You can use GPT-J, GPT-3 or Jurassic-1 to create human-like texts and automate SEO.
But not all that glitters is gold...what are the limitations of these language models, and how can you get the best of it?
In his speech, SEO Automation using GPT3 and Transformer based language models, he shows you some examples of how to "hack" text-to-text transformer-based models to combine human intuition with artificial intelligence and what returns you can get in terms of traffic for your website.
In this presentation I will:
- explain the transformer architecture to SEO specialists and marketers;
- show the limits of deep autoregressive language models created with this architecture;
- provide some tips on how you can use them and manage the conditions.
3. #brightonSEO
@cyberandy
"Hi, my name is Andrea
Volpini, I am the CEO of
WordLift, you can find
me on Twitter as
@cyberandy and here is a
webpage
https://wordlift.io/blog
/en/entity/andrea-volpin
i"
UNSTRUCTURED
35. “The future
of SEO”
EMBEDDINGS
DECODER
Intermediate state to
output sequence
GPT architecture
● One single model
for a variety of
tasks
● Very simple to
use (priming is
done with text)
● Primarily trained
on English but
works also on
other languages
(sort of)
36. ● 175 billion
parameters
● closed-source
● very flexible can
be applied to
different tasks
GPT-3
#brightonSEO
39. OUTPUT
INPUT “LANGUAGE
MODELS DON’T
KNOW WHAT
THEY ARE
TALKING ABOUT”
Gary Marcus
"they're approximations [...] of language use
rather than language understanding.”
47. few-shot
zero-shot
fine-tuning
Q: is Carbonara an Italian or an Indian recipe?
A: Carbonara is an Italian recipe.
Carbonara is an Italian recipe.
Duck Samosas is an Indian recipe.
Burritos are a Mexican recipe.
ONLY FEW
SAMPLES
48. few-shot
zero-shot
fine-tuning
Q: is Carbonara an Italian or an Indian recipe?
A: Carbonara is an Italian recipe.
Carbonara is an Italian recipe.
Duck Samosas is an Indian recipe.
Burritos are a Mexican recipe.
Carbonara is an Italian recipe.
Duck Samosas is an Indian recipe.
Burritos are a Mexican recipe.
Shish Tawook is an Egyptian recipe.
Sichuan Pork is a Chinese recipe.
Pizza Margherita is an Italian recipe.
Chicken tikka is an Indian recipe.
HUNDREDS
OF
SAMPLES
50. title tag generation
search intent classification
frequently asked questions
product descriptions
query: I want to subscribe to wordlift
classes:
[informational,commercial,navigational,transactional]
label: transactional
topic: vaccine hesitancy, america
title: How America’s Vaccine Hesitancy is Hurting
Our Children
Q: how many boxes of daily contacts is one year of
supply?
title: One box is a year’s supply of contacts.
Montblanc | minimalist wallet | made of top quality
| abrasion-proof fabric | 5 times thinner than a
traditional leather wallet | holds up to 15 cards |
available in any color
Montblanc is a slim, lightweight wallet that allows
you to carry more credit cards at once and safely
walk around without fear of damaging your wallet.
51. TIPS & TRICKS
1 Provide an informative
context at the beginning of
the prompt to explain the
task and to condition the
tone of voice 👇
Describe the iconic
sunglasses, from a
list of features, to a
generation Y.
1
52. TIPS & TRICKS
1 Always remember
that any form of
incorrect grammar,
spelling, and
punctuation mistakes
will badly affect
completion.
2
53. TIPS & TRICKS
1 Add as many
examples as the
prompt will allow.
Diversification is
highly needed to
mitigate overfitting.
3
54. TIPS & TRICKS
1
Always consider for
your task an alternative
0-shot implementation.
This will improve the
ability of the model to
generalize the task and
force it to rely on its
vast acquired
knowledge rather than
on few examples.
4
55. TIPS & TRICKS
1
Add always a clear
stop sequence
(usually ### or ↵↵)
to prevent semantic
contamination.
5
56. TIPS & TRICKS
generate -> curate ->
generate is your
mantra.
“To generate
high-quality writing
you need to create a
prompt that seeds it
with high-quality
writing”
6
59. GOOGLE DOESN’T LIKE AUTOMATICALLY
GENERATED CONTENT WHEN...
IT’S INTENDED TO MANIPULATE SEARCH
RANKINGS AND NOT TO HELP USERS
60. R
O
I THE ROI OF AI-GENERATED FAQ?
Source: Google Search Console data from an e-commerce in
2K
CLICKS
2
MONTHS
$ 6.7K
CPC ADV
In 2 months, we generated
2.08K additional clicks that
would have otherwise cost
our client $ 6.7K on Google
Ads.
#brightonSEO
62. 1
Content, even when
automatically generated,
needs to be updated and
maintained on regular basis.
2
We can seed questions from
both PAA or
what|how|when|why queries
that we’re already intercepting
in GSC.
3
Validation is both for humans
and for machines and helps
improving content quality.
#brightonSEO