Google's New AI Creates Summaries of Your Documents in Google Docs

Google recently announced a new model for automatically generating summaries using machine learning, released in Google Docs that you can already use.

The model will try to understand the whole document and generate a short summary of the piece—something some movie professionals clearly still can’t do.

The model needs to achieve two things to achieve that, which you will learn in the video below!

►Read the full article: https://www.louisbouchard.ai/google-docs-summary/
►Google's blog post: https://ai.googleblog.com/2022/03/auto-generated-summaries-in-google-docs.html
►GPT-3 video: https://youtu.be/gDDnTZchKec
►Attention video: https://youtu.be/QcCJJOLCeJQ
►What are RNNs?: https://youtu.be/Z0pb3LjeIZg
►My Newsletter (A new AI application explained weekly to your emails!): https://www.louisbouchard.ai/newsletter/

0:00

do you find it hard to quickly summarize

0:02

a movie you just watched or a book you

0:04

read a few weeks ago sometimes you love

0:06

a book and if you can't manage to

0:09

remember its content which i often can't

0:11

you may end up boring your friend for

0:12

talking for an hour describing the many

0:15

chapters and important parts while your

0:17

friend just want to have a quick and

0:18

concise summary this is because doing a

0:21

great summary is challenging even for us

0:23

but necessary how useful it is to be

0:26

able to quickly know what the book is

0:28

about before buying it or simply to help

0:30

you go through all your emails and

0:32

documents in seconds taking the time to

0:34

recap and summarize your work is also

0:36

very important if you want to be taken

0:38

seriously at your job and weights and

0:41

biases the sponsor of this video will

0:43

help you with that with my favorite

0:45

feature weights and biases reports it

0:48

helps you create beautiful reports about

0:50

your previous runs with your hyper

0:51

parameters and any matrix tracked in

0:54

seconds it even allows you to create

0:56

clear and dynamic plots comparing the

0:58

runs your team members can leave

1:00

comments right on the report if this

1:02

doesn't make you the coolest ml engineer

1:05

on your team i don't know what will your

1:07

teammates will understand what you are

1:09

working on in a split second it will

1:11

allow them to help you out if they can

1:13

saving valuable time and possible

1:15

misunderstandings try out weights and

1:17

biases for free with the first link

1:19

below and profit from these benefits by

1:22

stepping up your engineering game

1:27

just like for creating clear reports you

1:29

need to have a great understanding of a

1:32

book movie or any content you are trying

1:34

to summarize to do it well omit all

1:36

unnecessary information while keeping

1:39

the essential making something as

1:41

concise as possible can be really

1:42

complicated or even impossible here i

1:46

try to explain research in a few minutes

1:48

and i often can't manage to make it

1:49

shorter than 5 minutes even if it's only

1:52

a summary of a 20 page piece it requires

1:55

hours of work and fine tuning and now i

1:57

may be replaced by an ai that does that

2:00

better in milliseconds indeed google

2:03

recently announced a new model for

2:05

automatically generating summaries using

2:07

machine learning released in google docs

2:09

that you can already use the model will

2:12

try to understand the whole document and

2:14

generate a short summary of the piece

2:16

something some movie professionals

2:18

clearly still can't do the model needs

2:21

to achieve two things understand the

2:23

text in the document called natural

2:25

language understanding and generate

2:28

coherent sentences using a natural

2:30

language or in other words perform

2:33

natural language generation but how can

2:36

you achieve that you guessed it with a

2:39

lot of data and compute power luckily

2:41

enough this is google research they

2:44

trained our model to replicate our

2:46

thought process for generating summaries

2:48

using way too many documents with

2:50

manually generated summaries seeing all

2:52

these examples the model does like any

2:55

good student and ends up being able to

2:57

generate relatively good summaries for

2:59

similar documents as it has seen during

3:02

its training phase you can see why we

3:04

need good quality data here the model

3:06

will learn from them it may only be as

3:09

good as the data that was used for

3:11

training it it will be like having a

3:13

really bad coach that doesn't know

3:15

anything about basketball trying to

3:17

teach a new player how could this new

3:19

player become any good if the coach

3:21

doesn't know anything about the sport

3:24

the newcomer's talent won't be optimized

3:26

and might be wasted only because of the

3:28

poor coaching the challenge comes with

3:31

generalizing to new documents

3:33

generalizing is something even difficult

3:35

for students that only learned how to

3:37

perform the given examples but did not

3:40

understand how to apply the formulas

3:42

it's the same thing here the model faces

3:45

difficulties as it cannot remember all

3:47

documents and summaries by heart it has

3:50

to understand them or at least know

3:52

which words to put its attention on in

3:55

order to produce a summary that reflects

3:57

the document well the latter will most

4:00

likely happen as the model doesn't

4:02

really understand the document it only

4:04

understands how to perform the task

4:06

which is unfortunately still far from

4:09

human level intelligence but well enough

4:11

for such a task i just mentioned

4:13

attention well this was not a

4:16

coincidence attention may be the most

4:18

important concept behind this model

4:21

indeed just like gpt3 this new model

4:24

also uses the transformer architecture

4:26

and attention mechanisms this is where

4:28

high computation is required as you know

4:31

transformers are big and powerful

4:34

networks but most time a bit too big for

4:36

fast and efficient tools that need to be

4:38

available online in seconds transformers

4:41

computation complexity also scales with

4:43

the input size which means that the

4:45

longer the input the heavier the

4:47

computation will be causing big issues

4:50

when we want to summarize a whole book

4:52

gpt3 works well for small inputs like

4:55

question and sharing tasks but the same

4:57

architecture won't be able to process

4:59

whole books efficiently instead they had

5:01

to use some tricks in order to have a

5:03

smaller and more efficient model while

5:06

keeping high quality results this

5:08

optimization was achieved by merging

5:10

transformers with rnns or recurrent

5:13

neural networks which are two concepts i

5:16

explained in previous videos that i

5:18

highly recommend watching for a better

5:19

understanding both videos are linked in

5:22

the description below in short it will

5:24

act similar to gpt3 which you should

5:26

understand by now from my video about it

5:29

but with a smaller version of the model

5:31

iterating over and over until the model

5:34

finishes the book the transformer part

5:37

of the architecture will be responsible

5:39

for understanding a small section of the

5:41

text and producing an encoded version of

5:44

it the rnn will be responsible for

5:46

stacking and keeping this knowledge in

5:48

memory iterating through the whole book

5:51

to end up with the most concise way of

5:53

summarizing its content working together

5:56

the attention mechanism added to the

5:58

recurrent architecture will be able to

6:00

go through long documents and find the

6:02

most important features to mention in

6:04

the summary as any human will do of

6:07

course the model is not perfect since

6:09

even professional writers aren't perfect

6:11

at summarizing their work but the

6:13

results are quite impressive and

6:15

produced extremely efficiently i'd

6:17

strongly recommend trying it for

6:19

yourself in google docs to make up your

6:21

mind about it

6:22

and voila this is how google docs

6:25

automatically summarize your documents

6:27

with their new machine learning based

6:29

model i hope you enjoyed this video if

6:32

so please take a second to leave a like

6:34

and comment your thoughts on the video

6:36

and on this new model will you use it

6:39

thank you for watching until the end and

6:41

i will see you next week with another

amazing paper