data analysis. Built data pipelines using Apache NiFi to automate data flow and processing. Developed frontend and backend for multiple clients using HTML, CSS, JavaScript, Django, Python, and Android Studio. received insta award for bug-free delivery. <unk> Developed data visualization dashboards using Tableau to provide insights into business trends and performance.</s>

Summary for Resume 3:

<pad> 5.7 years of Experience as a data engineer and data scientist in the automotive industry. Bachelor’s degree in Mechanical engineering from Pune University secured first class with distinction with an overall aggregate of 76%. strong knowledge of Pyspark SQL dataframes and RDD functions. Knowledge of data management, ETL, and RDBMS query language. worked on more than 30 data science projects from Kaggle, Scikit -learn & GitHub.</s>

Summary for Resume 4:

<pad> offer strong technical acumen with diverse abilities in the relational database at cloud-based
data warehouses and data lake. Process structured and semi-structured datasets using the PySpark ETL pipeline, which Apache Airflow automates under the Big data ecosystem. Managed multiple projects using rigor improvements, succession planning to de-risk programs, client engagement workshops, baseline expectations, and SLAs. Worked closely with upper management to ensure the project’s scope and direction were on schedule.</s>

Other Real World Applications

PDF summarization using transformers has numerous practical applications across industries. Let’s
explore some real-world scenarios where this technology can be utilized and discuss possibilities for further advancements:

  1. Confidential Document Summarization: In industries dealing with sensitive or personal information, such as finance or legal sectors, transformers can summarize critical details from confidential PDF documents. Summaries can be generated while preserving data privacy and ensuring compliance with security regulations. Future enhancements may involve developing secure summarization frameworks that protect sensitive information while providing valuable insights.
  2. Medical and Healthcare Reports: Medical professionals often struggle to extract crucial information from extensive medical reports and research papers. By leveraging transformers for PDF summarization, doctors, researchers, and healthcare providers can quickly obtain summaries highlighting essential findings, diagnoses, or treatment recommendations. Further advancements may include domain-specific models fine-tuned on medical literature to ensure accurate and contextually relevant summaries.
  3. Crisis Management and Emergency Response: During crises or emergencies, decision-makers must quickly process vast amounts of information. PDF summarization with transformers can assist in summarizing situation reports, incident updates, or risk assessments, enabling faster decision-making and effective emergency response coordination. Future enhancements may involve real-time summarization techniques to provide up-to-date and concise summaries during critical situations.

Limitations & Challenges

While discussing the limitations and challenges of PDF summarization using transformers, it’s essential to consider the broader context and acknowledge the potential complexities associated with this technology. Here, we highlight some factors that can impact the performance and effectiveness of PDF summarization:

  1. Complex Document Structures:  Firstly, transformers may face challenges with PDF documents containing intricate structures, such as tables, diagrams, or non-standard formatting. Extracting information from such complex layouts and representing them effectively in summaries can be demanding. The transformer models may struggle to maintain coherence or accurately capture the intended meaning, resulting in suboptimal summaries.
  2. Limited Context Understanding: Moreover, transformers excel at capturing local dependencies within a given context. However, they may struggle with broader contextual understanding, especially when summaries need to consider information from multiple sections or pages within a PDF document. Generating comprehensive and coherent summaries across different sections or chapters remains a challenge.
  3. Language and Domain Bias: Furthermore, transformer models trained on large-scale datasets can inadvertently reflect biases present in the training data. These biases can manifest in the summaries generated, leading to skewed or inaccurate results, mainly when dealing with specific domains, technical jargon, or cultural nuances. Addressing language and domain bias requires careful dataset curation, fine-tuning, and continuous monitoring.


Throughout this article, we have covered essential aspects of PDF summarization using transformers. We have delved into the capabilities and applications of transformers in natural language processing tasks, particularly in summarizing information from PDF documents. Readers have gained valuable knowledge and skills in this domain by exploring the provided code examples and step-by-step instructions.

Key Takeaways

  1. Understanding the fundamentals of transformers and their role in text summarization.
  2. Implementing PDF parsing and text extraction to obtain text content from PDF files.
  3. Leveraging pre-trained transformer models, such as T5, for generating concise summaries.
  4. Exploring real-world applications of PDF summarization in various industries.

By acquiring these skills, one can enhance information processing capabilities, streamline document analysis, and leverage the power of transformers to extract critical insights efficiently. Generating accurate and concise summaries from PDF documents allows for improved decision-making, faster information retrieval, and enhanced knowledge management.

Frequently Asked Questions

Q1. What is the significance of transformers in NLP?

A. Transformers are advanced models revolutionizing natural language processing (NLP). They employ
self-attention mechanisms to capture intricate word relationships, enabling accurate text representation and understanding.

Q2. How do transformers contribute to text summarization?

A. Transformers excel at condensing lengthy documents into concise summaries by identifying critical information and capturing contextual nuances. Their ability to understand language dynamics makes them highly effective in automating the summarization process.

Q3. How do transformers impact PDF summarization?

A. Transformers are pivotal in PDF summarization, extracting essential details, and capturing sentence relationships. Their adaptability allows the seamless processing of diverse PDF formats and layouts, facilitating automated information extraction from PDF documents.

Q4. Can transformers be tailored for domain-specific summarization?

A. Transformers can be fine-tuned for specific domains. By training them on domain-specific data, transformers can generate more accurate summaries tailored to industries like finance, law, or scientific research.

Q5. What advantages do transformers offer for text summarization?

A. Transformers provide efficiency, accuracy, and the ability to handle complex documents. They streamline information extraction, enhance decision-making processes, and optimize document analysis across industries. Transformers empower users with automated, insightful summaries.

The media shown in this article is not owned by Analytics Vidhya and is used at the Author’s discretion.

Akshit Behera 12 Jul 2023

Frequently Asked Questions

Lorem ipsum dolor sit amet, consectetur adipiscing elit,

Responses From Readers

  • [tta_listen_btn class="listen"]