A study on generating notes from slides to support slide presentation

—— number of words ~ number of crutch words number of word... IV/ RESULTS AND DISCUSSION “Hello everyone, it's me, Luc Debruyne, President of GSK Vaccines, and I'm thrilled to... “

ý

USTH

VIETNAM FRANCE UNIVERSITY

“Doan Thuan An BI11 002”

“A study on generating notes from slides to support

slide presentation”

External supervisor: “Pham Quang Nhat Minh Aimesoft”

Internal supervisor: “Tran Hoang Tung — ICT lab”

TABLE OF CONTENTS

ABBREVIATIONS

with you and I’m proud of

Besides my supervisor, I’d also like to thank my thesis reviewers: Assoc Prof

LIST OF TABLES

LIST OF FIGURES

Figure 7: Histogram of speech notes’ scores

ABSTRACT

I/ INTRODUCTION

ith today’s state of the art

HI/ MATERIALS AND METHODS

Feed forward

Add & norm Add & norm Multi-head Masked multi- attention head attention

Input embedding Output embedding

pos

\ — oi PE(pos, 2i) SiN 0002 dnoaua

/ pos PE(pos, 2i + 1) = cos aa

rds of the speech note to be the slides’ textual content

tf

f(t d) _ count of tind number of words ind idf

Pegasus is |MASK2] [M It [MASK2] the model <s> It is pure white |

Pegasus is |mythical] \It is pure white ; It}names|the model

and utterance’s (the sequence

Score = sum([U S c] x WE.„e )

—— number of words ~ number of crutch words number of word

_ ay éL Wirt = We ( + e)1Z x law,

problem by computing ‘bias corrected’ m, vt

^ Thị Vr

Thị VE

IT; I, To 1, -T3 1 I, Ty

IyT, | IyT2 | T3) |IyTụy lạTị (IyT; |lạTy | (Ty Ty

INT, | InT2 | InT3 |IVTN

(2) Create dataset classifier from label text

cÍ = c],e;, cị

pị, pụ = F(CLIP(x))

zy, = ly, DiC ,Cị Ny

Nol max YY logpe(c; |x, c, wo G a) jel

i=1

N ol

Ly =-YYlogpe(G |p °, p‡,cj, ej—3)

i=1 j=1

IV/ RESULTS AND DISCUSSION

“Hello everyone, it's me, Luc Debruyne, President of GSK Vaccines, and I'm thrilled to

“So, um, Vaccines are, you know, something that really benefit all phases of life

kids They help maintain, uh, immunity and guard us against, uh, various diseases.”

mostly not aligned with the intention of the slides’ author

es’ topic However, some times the outputs are still not aligned with the intention of the slides’

today’s presentation We will

better night’s rest We'll wrap

Know when to contact your doctor

your sleep quality, it’s

@ The National Sleep Foundation (NSF)

@ Sleep and mood are closely connected,

Sleep Health Foundation

@ Sleep deprivation, Better Health Channel

@ Q&Me survey

V/ CONCLUSION

Sutskever, “Language Models are Unsupervised Multitask Learners”

[3] Yew Lin, “ROUGE: A Package for Automatic Evaluation of summaries’,

Method for Automatic Evaluation of Machine Translation”, Proceedings of the 40th

[5]

“PyTorch Distributed: Experiences on Accelerating Data Parallel Training”,

[6] y, Amir Hertz, Amit H Bermano, “ClipCap: CLIP Prefix for Image Captioning”, arXiv:2111.09734 [cs.CV]

[7] Jesse Vig, “Visualizing Attention in Transformer Representation Models”, arXiv: 1904.02679v2 [cs.HC] 11 Apr 2019 [8] y Lei Ba, “ADAM: A METHOD FOR STOCHASTIC OPTIMIZATION”, arXiv:1412.6980v9 [cs.LG] 30 Jan 2017

[9] W S McCulloch and W Pitts, “A logical calculus of the ideas immanent in nervous activity”, en, The bulletin of mathematical biophysics, vol 5, no 4, pp 11 —

[10] D E Rumelhart, G E Hinton, and R J Williams, “Learning representations

by backpropagating errors’, en, Nature, vol 323, no 6088, pp 533—

[11] R Sathya, Annamma Abraham, “Comparison of Supervised and

Learning Algorithms for Pattern Classification”, (JARAI) International Journal of

[12]

[13] Dr S Viayarani, Ms J Ilamathi, Ms Nithya, “Preprocessing Techniques for

An Overview”, International Journal of Computer Science &

[14] ao Zhong, “Cross

Theoretical Analysis and Applications”, arXiv:2304.07288 [cs.LG]

[15] Rico Sennrich, Barry Haddow, Alexandra Birch, “Improving Neural Machine

Translation Models with Monolingual Data’, [16] , Astatistical interpretation of term specificity and its

[17] >

[18]

APPENDIX

per slide’s notes

1 | GSK Vaccines | 29 November | Luc Debruyne (President, GSK gsk_27.csv meet the 2016 Vaccines): Hello everyone, good

management | GSK Vaccines | afternoon, welcome to this "meet

29 November | meet the the management" session for

2016 management | Vaccines | hope you have been

impressed, like every morning when | drive on this site, | am still impressed after three years how vast this is, this site, and you have only seen a part of it, because we have 9,000 people here in Belgium,

it is one of the biggest sites in the world, where we do biologicals |

am sure you have seen, in your tour, the complexity of making vaccines, but | hope that you have also been impressed by the quality

of the people you have met on the tour For those who don't know

me, | am Luc Debruyne, | am 25

years working for GSK andl ama Belgian