Presentify: Automated Presentation Slide Generation from

Research Papers using NLP and Deep Learning (May 2024)

Atul Shreewastav1 , Bidhan Acharya1 , Nischal Paudel1 , and Yugratna Humagain1
1
Department of Electronics and Computer Engineering, IOE, Thapathali Campus

May 13, 2024

Posted on 13 May 2024 — CC-BY 4.0 — https://doi.org/10.36227/techrxiv.171561007.70987286/v1 — e-Prints posted on TechRxiv are preliminary reports that are not peer reviewed. They should not b...

1
 IOE, Thapathali Campus

Presentify: Automated Presentation Slide

Generation from Research Papers using
NLP and Deep Learning
(May 2024)
Atul Shreewastav (THA077BCT013)1 , Bidhan Acharya (THA077BCT015)2 , Nischal Paudel (THA077BCT028)3 and
Yugratna Humagain(THA077BCT047)4
1,2,3
Department of Electronics and Computer Engineering, IOE, Thapathali Campus, Kathmandu 44600, Nepal

ABSTRACT This study presents a novel approach for automating the generation of presentation slides
from academic research papers. We leverage Natural Language Processing (NLP) techniques, particularly
a fine-tuned T5 transformer model, to streamline knowledge dissemination within the computer science
domain. The model, trained on a custom dataset of research articles, extracts key information and
transforms it into well-structured presentation slides. The T5 model’s proficiency in text-to-text transfer
tasks is enhanced through fine-tuning with domain-specific metrics. This approach enables the automatic
conversion of research articles into concise and informative presentation slides. This research has the
potential to significantly impact how technical information is summarized and communicated within the
research community, fostering greater efficiency and clarity.

INDEX TERMS NLP, T5 model, Gemini model ,Slide generation, Text extraction and Section extraction

I INTRODUCTION II RELATED WORK

The effective communication of complex information in Automating the generation of presentation slides from re-
professional and educational settings often relies on clear search papers has become an active area of research, with
and well-structured presentations. However, manually several approaches emerging in recent years. One notable
creating these presentations can be time-consuming and example is ”PPSGen, Learning-Based Presentation Slides
resource-intensive, hindering efficient knowledge dissem- Generation for Academic Papers” [1] that emphasizes fac-
ination. Recent advancements in Natural Language Pro- tual accuracy by employing sentence selection algorithms.
cessing (NLP) offer promising avenues to streamline this Another approach, titled ”Learning Based Slide Gener-
process by automating content extraction and organiza- ator” [2] utilizes machine learning techniques to extract
tion. relevant information for slide generation . Additionally,
This research addresses the challenges associated with research presented in ”Technique for Generating Auto-
manual slide creation by proposing a novel methodol- matic Slides on the basis of Paper Structure Analysis” [3]
ogy for the automated generation of presentation slides focuses on analyzing the structure of research papers to
from academic research papers in the computer science generate slides, highlighting the relationship between sec-
domain. We leverage NLP techniques, specifically a fine- tions and their content. Furthermore, ”DOC2PPT: Auto-
tuned T5 transformer model, to extract key information, matic Presentation Slides Generation from Scientific Doc-
summarize content, and generate well-structured presen- uments” explores automatic slide generation from sci-
tation slides. This approach aims to empower users, such entific documents, emphasizing paraphrasing and layout
as researchers, educators, and students, to focus on con- within a hierarchical model [4].
tent creation and knowledge dissemination by automating
the time-consuming tasks of information extraction and These efforts collectively demonstrate the potential of
formatting. Ultimately, this project aspires to contribute NLP techniques to streamline the often time-consuming
to a more efficient and effective presentation creation pro- process of slide creation. However, existing approaches
cess, benefiting diverse professional and educational con- face challenges in ensuring the accuracy, relevance, and
texts. visual appeal of the generated slides. Our proposed sys-
tem addresses these challenges by leveraging a fine-tuned

1 Presentify, May 2024

 IOE, Thapathali Campus A. Shreewastav, B. Acharya, N. Paudel and Y. Humagain: Presentify

T5 transformer model for content summarization and a summaries for each section. The T5 model has been
dedicated section extraction model (e.g., Gemini model) specifically adjusted to optimize the summarization pro-
to preserve the structure and key points of the original re- cess for research articles.
search paper.
4 Slide Generation
III METHODOLOGY The summarized content for each section is then passed
A DATASET DESCRIPTION to the presentation module, which formats the information
This research leverages a domain-specific dataset sourced into a standardized presentation slide format. This ensures
exclusively from arXiv.org, a reputable online repository the slides have a clear and coherent structure, with each
for computer science research articles. This targeted se- slide covering a specific section of the research paper.
lection ensures the dataset’s relevance to the research fo- 5 Customization Options
cus and provides a well-defined corpus for model training The system also provide options for users to customize the
and evaluation. appearance and layout of the generated slides, allowing
The dataset consists of 15,655 entries meticulously di- them to tailor the presentation to their specific needs.
vided into training (10,958), testing (3,131), and valida-
tion sets (1,566) to facilitate robust model performance as- C SYSTEM REQUIREMENTS
sessment. Notably, ”Computer Vision and Pattern Recog- Following are the basic requirements for the system:
nition” emerges as the most prevalent category, with 675 1 Hardware Requirements
instances. • Multi-core CPU (Intel Xeon or AMD Ryzen)
Each data point comprises two key features: text and sum-
mary. The ”text” column houses the extracted content • RAM (16GB or more)
from research articles, potentially including sections like
introductions, literature reviews, methodologies, results, • System Type – 64-bit Operating System
and conclusions. The corresponding ”summary” column
contains a concise representation of the extracted text, of- 2 Software Requirements
fering a succinct version of the original content. This • Operating System – Windows
dataset serves as the foundation for training and evalu-
ating the employed NLP models. Ultimately, it facilitates • RAM Programming Language – Python
the optimization of the system’s ability to generate infor- • Python Packages - Numpy, Pandas, BeautifulSoup,
mative presentation slides tailored for academic research PyMuPDF
papers within the computer science domain.
B PROPOSED METHODOLOGY Python
This research proposes a novel methodology for automat- Python is a high level, general purpose, interpreted, dy-
ing the generation of presentation slides from academic namic programming language. Python supports multiple
research papers in the computer science domain. The programming paradigms, including object-oriented, im-
methodology utilizes Natural Language Processing (NLP) perative and functional programming or procedural styles.
techniques to achieve information extraction, summariza- NumPy
tion, and presentation structure creation. NumPy is a Python library for numerical computing, of-
fering powerful data structures and tools for working with
1 Text Extraction arrays, matrices, and mathematical functions. Its efficient
The system first extracts the textual content from the PDF operations make it essential for scientific computing, data
file using PyMuPDF, a powerful Python library for work- analysis, and machine learning in Python.
ing with PDFs. This involves parsing the PDF and ex- Pandas
tracting the text, including sections, paragraphs, and other A Python library specializing in data manipulation and
relevant information. analysis, providing high performance, easy-to-use data
2 Section Extraction with Gemini structures and tools for working with structured data like
tables and time series.
The extracted text is then passed to Gemini, a tool capa-
BeautifulSoup
ble of identifying and extracting distinct sections from the
Beautiful Soup is a Python package for parsing HTML
research paper, such as Introduction, Literature Review,
and XML documents (including having malformed
Methodology, Results, and Conclusions.
markup, i.e., non-closed tags, so named after tag soup).
3 Summarization with Fine-Tuned T5 Model It creates a parse tree for parsed pages that can be used to
The sections extracted by Gemini are then fed into a fine- extract data from HTML, which is useful for web scrap-
tuned T5 model, which transforms the text into concise ing. Beautiful Soup

2 Presentify, May 2024

 IOE, Thapathali Campus A. Shreewastav, B. Acharya, N. Paudel and Y. Humagain: Presentify

PythonPPTX IV RESULTS
PythonPPTX is a Python library used for creating and A User Interface
manipulating Microsoft PowerPoint (.pptx) files program-
matically. With this library, you can automate the pro-
cess of generating presentations by adding slides, insert-
ing text, images, shapes, tables, and charts, and customiz-
ing their properties such as formatting, styles, and layouts.
It enables tasks like batch processing multiple presenta-
tions, dynamically generating slides based on data, and
applying design templates for consistent styling. Over-
all, PythonPPTX streamlines the creation and modifica-
tion of PowerPoint files, making it a valuable tool for re-
port generation, data visualization, and automated presen-
tation generation.

D SYSTEM BLOCK DIAGRAM

Figure 2: UI for PDF Uploader

Figure 1: Block Diagram for Presntify

3 Presentify, May 2024

 IOE, Thapathali Campus A. Shreewastav, B. Acharya, N. Paudel and Y. Humagain: Presentify

Figure 5: Slide 1

Figure 3: UI for Link Uploader

Figure 6: Slide 2

Figure 4: UI for Slide Preview

B Presentation slide result

Upon uploading the research paper ”A Survey on Cross-
Domain Sequential Recommendation” the following re-
sults were obtained:

Figure 7: Slide 3

4 Presentify, May 2024

 IOE, Thapathali Campus A. Shreewastav, B. Acharya, N. Paudel and Y. Humagain: Presentify

Figure 8: slide4 Figure 11: slide7

V CONCLUSION
This research presented a novel tool that utilizes Natural
Language Processing (NLP) techniques to automate the
generation of presentation slides from academic research
papers in the computer science domain. The system
leverages a fine-tuned T5 model for content
summarization and the Gemini model for section
extraction, enabling the extraction of key information
from research articles. This information is then
transformed into well-structured and visually appealing
presentation slides. The system architecture incorporates
Figure 9: slide5 functionalities for data preparation, model training, text
extraction, summarization, and slide generation. The user
interface allows for user interaction through features like
PDF upload, URL input, slide preview, and theme
selection.
The system’s performance was evaluated using loss
curves and ROUGE scores, demonstrating the
effectiveness of the T5 model in summarizing research
articles. This evaluation ensures the accuracy and
conciseness of the generated summaries. Additionally,
the user interface offers a user-friendly experience for
researchers, educators, and professionals, empowering
them to create informative presentations with greater
efficiency.
Figure 10: slide6 A ADVANTAGE
• Demonstrates precision in extracting relevant
information from documents.

• Efficiently condenses detailed content into

impactful slides.

• Versatile in engineering fields.

B DISADVANTAGE
• Projects success depends on Gemini’s accuracy.

• Users ability to customize the PowerPoint maybe

limited.

5 Presentify, May 2024

 IOE, Thapathali Campus A. Shreewastav, B. Acharya, N. Paudel and Y. Humagain: Presentify

• Number of bullets per slide and length of bullet is

Bidhan Acharya is currently
not configurable by user.
pursuing his undergraduate de-
gree in Computer Engineering at
• Resolution and positioning of image is inconsistent. IOE, Thapathali Campus. His re-
search interests encompass vari-
VI FUTURE SCOPE ous areas, including data mining,
This work lays the foundation for a user-customizable network communications, and
summarization system. Future iterations can focus on optimization theory and technol-
expanding user control over the front-end experience. ogy.(THA077BCT015)
This could include implementing a wider variety of
themes for visual appeal and allowing users to adjust text Nischal Paudel is currently pur-
size, font, and stylistic elements directly within the suing his undergraduate degree
website. Additionally, ongoing efforts will prioritize in Computer Engineering at IOE,
further refinement of the machine learning models used Thapathali Campus. His re-
for summarization. This includes continuous parameter search interests encompass vari-
optimization, exploration of advanced summarization ous areas, including data mining,
techniques, and a focus on improving the accuracy, network communications, and
quality, and precision of the generated key insights. By optimization theory and technol-
prioritizing accuracy improvements, we aim to deliver a ogy.(THA077BCT028)
robust system that consistently meets user expectations Yugratna Humagain is cur-
and effectively addresses their summarization needs. rently pursuing his undergradu-
ate degree in Computer Engi-
REFERENCES neering at IOE, Thapathali Cam-
[1] Y. Hu and X. Wan, “Ppsgen: Learning-based pus. His research interests en-
presentation slides generation for academic papers,” compass various areas, includ-
IEEE transactions on knowledge and data ing data mining, Artificial Intel-
engineering, vol. 27, no. 4, pp. 1085–1097, 2014. ligence (AI) and Machine Learn-
ing (ML) .(THA077BCT047)
[2] K. Shaj and S. S. John, “Learning based slide
generator.”

[3] E. Meshram and D. Phalke, “Technique for

generating automatic slides on the basis of paper
structure analysis,” International Journal of
Innovative Research in Science, Engineering and
Technology, vol. 5, no. 6, 2016.

[4] T.-J. Fu, W. Y. Wang, D. McDuff, and Y. Song,

“Doc2ppt: Automatic presentation slides generation
from scientific documents,” in Proceedings of the
AAAI Conference on Artificial Intelligence, vol. 36,
no. 1, 2022, pp. 634–642.

Atul Shreewastav is currently

pursuing his undergraduate de-
gree in Computer Engineer-
ing at IOE, Thapathali Cam-
pus. His research interests en-
compass various areas, includ-
ing data mining,automation, and
AI/ML.(THA077BCT013)

6 Presentify, May 2024