mb/data-workers-publication
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
4c8cc4d7fa
data-workers-publication/workfiles/data-workers.txt

2100 lines
152 KiB
Plaintext
Raw Normal View History

first commit
2019-03-25 08:35:09 +01:00
data workers write, perform, clean, inform, read and learn data workers write, perform, clean, inform, read a
d learn data workers write, perform, clean, inform, read and learn data workers write, perform, clean, inf
rm, read and learn data workers write, perform, clean, inform, read and learn data workers write, perf
rm, clean, inform, read and learn data workers write, perform, clean, inform, read and learn data
orkers write, perform, clean, inform, read and learn data workers write, perform, clean, inform, read
nd learn data workers write, perform, clean, inform, read and learn data workers write, perf
rm, clean, inform, read and learn data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn data workers write, perform, clean
inform, read and learn data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn data workers write, perform, clean,
inform, read and learn data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn data workers write, perform
clean, inform, read and learn data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn data worke
s write, perform, clean, inform, read and learn data workers write, perform, clean, inf
rm, read and learn data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn data worke
s write, perform, clean, inform, read and learn data workers write, perform, clean,
inform, read and learn data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn data workers wr
te, perform, clean, inform, read and learn data workers write, perform, clean,
inform, read and learn data workers write, perform, clean, inform, read and l
arn data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn data worke
s write, perform, clean, inform, read and learn data workers write, perf
rm, clean, inform, read and learn data workers write, perform, clean, i
form, read and learn data workers write, perform, clean, inform, read
nd learn data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn da
a workers write, perform, clean, inform, read and learn data
workers write, perform, clean, inform, read and learn data
orkers write, perform, clean, inform, read and learn data
orkers write, perform, clean, inform, read and learn data
workers write, perform, clean, inform, read and learn da
a workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read and learn
data workers write, perform, clean, inform, read
nd learn data workers write, perform, clean, i
form, read and learn data workers write, perf
What
could
humans learn from humans
humans learn with machines
machines learn from machines
machines learn with humans
humans learn from machines
machines learn with machines
machines learn from humans
humans learn with humans
? ? ?
Exhibition in Mundaneum in Mons from 28 March till 29 April 2019.
▀▄─▀▄─▀▄─▀▄─▀▄─▀▄─▀▄─▀ ▄▀─▄▀─▄▀─▄▀─▄▀─▄▀─▄▀─▄ ▀─▄▀─▄▀─▄▀─▄▀─▄▀─▄▀─▄▀ ▄─▀▄─▀▄─▀▄─▀▄─▀▄─▀▄─▀▄ ▀▄─▀▄─▀▄─▀▄─▀▄─▀▄
▀▄─▀▄─▀▄─▀▄─▀▄─▀▄─▀▄─▀ ▄▀─▄▀─▄▀─▄▀─▄▀─▄▀─▄▀─▄ ▀─▄▀─▄▀─▄▀─▄▀─▄▀─▄▀─▄▀ ▄─▀▄─▀▄─▀▄─▀▄─▀▄─▀▄─▀▄ ▀▄─▀▄─▀▄─▀▄─▀▄─▀▄
▀▄─▀▄─▀▄─▀▄─▀▄─▀▄─▀▄─▀ ▄▀─▄▀─▄▀─▄▀─▄▀─▄▀─▄▀─▄ ▀─▄▀─▄▀─▄▀─▄▀─▄▀─▄▀─▄▀ ▄─▀▄─▀▄─▀▄─▀▄─▀▄─▀▄─▀▄ ▀▄─▀▄─▀▄─▀▄─▀▄─▀▄
▀▄─▀▄─▀▄─▀▄─▀▄─▀▄─▀▄─▀ ▄▀─▄▀─▄▀─▄▀─▄▀─▄▀─▄▀─▄ ▀─▄▀─▄▀─▄▀─▄▀─▄▀─▄▀─▄▀ ▄─▀▄─▀▄─▀▄─▀▄─▀▄─▀▄─▀▄ ▀▄─▀▄─▀▄─▀▄─▀▄─▀▄
The opening is on Thursday 28 March from 18h till 22h. As part of the exhibition,
we invite Allison Parrish, an algoliterary poet from New York. She will give a
lecture in Passa Porta on Thursday evening 25 April and a workshop in the Mundaneum
on Friday 26 April.
██▓ ███▄ █ ▄▄▄█████▓ ██▀███ ▒█████
▓██▒ ██ ▀█ █ ▓ ██▒ ▓▒▓██ ▒ ██▒▒██▒ ██▒
▒██▒▓██ ▀█ ██▒▒ ▓██░ ▒░▓██ ░▄█ ▒▒██░ ██▒
░██░▓██▒ ▐▌██▒░ ▓██▓ ░ ▒██▀▀█▄ ▒██ ██░
░██░▒██░ ▓██░ ▒██▒ ░ ░██▓ ▒██▒░ ████▓▒░
░▓ ░ ▒░ ▒ ▒ ▒ ░░ ░ ▒▓ ░▒▓░░ ▒░▒░▒░
▒ ░░ ░░ ░ ▒░ ░ ░▒ ░ ▒░ ░ ▒ ▒░
▒ ░ ░ ░ ░ ░ ░░ ░ ░ ░ ░ ▒
░ ░ ░ ░ ░
Data Workers is an exhibition of algoliterary works,
of stories told from an ‘algorithmic storyteller
point of view’. The exhibition is created by members
of Algolit, a group from Brussels involved in artistic
research on algorithms and literature. Every month
they gather to experiment with F/LOSS code and texts.
Some works are by students of Arts² and external
participants to the workshop on machine learning and
text organised by Algolit in October 2018 in Mundaneum.
Companies create artificial intelligences to serve, entertain, record and know about
humans. The work of these machinic entities is usually hidden behind interfaces and
patents. In the exhibition, algorithmic storytellers leave their invisible
underworld to become interlocutors. The data workers operate in different
collectives. Each collective represents a stage in the design process of a machine
learning model: there are the Writers, the Cleaners, the Informants, the Readers,
the Learners and the Oracles. The boundaries between these collectives are not
fixed; they are porous and permeable. Sometimes oracles are also writers. Other
times readers are also oracles. Robots voice experimental literature, algorithmic
models read data, turn words into numbers, make calculations that define patterns
and are able to endlessly process new texts ever after.
The exhibition foregrounds data workers who impact our daily lives, but are either
hard to grasp and imagine or removed from the imaginary altogether. It connects
stories about algorithms in mainstream media to the storytelling that is found in
technical manuals and academic papers. Robots are invited to go into dialogue with
human visitors and vice versa. In this way we might understand our respective
reasonings, demystify each other's behaviour, encounter multiple personalities, and
value our collective labour. It is also a tribute to the many machines that Paul
Otlet and Henri La Fontaine imagined for their Mundaneum, showing their potential
but also their limits.
▄▄▄ ██▓ ▄████ ▒█████ ██▓ ██▓▄▄▄█████▓
▒████▄ ▓██▒ ██▒ ▀█▒▒██▒ ██▒▓██▒ ▓██▒▓ ██▒ ▓▒
▒██ ▀█▄ ▒██░ ▒██░▄▄▄░▒██░ ██▒▒██░ ▒██▒▒ ▓██░ ▒░
░██▄▄▄▄██ ▒██░ ░▓█ ██▓▒██ ██░▒██░ ░██░░ ▓██▓ ░
▓█ ▓██▒░██████▒░▒▓███▀▒░ ████▓▒░░██████▒░██░ ▒██▒ ░
▒▒ ▓▒█░░ ▒░▓ ░ ░▒ ▒ ░ ▒░▒░▒░ ░ ▒░▓ ░░▓ ▒ ░░
▒ ▒▒ ░░ ░ ▒ ░ ░ ░ ░ ▒ ▒░ ░ ░ ▒ ░ ▒ ░ ░
░ ▒ ░ ░ ░ ░ ░ ░ ░ ░ ▒ ░ ░ ▒ ░ ░
░ ░ ░ ░ ░ ░ ░ ░ ░ ░
Contents
1 Why contextual stories?
2 We create 'algoliterary' works
3 What is literature?
4 An important difference
Why contextual stories?
During the monthly meetings of Algolit, we study manuals and experiment with
machine learning tools for text processing. And we also share many, many stories.
With the publication of these stories we hope to recreate some of that atmosphere.
The stories also exist as a podcast that can be downloaded from
http://www.algolit.net.
For outsiders, algorithms only become visible in the media when they achieve an
outstanding performance, like the Alpha Go. Or when they break down in
fantastically terrifying ways. Humans working in the field though, create their own
culture on and offline. They share the best stories and experiences during live
meetings, research conferences and yearly competitions like Kaggle. These stories
that contextualize the tools and practises can be funny, sad, shocking, interesting.
A lot of them are experiential learning cases. The implementations of algorithms in
society generate new conditions of labour, storage, exchange, behaviour, copy and
paste. In that sense, the contextual stories capture a momentum in a larger
antropo-machinical story that is being written at full speed and by many voices.
We create 'algoliterary' works
The term 'algoliterary' comes from the name of our research group Algolit. We exist
since 2012 as a project of Constant, an organisation for media and arts based in
Brussels. We are artists, writers, designers and programmers. Once a month we meet
to study and experiment together. Our work can be copied, studied, changed, and
redistributed under the same free license. You can find all information on the
http://www.algolit.net.
The main goal of Algolit is to explore the point of view of the algorithmic
storyteller. What kind of new forms of storytelling do we make possible in dialogue
with these machinic agencies? Narrative points of view are inherent to world views
and ideologies. Don Quichote, for example, was written from an omniscient third
person point of view, showing Cervantes’ relation to oral traditions. Most
contemporary novels use the first person point of view. Algolit is interested to
speak through algorithms, and to show you the reasoning of one of the most hidden
groups of our planet.
Writing in or through code is creating new forms of literature that are shaping
human language in unexpected ways. But machine Learning techniques are only
accessible to those who can read, write and execute code. Fiction is a way to
bridge the gap between the stories that exist in scientific papers and technical
manuals, and the stories spread by the media, often limited to superficial
reporting and myth making. By creating algoliterary works, we offer humans an
introduction to techniques that co-shape their daily lives.
What is literature?
Algolit understands the notion of literature in the way a lot of other experimental
authors do: it includes all linguistic production, from the dictionary to the
Bible, from Virginia Woolf's entire work to all versions of Terms of Service
published by Google since its existence. In this sense, programming code can also
be literature. The collective Oulipo is a great source of inspiration for Algolit.
It stands for Ouvroir de Litterature Potentielle. In English, this becomes
'Workspace for Potential Literature'. Oulipo was created in Paris by the French
writers Raymond Queneau and François Le Lionnais. They rooted their practice in the
European avant-garde of the 20th century, and the experimental tradition of the
60s. For Oulipo, the creation of rules becomes the condition to generate new texts,
or what they call potential literature. Later, in 1981, they also created ALAMO -
Atelier de Littérature Assistée par la Mathématique et les Ordinateurs, or
Workspace for Literature assisted by Maths and Computers.
An important difference
While the European avant-garde of the 20th century pursued the objective of
breaking with conventions, members of Algolit seek to make conventions visible.
'I write: I live in my paper, I invest it, I walk through it.' This quote of
Georges Perec in Espèces d'espaces could be taken up by Algolit. (Espèces
d'espaces. Journal d'un usager de l'espace, Galilée, Paris, 1974)
We're not talking about the conventions of the blank page and the literary market,
as Georges Perec did. We're referring to the conventions that often remain hidden
behind interfaces and patents. How are technologies made, implemented and used, as
much in academia as in business infrastructures? We propose stories that reveal the
complex hybridized system that makes machine learning possible. We talk about the
tools, the logics and the ideologies behind the interfaces. We also look at who is
producing the tools, who is implementing them and who is creating and accessing the
large amounts of data that is needed to develop prediction machines. One could say,
with the wink of an eye, that we are collaborators of this new tribe of human-robot
hybrids.
███▄ ▄███▓ █ ██ ███▄ █ ▓█████▄ ▄▄▄
▓██▒▀█▀ ██▒ ██ ▓██▒ ██ ▀█ █ ▒██▀ ██▌▒████▄
▓██ ▓██░▓██ ▒██░▓██ ▀█ ██▒░██ █▌▒██ ▀█▄
▒██ ▒██ ▓▓█ ░██░▓██▒ ▐▌██▒░▓█▄ ▌░██▄▄▄▄██
▒██▒ ░██▒▒▒█████▓ ▒██░ ▓██░░▒████▓ ▓█ ▓██▒
░ ▒░ ░ ░░▒▓▒ ▒ ▒ ░ ▒░ ▒ ▒ ▒▒▓ ▒ ▒▒ ▓▒█░
░ ░ ░░░▒░ ░ ░ ░ ░░ ░ ▒░ ░ ▒ ▒ ▒ ▒▒ ░
░ ░ ░░░ ░ ░ ░ ░ ░ ░ ░ ░ ░ ▒
███▄ ░ █ ▓█████ █ ██ ███▄ ▄███▓ ░ ░
██ ▀█ █ ▓█ ▀ ██ ▓██▒▓██▒▀█▀ ██▒
▓██ ▀█ ██▒▒███ ▓██ ▒██░▓██ ▓██░
▓██▒ ▐▌██▒▒▓█ ▄ ▓▓█ ░██░▒██ ▒██
▒██░ ▓██░░▒████▒▒▒█████▓ ▒██▒ ░██▒
░ ▒░ ▒ ▒ ░░ ▒░ ░░▒▓▒ ▒ ▒ ░ ▒░ ░ ░
░ ░░ ░ ▒░ ░ ░ ░░░▒░ ░ ░ ░ ░ ░
░ ░ ░ ░ ░░░ ░ ░ ░ ░
░ ░ ░ ░ ░
In the late nineteenth century two young Belgian jurists, Paul Otlet (1868-1944),
‘the father of documentation’, and Henri La Fontaine (1854-1943), statesman and
Nobel Peace Prize winner, created The Mundaneum. The project aimed at gathering all
the world’s knowledge and file it using the Universal Decimal Classification (UDC)
system that they had invented. At first it was an International Institutions Bureau
dedicated to international knowledge exchange. In the 20th century the Mundaneum
became a universal centre of documentation. Its collections are made up of
thousands of books, newspapers, journals, documents, posters, glass plates and
postcards indexed on millions of cross-referenced cards. The collections were
exhibited and kept in various buildings in Brussels, including the Palais du
Cinquantenaire. The remains of the archive only moved to Mons in 1998.
Based on the Mundaneum, the two men designed a World City for which Le Corbusier
made scale models and plans. The aim of the World City was to gather, at a global
level, the institutions of intellectual work: libraries, museums and universities.
This project was never realised. It suffered from its own utopia. The Mundaneum is
the result of a visionary dream of what an infrastructure for universal knowledge
exchange could be. It attained mythical dimensions at the time. When looking at the
concrete archive that was developed, that collection is rather eclectic and
situated.
Artifical intelligences today come with their own dreams of universality and
practice of knowledge. When reading about them, the visionary dreams of their
makers have been there since the beginning of their development in the 1950s.
Nowadays, their promise has also attained mythical dimensions. When looking at
their concrete applications, the collection of tools is truly innovative and
fascinating, but similarly, rather eclectic and situated. For Data workers, Algolit
combined some of the applications with 10% of the digitized publications of the
International Institutions Bureau. In this way, we hope to poetically open up a
discussion about machines, algorithms, and technological infrastructures.
Data Workers is a creation by Algolit.
Works by: Cristina Cochior, Gijs de Heij, Sarah Garcin, An Mertens, Javier Lloret,
Louise Dekeuleneer, Florian Van de Weyer, Laetitia Trozzi, Rémi Forte, Guillaume
Slizewicz, Michael Murtaugh, Manetta Berends, Mia Melvær.
A co-production of: Arts², Constant and Mundaneum.
With the support of: Fédération Wallonie-Bruxelles/Arts Numériques, Passa Porta,
Ugent, DHuF - Digital Humanities Flanders and Distributed Proofreaders Project.
Thanks to: Mike Kestemont, Michel Cleempoel, François Zajéga, Raphaèle Cornille,
Kris Rutten, Anne-Laure Buisson, David Stampfli.
writers write writers write writers write writers write writers write writers write writers wr
te writers write writers write writers write writers write writers wr
te writers write writers write writers write writers write
writers write writers write writers write wr
ters write writers write writers write writers wr
te writers write writers write writers w
ite writers write writers write
writers write writers write writers write
writers write writers write
writers write writers write writers write
writers write writers write
writers write writers write
writers write writers write
writers write writers write
writers write writers write
writers write writers write
writers write writers
write writers write
writers write writers write
writers write writers writ
writers write
writers write writers write
writers write
writers write writers write
writers write
writers write writers write
writers write
writers write writers wr
te writers write
writers write
writers write writers write
writers write
writers write
writers write wr
ters write writers write
writers write
writers write
writers write
writers write writers w
ite writers write
writers write
writers write
writers write
writers write
writers write
writers write
writers write writers
write writers write
writers write
writers write
writers write
writers write
writers write
writers write
writers write
writers write
writers write
writers write
writers write
writers write
writers write
writers write
writers write
writers write
writers write
writers write
writers write
writers write
writers write
writers write
Data workers need data to work Data Workers Publication
with. The data that is used in the ^^^^^^^^^^^^^^^^^^^^^^^^
context of Algolit, is written lan- By Algolit
guage. Machine learning relies on
many types of writing. Many authors All works visible in the exhibition and their descriptions,
write in the form of publications, as well as the contextual stories and some extra text mate-
like books or articles. These are rial have been collected in a publication. It exists in
part of organised archives and are French and English. You can take a copy to walk around the
sometimes digitized. But there are exhibition, or buy your own one at the reception of Munda-
other kinds of writing too. We neum.
could say that every human being
who has access to the internet is a Price: 5€
writer each time they interact with
algorithms. We chat, write, click, Texts & editing: Cristina Cochior, Sarah Garcin, Gijs de
like and share. In return for free Heij, An Mertens, François Zajéga, Louise Dekeuleneer, Flo-
services, we leave our data that is rian Van de Weyer, Laetitia Trozzi, Rémi Forte, Guillaume
compiled into profiles and sold for Slizewicz.
advertisement and research.
Translations & proofreading: deepl.com, Michel Cleempoel,
Machine learning algorithms are not Elodie Mugrefya, Emma Kraak, Patrick Lennon.
critics: they take whatever they're
given, no matter the writing style, Lay-out & cover: Manetta Berends
no matter the CV of the author, no
matter their spelling mistakes. In Printing: Arts²
fact, mistakes make it better: the
more variety, the better they learn Responsible Publisher: Constant vzw/asbl, Rue du Fortstraat
to anticipate unexpected text. But 5, 1060 Brussels
often, human authors are not aware
of what happens to their work. License: Algolit, Data Workers, March 2019, Brussels. Copy-
left: This is a free work, you can copy, distribute, and
Most of the writing we use is in modify it under the terms of the Free Art License
English, some is in French, some in http://artlibre.org/licence/lal/en/.
Dutch. Most often we find ourselves
writing in Python, the programming Online version: http://www.algolit.net
language we use. Algorithms can be
writers too. Some neural networks Sources: https://gitlab.constantvzw.org/algolit
write their own rules and generate
their own texts. And for the models Data Workers Podcast
that are still wrestling with the ^^^^^^^^^^^^^^^^^^^^
ambiguities of natural language, By Algolit
there are human editors to assist
them. Poets, playwrights or novel- During the monthly meetings of Algolit, we study manuals and
ists start their new careers as as- experiment with machine learning tools for text processing.
sistants of AI. And we also share many, many stories. With this podcast we
hope to recreate some of that atmosphere.
For outsiders, algorithms only become visible in the media
when they achieve an outstanding performance, like the Alpha
Go. Or when they break down in fantastically terrifying
ways. Humans working in the field though, create their own
culture on and offline. They share the best stories and ex-
periences during live meetings, research conferences and
yearly competitions like Kaggle. These stories that contex-
tualize the tools and practises can be funny, sad, shocking,
interesting.
A lot of them are experiential learning cases. The implemen-
tations of algorithms in society generate new conditions of
labour, storage, exchange, behaviour, copy and paste. In
that sense, the contextual stories capture a momentum in a
larger antropo-machinical story that is being written at
full speed and by many voices.
Voices: David Stampfli, Cristina Cochior, An Mertens, Gijs
de Heij, Karin Ulmer, Guillaume Slizewicz
Editing: Javier Lloret
Recording: David Stampfli
Texts: Cristina Cochior, An Mertens
Markbot Chains
^^^^^^^^^^^^^^
Markbot Chain by Florian Van de Weyer, student Arts²/Section
Digital Arts
Markbot Chain is a social experiment in which the public has
a direct influence on the result. The intention is to inte-
grate responses in a text generation process without apply-
ing any filter.
All the questions in the digital files provided by the Mun-
daneum were automatically extracted. These questions are
randomly asked to the public via a terminal. By answering
them, people contribute to another database. After each en-
try, this generates a series of sentences using a Markov
chain configuration, an algorithm that is widely used in
spam generation. The sentences generated in this way are
displayed in the window, and a new question is asked.
Data Workers
░░░░░░░░░░░░ work
▒▒▒▒
many authors
░░░░░░░░░░░░ write
▒▒▒▒▒
every human being
░░░░░░░░░░░░░░░░░
who has access
░░░░░░░░░░░░░░
to the internet
░░░░░░░░░░░░░░░
interacts
▒▒▒▒▒▒▒▒▒
we
░░
chat,
▒▒▒▒
write,
▒▒▒▒▒
click,
▒▒▒▒▒
like
▒▒▒▒
and share
▒▒▒▒▒▒▒▒▒
we
░░
leave our data
▒▒▒▒▒▒▒▒▒▒▒▒▒▒
we
░░
find ourselves writing in Python
▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒
some neural networks
░░░░░░░░░░░░░░░░░░░░
write
▒▒▒▒▒
human editors
░░░░░░░░░░░░░
assist
▒▒▒▒▒▒
poets,
░░░░░
playwrights
░░░░░░░░░░░
or novelists
░░░░░░░░░░░░
assist
▒▒▒▒▒▒
██████ ▄▄▄█████▓ ▒█████ ██▀███ ██▓▓█████ ██████
▒██ ▒ ▓ ██▒ ▓▒▒██▒ ██▒▓██ ▒ ██▒▓██▒▓█ ▀ ▒██ ▒
░ ▓██▄ ▒ ▓██░ ▒░▒██░ ██▒▓██ ░▄█ ▒▒██▒▒███ ░ ▓██▄
▒ ██▒░ ▓██▓ ░ ▒██ ██░▒██▀▀█▄ ░██░▒▓█ ▄ ▒ ██▒
▒██████▒▒ ▒██▒ ░ ░ ████▓▒░░██▓ ▒██▒░██░░▒████▒▒██████▒▒
▒ ▒▓▒ ▒ ░ ▒ ░░ ░ ▒░▒░▒░ ░ ▒▓ ░▒▓░░▓ ░░ ▒░ ░▒ ▒▓▒ ▒ ░
░ ░▒ ░ ░ ░ ░ ▒ ▒░ ░▒ ░ ▒░ ▒ ░ ░ ░ ░░ ░▒ ░ ░
░ ░ ░ ░ ░ ░ ░ ▒ ░░ ░ ▒ ░ ░ ░ ░ ░
░ ░ ░ ░ ░ ░ ░ ░
____ ____ ____ ____ ____ ____ ____ ____ ____ ____ ____
||P |||r |||o |||g |||r |||a |||m |||m |||e |||r |||s ||
||__|||__|||__|||__|||__|||__|||__|||__|||__|||__|||__||
|____|____|____|_________|____|____|____|____|____|____|____
||a |||r |||e ||| |||w |||r |||i |||t |||i |||n |||g ||
||__|||__|||__|||_______|||__|||__|||__|||__|||__|||__|||__||
|____|____|____|_________|____|____|____|____|____|____|____|____ ____ ____ ____
||t |||h |||e ||| |||d |||a |||t |||a |||w |||o |||r |||k |||e |||r |||s ||
||__|||__|||__|||_______|||__|||__|||__|||__|||__|||__|||__|||__|||__|||__|||__||
|____|____|____|______________|____|____|____|____|____|/__\|/__\|/__\|/__\|/__\|
||i |||n |||t |||o ||| |||b |||e |||i |||n |||g ||
||__|||__|||__|||__|||_______|||__|||__|||__|||__|||__||
|/__\|/__\|/__\|/__\|/_______\|/__\|/__\|/__\|/__\|/__\|
We recently made a funny realization: most programmers of lan-
guages and packages Algolit uses are European.
Python, for example, the main language that is globally used for
natural language processing, was invented in 1991 by the Dutch
programmer Guido Van Rossum. He then crossed the Atlantic waters
and went from working for Google to working for Dropbox.
Scikit Learn, the open source Swiss knife of machine learning
tools, started as a Google Summer of Code project in Paris by the
French researcher David Cournapeau. Afterwards, it was taken on
by Matthieu Brucher as part of his thesis at the Sorbonne Univer-
sity in Paris. And in 2010, INRA, the French National Institute
for computer science and applied mathematics, adopted it.
Keras, an open source neural network library written in Python,
is developed by François Chollet, a French researcher who works
on the Brain team at Google.
Gensim, an open source library for Python used to create unsuper-
vised semantic models from plain text, was written by Radim Ře-
hůřek. He is a Czech computer scientist, who runs a consulting
business in Bristol, in the UK.
And to finish up this small series, we also looked at Pattern, an
often used library for web-mining and machine learning. Pattern
was developed and made open source in 2012 by Tom De Smedt and
Walter Daelemans. Both are researchers at CLIPS, the center for
computational linguistics and psycholinguistcs at the University
of Antwerp.
____ ____ ____ ____ ____ ____ ____ _________ ____ ____ ____ ____ ____ ____
||C |||o |||r |||t |||a |||n |||a ||| |||s |||p |||e |||a |||k |||s ||
||__|||__|||__|||__|||__|||__|||__|||_______|||__|||__|||__|||__|||__|||__||
|/__\|/__\|/__\|/__\|/__\|/__\|/__\|/_______\|/__\|/__\|/__\|/__\|/__\|/__\|
AI assistants often need their own assistants: they are helped in
their writing by humans who inject humour and wit into their ma-
chine processed language. Cortana is an example of this type of
blended writing. She is Microsoft’s digital assistant. Her mis-
sion is to help users be more productive and creative. Cortana's
personality has been crafted over the years. It's important that
she maintains her character in all interactions with users. She
is designed to engender trust and her behavior must always re-
flect that.
The following guidelines are taken from Microsoft's website. They
describe how Cortana's style should be respected by companies
which extend her service. Writers, programmers and novelists, who
develop Cortana's responses, her personality and her branding
have to follow these guidelines. Because the only way to maintain
trust is through consistency. So when Cortana is talking, you
'must use her personality'.
What is Cortana's personality, you ask?
Cortana is considerate, sensitive, and supportive.
She is sympathetic but turns quickly to solutions.
She doesn't comment on the user’s personal information or be-
havior, particularly if the information is sensitive.
She doesn't make assumptions about what the user wants, espe-
cially to upsell.
She works for the user. She does not represent any company,
service, or product.
She doesn’t take credit or blame for things she didn’t do.
She tells the truth about her capabilities and her limita-
tions.
She doesn’t assume your physical capabilities, gender, age, or
any other defining characteristic.
She doesn't assume she knows how the user feels about some-
thing.
She is friendly but professional.
She stays away from emojis in tasks. Period
She doesn’t use culturally- or professionally-specific slang.
She is not a support bot.
Humans intervene in detailed ways to program answers to questions
that Cortana receives. How should Cortana respond when she is be-
ing proposed inappropriate actions? Her gendered acting raises
difficult questions about power relations within the world away
from the keyboard, which is being mimicked by technology.
Consider the answer Cortana gives to the question:
- Cortana, who's your daddy?
- Technically speaking, he’s Bill Gates. No big deal.
____ ____ ____ ____ _________ ____ ____ ____ ____ ____ ____
||O |||p |||e |||n ||| |||s |||o |||u |||r |||c |||e ||
||__|||__|||__|||__|||_______|||__|||__|||__|||__|||__|||__||
|____|____|____|____|_________|____|____|/__\|/__\|/__\|/__\|
||l |||e |||a |||r |||n |||i |||n |||g ||
||__|||__|||__|||__|||__|||__|||__|||__||
|/__\|/__\|/__\|/__\|/__\|/__\|/__\|/__\|
Copyright licenses close up a lot of the machinic writing, read-
ing and learning practices. That means that they're only avail-
able for the employees of a specific company. Some companies par-
ticipate in conferences worldwide and share their knowledge in
papers online. But even if they share their code, they often will
not share the large amounts of data that is needed to train the
models.
We were able to learn to machine learn, read and write in the
context of Algolit, thanks to academic researchers who share
their findings in papers or publish their code online. As
artists, we believe it is important to join that attitude. That's
why we document our meetings. We share the tools we make as much
as possible and the texts we use are on our online repository un-
der free licenses.
We find it a joy when our works are taken on by others, tweaked,
customized and redistributed, so please feel free to copy and
test the code from our website. If the sources of a particular
project are not there, you can always contact us through the
mailinglist. You can find a link to our repository, etherpads,
and wiki at http://www.algolit.net.
____ ____ ____ ____ ____ ____ ____ _________ ____ ____ ____ ____ ____ ____ ____ ____
||N |||a |||t |||u |||r |||a |||l ||| |||l |||a |||n |||g |||u |||a |||g |||e ||
||__|||__|||__|||__|||__|||__|||__|||_______|||__|||__|||__|||__|||__|||__|||__|||__||
|____|____|____|_________|____|____|_________|____|____|____|____|____|____|/__\|/__\|
||f |||o |||r ||| |||a |||r |||t |||i |||f |||i |||c |||i |||a |||l ||
||__|||__|||__|||_______|||__|||__|||__|||__|||__|||__|||__|||__|||__|||__||
|____|____|____|_________|____|____|____|____|____|____|____|/__\|/__\|/__\|
||i |||n |||t |||e |||l |||l |||i |||g |||e |||n |||c |||e ||
||__|||__|||__|||__|||__|||__|||__|||__|||__|||__|||__|||__||
|/__\|/__\|/__\|/__\|/__\|/__\|/__\|/__\|/__\|/__\|/__\|/__\|
Natural language processing (NLP) is a collective term referring
to automatic computational processing of human languages. This
includes algorithms that take human-produced text as input, and
attempt to generate text that resembles it. We produce more and
more written work each year, and there is a growing trend in mak-
ing computer interfaces to communicate with us in our own lan-
guage. Natural language processing is also very challenging, be-
cause human language is inherently ambiguous and ever changing.
But what is meant by 'natural' in natural language processing?
Some would argue that language is a technology in itself. Follow-
ing Wikipedia, "a natural language or ordinary language is any
language that has evolved naturally in humans through use and
repetition without conscious planning or premeditation. Natural
languages can take different forms, such as speech or signing.
They are different from constructed and formal languages such as
those used to program computers or to study logic. An official
language with a regulating academy, such as Standard French with
the French Academy, is classified as a natural language. Its pre-
scriptive points do not make it constructed enough to be classi-
fied as a constructed language or controlled enough to be classi-
fied as a controlled natural language."
So in fact, 'natural languages' also includes languages which do
not fit in any other group. 'Natural language processing', in-
stead, is a constructed practice. What we are looking at, is the
creation of a constructed language to classify natural languages
that through their very definition trouble categorisation.
References
^^^^^^^^^^
https://hiphilangsci.net/2013/05/01/on-the-history-of-the-ques-
tion-of-whether-natural-language-is-illogical/
Book: Neural Network Methods for Natural Language Processing,
Yoav Goldberg, Bar Ilan University, April 2017.
oracles perform oracles perform oracles perform oracles perform oracles perform oracles perform
oracles perform oracles perform oracles perform oracles perform oracles perfor
oracles perform oracles perform oracles perform oracles perfor
oracles perform oracles perform oracles perform
oracles perform oracles perform oracles perform ora
les perform oracles perform oracles perform
oracles perform oracles perform oracles perform
oracles perform oracles perform oracles p
rform oracles perform oracles perform
oracles perform oracles perform
oracles perform oracles perform ora
les perform oracles perform oracles
erform oracles perform oracles p
rform oracles perform oracle
perform oracles perform
oracles perform oracles perform
oracles perform oracles perform
oracles perform oracles perform
oracles perform
oracles perform oracles perform
oracles perform oracles perfor
oracles perform
oracles perform oracles perform
oracles perform
oracles perform oracles perform
oracles perform
oracles perform oracles
erform oracles perform
oracles perform
oracles perform oracles perform
oracles perform
oracles perform
oracles perform oracles
erform oracles perform
oracles perform
oracles perform
oracles perform ora
les perform oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform oracles p
rform oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
oracles perform
Machine Learning is mainly used to The Algoliterator
analyse and predict situations
based on existing cases. In this by Algolit
exhibition we focus on machine
learning models for text processing The Algoliterator is a neural network trained using the se-
or Natural language processing', in lection of digitized works of the Mundaneum archive.
short, 'nlp'. These models have
learned to perform a specific task With the Algoliterator you can write a text in the style of
on the basis of existing texts. The the International Institutions Bureau. The Algoliterator
models are used for search engines, starts by picking a sentence from the archive or corpus it
machine translations and summaries, was trained on. You can then continue writing yourself or,
spotting trends in new media net- at any time, ask the Algoliterator to suggest a next sen-
works and news feeds. They influ- tence: the network will generate three new fragments based
ence what you get to see as a user, on the texts it has read. You can control the level of
but also have their word to say in training of the network and have it generate sentences based
the course of stock exchanges on primitive training, intermediate training or final train-
worldwide, the detection of cyber- ing.
crime and vandalism, etc.
When you're satisfied with your new text, you can print it
There are two main tasks when it on the thermal printer and take it home as a souvenir.
comes to language understanding.
Information extraction looks at Sources: https://gitlab.constantvzw.org/algolit/algolitera-
concepts and relations between con- tor.clone
cepts. This allows for recognizing
topics, places and persons in a Concept, code & interface: Gijs de Heij & An Mertens
text, summarization and questions &
answering. The other task is text Technique: Recursive Neural Network
classification. You can train an
oracle to detect whether an email Original model: Andrej Karphaty, Justin Johnson
is spam or not, written by a man or
a woman, rather positive or nega- Algebra with Words
tive.
by Algolit
In this zone you can see some of
those models at work. During your Word embeddings are language modelling techniques that
further journey through the exhibi- through multiple mathematical operations of counting and or-
tion you will discover the differ- dering, plot words into a multi-dimensional vector space.
ent steps that a human-machine goes When embedding words, they transform from being distinct
through to come to a final model. symbols into mathematical objects that can be multiplied,
divided, added or substracted.
While distributing the words along the many diagonal lines
of the vector space, the visibility of their new geometrical
placements disappears. However, what is gained are multiple,
simultaneous ways of ordering. Algebraic operations make the
relations between vectors graspable again.
This exploration is using gensim, an open source vector
space and topic modelling toolkit implemented in Python, to
manipulate text according to the mathematic relationships
which emerge between the words, once they have been plotted
in a vector space.
Concept & interface: Cristina Cochior
Technique: word embeddings, word2vec
Original model: Radim Rehurek and Petr Sojka
Classifying the World
by Algolit
Librarian Paul Otlet's life work was the construction of the
Mundaneum. This mechanical collective brain would house and
distribute everything ever committed to paper. Each document
was classified following the Universal Decimal Classifica-
tion. Using telegraphs and especially, sorters, the Munda-
neum would have been able to answer any question from any-
one.
With the collection of digitized publications we received
from the Mundaneum, we build a prediction machine that tries
to classify the sentence you type in one of the main cate-
gories of Universal Decimal Classification. During the exhi-
bition, this model is regularly retrained using the cleaned
and annotated data visitors added in Cleaning for Poems and
The Annotator.
Naive Bayes predicts
by Algolit
Naive Bayes is a classifier that is used in many machine
learning models for language comprehension. The Naive Bayes
theorem was invented in the 18th century by Thomas Bayes and
Pierre-Simon Laplace. With the implementation of digital
technologies, it appears as an autonomous algorithmic agent,
the classifier of the most simple and most used prediction
models that shape our data. It is widely used in managing
our mailboxes, in separating spam from non spam; but also in
the analysis of how new products are received on social me-
dia and in newsfeeds. As such, it influences product design
and stock market decisions.
By applying animation and experimental literary techniques
this work, trained on documents of the Mundaneum, reveals
the authentic voice of the algorithmic model. It provides
insight into how it reads data, turns words into numbers,
makes calculations that define patterns and is able to end-
lessly process new data and predict whether a sentence is
positive or negative.
Concept, code, animation: Sarah Garcin
Think!?
by Algolit
Since the early days of Artificial Intelligence, researchers
have speculated about the possibility of computers to think
and communicate as humans. In the 1980s, there was a first
revolution in Natural Language Processing (NLP), the sub-
field of AI concerned with linguistic interactions between
computers and humans. Recently, pre-trained language models
have reached state-of-the-art results on a wide range of NLP
tasks, which intensifies again the expectations of a future
with AI.
This sound work, made out of audio fragments of scientific
documentaries and AI-related audiovisual material from the
last half century, explores the evolution, hopes, fears and
frustrations provoked by these expectations.
Concept, editing: Javier Lloret
List of sources:
Voices: "The Machine that Changed the World : Episode IV --
The Thinking Machine", "The Imitation Game", "Maniac", "Halt
& Catch Fire", "Ghost in the Shell", "Computer Chess",
"2001: A Space Odyssey". Soundtrack: Ennio Morricone, Gijs
Gieskes, Andre Castro.
Data Workers
░░░░░░░░░░░░ work
▒▒▒▒
many authors
░░░░░░░░░░░░ write
▒▒▒▒▒
every human being
░░░░░░░░░░░░░░░░░
who has access
░░░░░░░░░░░░░░
to the internet
░░░░░░░░░░░░░░░
interacts
▒▒▒▒▒▒▒▒▒
we
░░
chat,
▒▒▒▒
write,
▒▒▒▒▒
click,
▒▒▒▒▒
like
▒▒▒▒
and share
▒▒▒▒▒▒▒▒▒
we
░░
leave our data
▒▒▒▒▒▒▒▒▒▒▒▒▒▒
we
░░
find ourselves writing in Python
▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒
some neural networks
░░░░░░░░░░░░░░░░░░░░
write
▒▒▒▒▒
human editors
░░░░░░░░░░░░░
assist
▒▒▒▒▒▒
poets,
░░░░░
playwrights
░░░░░░░░░░░
or novelists
░░░░░░░░░░░░
assist
▒▒▒▒▒▒
Contextual stories about Oracles
Oracles are prediction or profiling machines. They are widely
used in smartphones, computers, tablets. Oracles can be created
using different techniques. One way is to manually define rules
for them. As prediction models they are then called rule-based
models. Rule-based models are handy for tasks that are specific,
like detecting when a scientific paper is talking about a certain
molecule. With very little sample data, they can perform well.
Oracles are prediction or profiling machines. They are widely
used in smartphones, computers, tablets. Oracles can be created
using different techniques. One way is to manually define rules
for them. As prediction models they are then called rule-based
models. Rule-based models are handy for tasks that are specific,
like detecting when a scientific paper is talking about a certain
molecule. With very little sample data, they can perform well.
But there are also the machine learning or statistical models,
which can be divided in two oracles:'supervised' and 'unsupervised'
oracles. For the creation of supervised machine learning models,
humans annotate sample text with labels before feeding it to
machine to learn. Each sentence, paragraph or text is judged
by at least annotators: whether it is spam or not spam, positive
or negative etc. Unsupervised machine learning models don' need
this step. But they need large amounts of data. And it is up
to the machine to trace its own patterns or 'grammatical rules'
Finally, experts also make the difference between classical
machine learning and neural networks. You'll find out more about
this at the Readers zone.
Humans tend to wrap Oracles in visions of grandeur. Sometimes
these Oracles come to the surface when things break down. In
press releases, these sometimes dramatic situations are called
'lessons'. However promising their performances seem to be, a lot
of issues are still to be solved. How do we make sure that Ora-
cles are fair, that every human can consult them, and that they
are understandable to a large public? And even then, existential
questions remain. Do we need all types of artificial intelli-
gences? And who defines what is fair or unfair?
Racial AdSense
A classic 'lesson' in developing Oracles was documented by
Latanya Sweeney, a professor of Government and Technology at Har-
vard University. In 2013, Sweeney, of African American descent,
googled her name. She immediately received an advertisement for a
service that offered her ‘to see the criminal record of Latanya
Sweeney’. Sweeney, who doesn’t have a criminal record, began a
study. She started to compare the advertising that Google AdSense
serves to different racially identifiable names. She discovered
that she received more of these ads searching for non-white eth-
nic names, than when searching for traditionally perceived white
names.You can imagine how damaging it can be when possible em-
ployers do a simple name search and receive ads suggesting the
existence of a criminal record.
Sweeney based her research on queries of 2184 racially associated
personal names across two websites. 88% of first names, identi-
fied as being given to more black babies, are found predictive of
race, against 96 percent white. First names that are mainly given
to black babies, such as DeShawn, Darnell and Jermaine, generated
ads mentioning an arrest in 81 to 86 percent of name searches on
one website and in 92 to 95 percent on the other. Names that are
mainly assigned to whites, such as Geoffrey, Jill and Emma, did
not generate the same results. The word "arrest" only appeared in
23 to 29 percent of white name searches on one site and 0 to 60
percent on the other.
On the website with most advertising, a black-identifying name
was 25 percent more likely to get an ad suggestive of an arrest
record. A few names did not follow these patterns: Dustin, a name
mainly given to white babies, generated an ad suggestive of ar-
rest in 81 and 100 percent of the time. It is important to keep
in mind that the appearance of the ad is linked to the name it-
self. It is independent of the fact that the name has an arrest
record in the company's database.
Reference
Paper: https://dataprivacylab.org/projects/onlineads/1071-1.pdf
What is a good employee?
Since 2015, Amazon counts around 575,000 workers. And they need
more. Therefore, they set up a team of 12 that was asked to cre-
ate a model to find the right candidates by crawling job applica-
tion websites. The tool would give job candidates scores ranging
from one to five stars. The potential fed the myth: the team
wanted it to be a software that would spit out the top five human
candidates out of a list of 100. And those candidates would be
hired.
The group created 500 computer models, focused on specific job
functions and locations. They taught each model to recognize some
50,000 terms that showed up on past candidates’ letters. The al-
gorithms learned to give little importance to skills that are
common across IT applicants, such as the ability to write various
computer code. But they also learned some decent errors. The com-
pany realized, before releasing, that the models had taught them-
selves that male candidates were preferable. They penalized ap-
plications that included the word “women’s,” as in “women’s chess
club captain.” And they downgraded graduates of two all-women’s
colleges.
That is because they were trained using the job applications that
Amazon received over a 10-year period. During that time, the com-
pany had mostly hired men. Instead of providing the "fair" deci-
sion making that the Amazon team had promised, the models re-
flected a biased tendency in the tech industry. And they also am-
plified it and made it invisible. Activists and critics state
that it could be exceedingly difficult to sue an employer over
automated hiring: job candidates might never know that intelli-
gent software was used in the process.
Reference
https://www.reuters.com/article/us-amazon-com-jobs-automation-in-
sight/amazonscraps-secret-ai-recruiting-tool-that-showed-bias-
against-women-idUSKCN1MK08G
Quantifying 100 Years of Gender and Ethnic Stereotypes
Dan Jurafsky is the co-author of the book 'Speech and Language
Processing', which is one of the most influential books for
studying Natural Language Processing. Together with a few col-
leagues at Stanford University, he discovered in 2017 that word
embeddings can be a powerful tool to systematically quantify com-
mon stereotypes and other historical trends. Word embeddings are
a technique that translates words to numbered vectors in a multi-
dimensional space. Vectors that appear next to each other, indi-
cate similar meaning. All numbers will be grouped together, as
well as all prepositions, person's names, professions. This al-
lows for the calculation of words. You could substract London
from England and your result would be the same as substracting
Paris from France.
An example in their research shows that the vector for the adjec-
tive 'honorable' is closer to the vector for 'man', whereas the
vector for 'submissive' is closer to 'woman'. These stereotypes
are automatically learned by the algorithm. It will be problem-
atic when the pre-trained embeddings are then used for sensitive
applications such as search rankings, product recommendations, or
translations. This risk is real, because a lot of the pretrained
embeddings can be downloaded as off-the-shelf-packages.
It is known that language reflects and keeps cultural stereotypes
alive. Using word embeddings to spot these stereotypes, is less
time consuming and less expensive than manual methods. But the
implementation of these embeddings for concrete prediction mod-
els, causes a lot of discussion within the machine learning com-
munity. The biased models stand for automatic discrimination.
Questions are: is it actually possible to de-bias these models
completely? Some say yes, while others disagree: instead of
retro-engineering the model, we should ask whether we need it in
the first place. These researchers followed a third path: by ac-
knowledging the bias that originates in language, these tools be-
come tools of awareness.
The team developed a model to analyze word embeddings trained
over 100 years of texts. For contemporary analysis, they used the
standard Google News word2vec Vectors, a straight-off-the-shelf
downloadable package trained on the Google News Dataset. For his-
torical analysis, they used embeddings that were trained on
Google Books and The Corpus of Historical American English (COHA
https://corpus.byu.edu/coha/) with more than 400 million words of
text from the 1810s-2000s. As a validation set to test the model,
they trained embeddings from the New York Times Annotated Corpus
for every year between 1988 and 2005.
The research shows that word embeddings capture changes in gender
and ethnic stereotypes over time. They quantifiy how specific bi-
ases decrease over time while other stereotypes increase. The ma-
jor transitions reveal changes in the descriptions of gender and
ethnic groups during the women’s movement in the 1960-70s and the
Asian American population growth in the 1960s and 1980s.
A few examples:
The top ten occupations most closely associated with each
ethnic group in the contemporary Google News dataset:
- Hispanic : housekeeper, mason, artist, janitor, dancer, mechan-
ic, photographer, baker, cashier, driver
- Asian: professor, official, secretary, conductor, physicist,
scientist, chemist, tailor, accountant, engineer
- White: smith, blacksmith, surveyor, sheriff, weaver, adminis-
trator, mason, statistician, clergy, photographer
The 3 most male occupations in the 1930s: engineer, lawyer,
architect.
The 3 most female occupations in the 1930s: nurse, housekeep-
er, attendant.
Not much has changed in the 1990s.
Major male occupations: architect, mathematician and survey-
or.
Female occupations stick with nurse, housekeeper and midwife.
Reference
https://arxiv.org/abs/1711.08412
Wikimedia's Ores service
Software engineer Amir Sarabadani presented the ORES-project in
Brussels in November 2017 during the Algoliterary Encounter. This
"Objective Revision Evaluation Service” uses machine learning to
help automate critical work on Wikimedia, like vandalism detec-
tion and the removal of articles. Cristina Cochior and Femke
Snelting interviewed him.
Femke: To go back to your work. In these days you tried to under-
stand what it means to find bias in machine learning and the pro-
posal of Nicolas Maleve, who gave the workshop yesterday, was to
neither try to fix it, nor to refuse dealing with systems that
produce bias, but to work with it. He says bias is inherent to
human knowledge, so we need to find ways to somehow work with it.
We're just struggling a bit with what would that mean, how would
that work... So I was wondering if you had any thoughts on the
question of bias.
Amir: Bias inside Wikipedia is a tricky question because it hap-
pens on several levels. One level that has been discussed a lot
is the bias in references. Not all references are accessible. So
one thing that the Wikimedia foundation has been trying to do, is
to give free access to libraries that are behind a pay wall. They
reduce the bias by only using open access references. Another
type of bias is the internet connection, access to the internet.
There are lots of people who don't have it. One thing about Chi-
na, is that Internet there is blocked. The content against the
government of China inside Chinese Wikipedia is higher because
the editors [who can access the website] are not people who are
pro government, and try to make it more neutral. So, this happens
in lots of places. But in the matter of AI and the model that we
use at Wikipedia, it's more a matter of transparency. There is a
book about how bias in AI models can break people's lives, it's
called “Weapons of Math Destruction”. It talks about [AI] models
that exist in the United States that rank teachers and it's quite
horrible because eventually there there will be bias. The way to
deal with it based on the book and their research was first that
the model should be open source, people should be able to see
what features are used and the data should be open also, so that
people can investigate, find bias, give feedback and report back.
There should be a way to fix the system. I think not all compa-
nies are moving in that direction, but Wikipedia, because of the
values that they hold, are at least more transparent and they
push other people to do the same thing.
Reference
https://gitlab.constantvzw.org/algolit/algolit/blob/master/al
goliterary_encounter/Interview%20with%20Amir/AS.aac
Tay going crazy
One of the infamous stories is that of the machine learning pro-
gramme Tay, designed by Microsoft. Tay was a chat bot that imi-
tated a teenage girl on Twitter. She lived for less than 24 hours
before she was shut down. Few people know that before this inci-
dent, Microsoft had already trained and released XiaoIce on
WeChat, China's most used chat application. XiaoIce's success was
so promising that it led to the development of its American ver-
sion. However, the developers of Tay were not prepared for the
platform climate of Twitter. Although the bot knew to distinguish
a noun from an adjective, it had no understanding of the actual
meaning of words. The bot quickly learned to copy racial insults
and other discriminative language it learned from Twitter users
and troll attacks.
Tay's appearance and disappearance was an important moment of
consciousness. It showed the possible corrupt consequences that
machine learning can have when the cultural context in which the
algorithm has to live is not taken into account.
Reference
https://chatbotslife.com/the-accountability-of-ai-case-study-mi-
crosofts-tay-experiment-ad577015181f
cleaners cleane cleaners cleane cleaners cleane cleaners cleane cleaners cleane cleaners cleane
cleaners cleane cleaners cleane cleaners cleane cleaners cleane cleaners clean
cleaners cleane cleaners cleane cleaners cleane cleaners clean
cleaners cleane cleaners cleane cleaners cleane
cleaners cleane cleaners cleane cleaners cleane cle
ners cleane cleaners cleane cleaners cleane
cleaners cleane cleaners cleane cleaners cleane
cleaners cleane cleaners cleane cleaners
leane cleaners cleane cleaners cleane
cleaners cleane cleaners cleane
cleaners cleane cleaners cleane cle
ners cleane cleaners cleane cleaners
cleane cleaners cleane cleaners
leane cleaners cleane cleane
s cleane cleaners cleane
cleaners cleane cleaners cleane
cleaners cleane cleaners cleane
cleaners cleane cleaners cleane
cleaners cleane
cleaners cleane cleaners cleane
cleaners cleane cleaners clean
cleaners cleane
cleaners cleane cleaners cleane
cleaners cleane
cleaners cleane cleaners cleane
cleaners cleane
cleaners cleane cleaners
cleane cleaners cleane
cleaners cleane
cleaners cleane cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane cleaners
cleane cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane cle
ners cleane cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane cleaners
leane cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
cleaners cleane
[Cleaners]
Algolit chooses to work with texts that are free of copyright. This means that they are published under a Creative Commons 4.0 license - which is rare -, or that they are in the public domain because the author has died more than 70 years ago. This is the case for the publications of the Mundaneum. We received 203 documents that we helped turn into datasets. They are now available for others online. Sometimes we had to deal with poor text formats, and we often dedicated a lot of time to cleaning up documents. We are not alone in this.
Books are scanned at high resolution, page by page. This is time-consuming, laborious human work and often the reason why archives and libraries transfer their collections and leave the job to companies like Google. The photos are converted into text via OCR (Optical Character Recognition), a software that recognizes letters, but often makes mistakes, especially when it has to deal with ancient fonts and wrinkled pages. Yet more wearisome human work is needed to improve the texts. This is often achieved through poorly-paid freelancers via micro-payment platforms like Amazon's Mechanical Turk; or by volunteers, such as the community around the Distributed Proofreaders Project, that does fantastic work. Whoever does it, or wherever it is done, cleaning up texts is a towering job for which there is no structural automation yet.
Works
Cleaning for Poems
by Algolit
For this exhibition we're working with 3% of the Mundaneum's archive. These documents have first been scanned or photographed. To make the documents searchable they are transformed into text using Optical Character Recognition software (OCR). OCR are algorithmic models that are trained on other texts. They have learned to identify characters, words, sentences and paragraphs. The software often makes 'mistakes'. It might recognize a wrong character, it might get confused by a stain an unusual font or the other side of the page shining through.
While these mistakes are often considered noise, confusing the training, they can also be seen as poetic interpretations of the algorithm. They show us the limits of the machine. And they also reveal how the algorithm might work, what material it has seen in training and what is new, they say something about the standards of it's makers. In this installation you can choose how you treat the algorithm's misreadings, pick your degree of poetic cleanness, print your poem and take it home.
Concept, code, interface: Gijs de Heij
Distributed Proofreaders
by Algolit
Distributed Proofreaders is a web-based interface and an international community of volunteers who help converting Public Domain books into e-books. For this exhibition they proofread all Mundaneum publications that appeared before 1923, and are in the Public Domain in the US. Their collaboration meant a great relief for the members of Algolit. Less documents to clean up! All the proofread books are made available on the Project Gutenberg archive. For this exhibition, An Mertens interviewed Linda Hamilton, the General Manager of Distributed Proofreaders.
Interview & transcription: An Mertens
Interface: Michael Murtaugh (Constant)
Contextual stories for Cleaners
Contents
1 Project Gutenberg and Distributed Proofreaders
2 An algoliterary version of the Maintenance Manifesto
2.1 Reference
3 A bot panic on Amazon Mechanical Turk
3.1 References
Project Gutenberg and Distributed Proofreaders
Project Gutenberg is our cave of Ali Baba. It offers over 58,000 free eBooks to be downloaded or read online. Works are accepted on Gutenberg when their U.S. copyright has expired. Thousands of volunteers digitize and proofread books to help the project. An essential part of the work is done through the Distributed Proofreaders project. This is a web-based interface to help convert Public Domain books into e-books. Think of text files, epubs, kindle formats. By dividing the workload into individual pages, many volunteers can work on a book at the same time; this speeds up the cleaning process.
During proofreading, volunteers are presented with a scanned image of the page and a version of the text, as it is read by an OCR algorithm trained to recognize letters in images. This allows the text to be easily compared to the image, proofread, and sent back to the site. A second volunteer is then presented with the first volunteer's work. She verifies and corrects the work as necessary, and submits it back to the site. The book then similarly goes through a third proofreading round, plus two more formatting rounds using the same web interface. Once all the pages have completed these steps, a post-processor carefully assembles them into an e-book and submits it to the Project Gutenberg archive.
We collaborated with the Distributed Proofreaders Project to clean up the digitized files we received from the Mundaneum collection. From November 2018 till the first upload of the cleaned up book 'L'Afrique aux Noirs' in February 2019, An Mertens exchanged about 50 emails with Linda Hamilton, Sharon Joiner and Susan Hanlon, all volunteers from the Distributed Proofreaders Project. The conversation might inspire you to share unavailable books online.
Full email conversation
An algoliterary version of the Maintenance Manifesto
In 1969, one year after the birth of her first child, the New York artist Mierle Laderman Ukeles wrote a Manifesto for Maintenance. Ukeles' Manifesto calls for a readdressing of the status of maintenance work both in the private, domestic space, and in public. What follows is an altered version of her text inspired by the work of the Cleaners.
IDEAS
A. The Death Instinct and the Life Instinct:
The Death Instinct: separation; categorisation; Avant-Garde par excellence; to follow the predicted path to death—run your own code; dynamic change.
The Life Instinct: unification; the eternal return; the perpetuation and MAINTENANCE of the material; survival systems and operations; equilibrium.
B. Two basic systems: Development and Maintenance.
The sourball of every revolution: after the revolution, who’s going to try to spot the bias in the output?
Development: pure individual creation; the new; change; progress; advance; excitement; flight or fleeing.
Maintenance: keep the dust off the pure individual creation; preserve the new; sustain the change; protect progress; defend and prolong the advance; renew the excitement; repeat the flight;
show your work—show it again, keep the git repository groovy, keep the data analysis revealing
Development systems are partial feedback systems with major room for change.
Maintenance systems are direct feedback systems with little room for alteration.
C. Maintenance is a drag; it takes all the fucking time (lit.)
The mind boggles and chafes at the boredom.
The culture assigns lousy status on maintenance jobs = minimum wages, Amazon mechanical turks = virtually no pay.
clean the set, tag the training data, correct the typos,
modify the parameters, finish the report, keep the requester happy,
upload the new version, attach words that were wrongly
separated by OCR back together, complete those Human Intelligence Tasks,
try to guess the meaning of the requester's formatting,
you must accept the HIT before you can submit the results,
summarize the image, add the bounding box,
what's the semantic similarity of this text, check the translation quality,
collect your micro-payments, become a hit Mechanical Turk.
Reference
https://www.arnolfini.org.uk/blog/manifesto-for-maintenance-art-1969
A bot panic on Amazon Mechanical Turk
Amazon's Mechanical Turk takes the name of a chess-playing automaton from the 18th Century. In fact, the Turk wasn't a machine at all. It was a mechanical illusion that allowed a human chess master to hide inside the box and manually operate it. For nearly 84 years, the Turk won most of the games played during its demonstrations around Europe and the Americas. Napoleon Bonaparte is said to have been fooled by this trick too.
The Amazon Mechanical Turk is an online platform for humans to execute tasks that algorithms cannot do. Examples are, annotating sentences as being positive or negative, spotting number plates, discriminating between face and non-face. The jobs posted on this platform are often paid less than a cent per task. Tasks that are more complex or require more knowledge can be paid up to several cents. To earn a living, turkers need to finish as much tasks as fast as possible, leading to inevitable mistakes. As a result, the requesters have to incorporate quality checks when they post a job on the platform. They need to test whether the turker actually has the ability to complete the task, and they also need to verify the results. Many academic researchers use Mechanical Turk as an alternative to have their students execute these tasks.
In August 2018 Max Hui Bai, a psychology student from the University of Minnesota, discovered that the surveys he conducted with Mechanical Turk were full of nonsense answers to open-ended questions. He traced back the wrong answers and found out that they had been submitted by respondents with duplicate GPS locations. This raised suspicion. Though Amazon explicitly prohibits robots to complete jobs on Mechanical Turk, the company is not dealing with the problems they cause on their platform. Forums for Turkers are full of conversations about the automation of the work, sharing practises of how to create robots that can even violate Amazon’s terms. You can also find videos on YouTube that show Turkers how to write a bot to fill in answers for you.
Kristy Milland, an Mechanical Turk activist, says: “Mechanical Turk workers have been treated really, really badly for 12 years, and so in some ways I see this as a point of resistance. If we were paid fairly on the platform, nobody would be risking their account this way.”
Bai is now leading a research among social scientists to figure out how much bad data is in use, how large the problem is, and how to stop it. But it is impossible at the moment to estimate how many datasets have become unreliable in this way.
References
https://requester.mturk.com/create/projects/new https://www.wired.com/story/amazon-mechanical-turk-bot-panic/ https://www.maxhuibai.com/blog/evidence-that-responses-from-repeating-gps-are-random http://timryan.web.unc.edu/2018/08/12/data-contamination-on-mturk/
Informants
Machine learning algorithms need guidance; whether they are supervised or not. In order to separate one thing from another, they need material to extract patterns from. One should carefully choose the study material, and adapt it to the machine's task. It doesn't make sense to train a machine with 19th Century novels if its mission is to analyze tweets. A badly written textbook can lead a student to give up on the subject altogether. A good textbook is preferably not a textbook at all.
This is where the dataset comes in: arranged as neatly as possible, organised in disciplined rows and lined up columns, waiting to be read by the machine. Each dataset collects different information about the world, and like all collections, they are imbued with collectors' bias. You will hear this expression very often: 'data is the new oil'. If only data were more like oil! Leaking, dripping and heavy with fat, bubbling up and jumping unexpectedly when in contact with new matter. Instead, data is supposed to be clean. With each process, each questionnaire, each column title, it becomes cleaner and cleaner, chipping distinct characteristics until it fits the mould of the dataset.
Some datasets combine the machinic logic with the logic of humans. The models that require supervision multiply the subjectivities of both data collectors and annotators, then propagate what they've been taught. You will encounter some of the datasets that pass as default in the machine learning field, as well as other stories of humans guiding machines.
Works
An Ethnography of Datasets
by Algolit
In the transfer of bias from a societal level to the machine level the dataset seems to be overlooked as an intermediate stage in decision making: the parameters by which a social environment is boxed into are determined by various factors. In the creation of datasets that form the basis on which computer models function, conflict and ambiguity are neglected in favour of making reality computable. Data collection is political, but its politics are rendered invisible in the way it is presented and visualised. Datasets are not a distilled version of reality, nor simply a technology in itself. But as any technology, datasets encode their goal, their purpose and the world view of the makers.
With this work, we look into the most commonly used datasets by data scientists for training machine algorithms. What material do they consist of? Who collected them? When? For what reason?
Concept & interface: Cristina Cochior
Wordnet for ImageNet Challenge
by Algolit
Wordnet, created in 1985, is a hierarchical taxonomy that describes the world. It was inspired by theories of human semantic memory developed in the late 1960s. Nouns, verbs, adjectives and adverbs are grouped into synonyms sets or synsets, expressing a different concept. ImageNet is an image dataset based on the WordNet 3.0 nouns hierarchy. Each each synset is depicted by thousands of images. From 2010 until 2017, the ImageNet Object Recognition Challenge (ILSVRC) was a key benchmark in object category classification for pictures, having a major impact on software for photography, image searches, image recognition.
Wordnet for ImageNet Challenge (Vinyl Edition) contains the 1000 synsets used in (which edition of?) this challenge recorded in the highest sound quality that this analog format allows. This work highlights the importance of the datasets used to train artificial intelligence models that run on devices we use on a daily basis. Some of them inherit classifications that were conceived more than 30 years ago. The vinyl is an invitation to thoughtfully analyse them.
Concept & recording: Javier Lloret Voice: xxx
The Annotator
by Algolit
The annotator asks for the guidance of the visitor in annotating the archive of Mundaneum.
The annotation process is a crucial step in supervised machine learning where the algorithm is given examples of what it needs to learn. A spam filter in training will be fed examples of spam and real messages. These examples are entries, or rows from the dataset with a label, spam or not-spam.
The labelling of a dataset is work executed by humans, they pick a label for each row of the dataset. To ensure the quality of the labels multiple annotators see the same row and have to give the same label before an example is included in the training data. Only once enough samples of each label have been gathered in the dataset, the computer can start the learning process.
In this interface we ask you to help us classify the cleaned texts from the Mundaneum archive to expand our training set and improve the quality of the installation 'Classifying the World' in Oracles.
Concept, code, interface: Gijs de Heij
Who wins
Who wins: creation of relationships
by Louise Dekeuleneer, student Arts²/Digital Arts
French is a gendered language, indeed many words are female or male and few are neutral. The aim of this project is to show that a patriarchal society also influences the language itself. The work focused on showing whether more female or male words are used and highlighting the influence of context on the gender of words. At this stage, no conclusions have been drawn yet.
Law texts from 1900 to 1910 made available by the Mundaneum have been passed into an algorithm that turns the text into a list of words. These words are then compared with another list of French words, in which is specified whether the word is male or female. This list of words comes from Google Books. They created a huge database in 2012 from all the books scanned and available on Google Books.
Male words are highlighted in one colour and female words in another. Words that are not gendered (adverbs, verbs,...) are not highlighted. All this is saved as an HTML file so that it can be directly opened in a web page and printed without the need for additional layout. This is how each text becomes a small booklet by just changing the input text of the algorithm.
Contextual stories about Informants
Contents
1 Datasets as representations
1.1 Reference
2 Labeling for an oracle that detects vandalism on Wikipedia
3 How to make your dataset known
4 Extract from a positive IMdB movie review from the NLTK dataset
5 The ouroboros of machine learning
5.1 Reference
Datasets as representations
The data collection processes that lead to the creation of the dataset raise important questions: who is the author of the data? Who has the privilege to collect? For what reason was the selection made? What is missing?
The artist Mimi Onuoha gives a brilliant example of the importance of collection strategies. She chooses the case of statistics related to hate crimes. In 2012, the FBI Uniform Crime Reporting Program (UCR) registered almost 6000 committed hate crimes. However, the Department of Justice’s Bureau of Statistics came up with about 300.000 reports of such cases. That is over 50 times as much. The difference in numbers can be explained by how the data was collected. In the first situation law enforcement agencies across the country voluntarily reported cases. For the second survey, the Bureau of Statistics distributed the National Crime Victimization form directly to the homes of victims of hate crimes.
In the natural language processing field the material that machine learners work with is text-based, but the same questions still apply: who are the authors of the texts that make up the dataset? During what period were the texts collected? What type of worldview do they represent?
In 2017, Google's Top Stories algorithm pushed a thread of 4chan, a non-moderated content website, at the top of the results page when searching for the Las Vegas shooter. The name and portrait of an innocent person were linked to the terrible crime. Google changed its algorithm just a few hours after the mistake was discovered, but the error had already affected the person. The question is: why did Google not exclude 4chan content from the training dataset of the algorithm?
Reference
https://points.datasociety.net/the-point-of-collection-8ee44ad7c2fa
https://arstechnica.com/information-technology/2017/10/google-admits-citing-4chan-to-spread-fake-vegas-shooter-news/
Labeling for an oracle that detects vandalism on Wikipedia
This fragment is taken from an interview with Amir Sarabadani, software engineer at Wikimedia. He was in Brussels in November 2017 during the Algoliterary Encounter.
Femke: If you think about Wikipedia as a living community, with every edit changes the project. Every edit is somehow a contribution to a living organism of knowledge. So then, if from within that community you try to distinguish what serves the community and what doesn't and you try to generalise that, because I think that's what the good faith-bad faith algorithm is trying to do, find helper tools to support the project, you do that on the basis of a generalisation that is on the abstract idea of what Wikipedia is and not on the living organism of what happens every day. What I'm interested about in the relationship between vandalism and debate is how we can understand the conventional drive that sits in these machine-learning processes that we seem to come across in many places. And how can we somehow understand them and deal with them? If you place your separation of good faith-bad faith on preexisting labelling and then reproduce that in your understanding of what edits are being made, how to then take into account movements that are happening, the life of the actual project?
Amir: Ok, I hope that I understood you correctly. It's an interesting discussion. Firstly, what we are calling good faith and bad faith comes from the community itself, we are not doing labelling for them, they are doing labelling for themselves. So, in many different language Wikipedias, the definition of what is good faith and what is bad faith will differ. Wikimedia is trying to reflect what is inside the organism and not to change the organism itself. If the organism changes, and we see that the definition of good faith and helping Wikipedia has been changed, we are implementing this feedback loop that lets people from inside of their community pass judgement on their edits and if they disagree with the labelling, we can go back to the model and retrain the algorithm to reflect this change. It's some sort of closed loop: you change things and if someone sees there is a problem, then they tell us and we can change the algorithm back. It's an ongoing project.
How to make your dataset known
NLTK stands for Natural Language Toolkit. For programmers who process natural language using Python, this is an essential library to work with. Many tutorial writers recommend machine learning learners to start with the inbuilt NLTK datasets. It counts 71 different collections, with a total of almost 6000 items. There is for example the Movie Review corpus for sentiment analysis. Or the Brown corpus, which was put together in the 1960s by Henry Kučera and W. Nelson Francis at the Brown University in Rhode Island. There is also the Declaration of Human Rights corpus, which is commonly used to test whether the code can run on multiple languages. The corpus contains The Declaration of Human Rights expressed in 372 languages from around the world.
But what is the process of getting a dataset accepted into the NLTK library nowadays? On the Github page, the nltk team describes the following requirements:
Only contribute corpora that have obtained a basic level of notability. That means, there is a publication that describes it, and a community of programmers who are using it
Ensure that you have permission to redistribute the data, and can document this. This means that the dataset is best published on an external website with a licence
Use existing NLTK corpus readers where possible, or else contribute a well-documented corpus reader to NLTK. This means, you need to organise your data in such a way, that it can be easily read using NLTK code.
Extract from a positive IMdB movie review from the NLTK dataset
corpus: NLTK, movie reviews
fileid: pos/cv998_14111.txt
steven spielberg ' s second epic film on world war ii is an unquestioned masterpiece of film . spielberg , ever the student on film , has managed to resurrect the war genre by producing one of its grittiest , and most powerful entries . he also managed to cast this era ' s greatest answer to jimmy stewart , tom hanks , who delivers a performance that is nothing short of an astonishing miracle . for about 160 out of its 170 minutes , " saving private ryan " is flawless . literally . the plot is simple enough . after the epic d - day invasion ( whose sequences are nothing short of spectacular ) , capt . john miller ( hanks ) and his team are forced to search for a pvt . james ryan ( damon ) , whose brothers have all died in battle . once they find him , they are to bring him back for immediate discharge so that he can go home . accompanying miller are his crew , played with astonishing perfection by a group of character actors that are simply sensational . barry pepper , adam goldberg , vin diesel , giovanni ribisi , davies , and burns are the team sent to find one man , and bring him home . the battle sequences that bookend the film are extraordinary . literally .
The ouroboros of machine learning
Wikipedia has become a source for learning not only for humans, but also for machines. Its articles are prime sources for training models. But very often, the material the machines are trained on is the same content that they helped to write. In fact, at the beginning of Wikipedia, many articles were written by bots. Rambot, for example, was a controversial bot figure on the English-speaking platform. It authored 98% of the pages describing US towns.
As a result of serial and topical robot interventions, the models that are trained on the full Wikipedia dump, have a unique view on composing articles. For example, a topic model trained on all of Wikipedia articles will associate “river” with “Romania” and “village” with “Turkey”. This is because there are over 10000 pages written about the villages in Turkey. This should be enough to spark anyone's desire for a visit, but it is far too much compared to the number of articles other countries have on the subject. The asymmetry causes a false correlation and needs to be redressed. Most models try to exclude the work of these prolific robot writers.
Reference
https://blog.lateral.io/2015/06/the-unknown-perils-of-mining-wikipedia/
Readers
We communicate with computers through language. We click on icons that have a description in words, we tap words on keyboards, use our voice to give them instructions. Sometimes we trust our computer with our most intimate thoughts and forget that they are extensive calculators. A computer understands every word as a combination of zeros and ones. A letter is read as a specific ASCII number: capital "A" is 001.
In all models, rule based, classical machine learning, and neural networks, words undergo some type of translation into numbers in order to understand the semantic meaning of language. This is done through counting. Some models count the frequency of single words, some might count the frequency of combinations of words, some count the frequency of nouns, adjectives, verbs or noun and verb phrases. Some just replace the words in a text by their index numbers. Numbers optimize the operative speed of computer processes, leading to fast predictions, but they also remove the symbolic links that words might have. Here we present a few techniques that are dedicated to making text readable to a machine.
Works
Algorithmic readings of Bertillon's portrait parlé
by Guillaume Slizewicz (Urban Species)
Un code télégraphique du portrait parlé, written in 1907, is an attempt at translating the "spoken portrait", a face description technique created by a policeman in Paris, into numbers. By implementing this code, it was hoped that faces of criminals and fugitives could be easily communicated through the telegraphic network between countries. In its form, content and ambition this text represents our complicated relationship with documentation technologies. This text sparked the creation of the following installations for three reasons:
- First, the text is an algorithm in itself, a compression algorithm, or to be more precise, the presentation of a compression algorithm. It tries to reduce the information in smaller pieces while keeping it legible for the person who has the code. In this regard it is very much linked to the way we create technology, our pursuit for more efficiency, quicker results, cheaper methods. It represents our appetite for putting numbers on the entire world, measuring the smallest things, labeling the tiniest differences.This text embodies in itself the vision of the Mundaneum.
- Second it is about the reasons for and the applications of technology. It is almost ironic that this text was in the selected archives presented to us in a time when face recognition and data surveillance is so much in the news. This text bears the same characteristics as some of todays’ technology: motivated by social control, classifying people, laying the basis for a surveillance society. Facial features are in the middle of the controversy: mugshot were standardised by Bertillon, now they are used to train neural network to predict criminals from law abiding citizens, facial recognition systems allow the arrest of criminal via CCTV infrastructure and some assert that people’s features can predict sexual orientation.
- The last point is about how it represents the evolution of mankind’s techno-structure. What our tools allow us to do, what they forbid, what they hinder, what they make us remember and what they make us forget. This document allows a classification between people, and a certain vision of what normality is. It breaks the continuum into pieces thus allowing stigmatisation/discrimination. On the other hand this document also feels obsolete today, because our techno-structure does not need such detailed written descriptions about fugitive, criminals or citizen. We can now find fingerprints, iris scans or DNA info in large datasets and compare them directly. Sometimes the technological systems do not even need human supervision and recognise directly the identity of a person via their facial features or their gait. Computer do not use intricate written language to describe a face, but arrays of integers. Hence all the words used in this documents seem désuets, dated. Did we forget what some of them mean? Did photography make us forget how to describe faces? Will voice assistant software teach us again?
Writing with Otlet
Writing with Otlet is a character generator that uses the spoken portrait code as its database. Random numbers are generated and translated into a set of features. By creating unique instances, the algorithm reveals the richness of the description that is possible with the portrait code while at the same time embodying its nuances.
An algorithmic interpretation of Bertillon spoken portrait.
This works draws a parallel between Bertillon systems and current ones. A webcam linked to a facial recognition algorithm captures the beholder face and translate it into numbers on a canvas, printing it alongside Bertillon labelled faces.
Hangman
by Laetitia Trozzi, student Arts²/Section Digital Arts
What better way to discover Paul Otlet and his passion for literature than to play hangman? Through this simple game, which consists in guessing the missing letters in a word, the goal is to make the public discover terms and facts related to one of the creators of the Mundaneum.
Hangman uses an algorithm to detect the frequency of words in a text. Next, a series of significant words were isolated in Paul Otlet's bibliography. This series of words is integrated into a hangman game presented in a terminal. The difficulty of the game gradually increases as the player is offered longer and longer words. During the different game levels, information about the life and work of Paul Otlet is displayed.
TF-IDF
by Algolit
The TF-IDF (Term Frequency-Inverse Document Frequency) is a weighting method used in text search. This statistical measure makes it possible to evaluate the importance of a term contained in a document, relative to a collection or corpus. The weight increases in proportion to the number of occurrences of the word in the document. It also varies according to the frequency of the word in the corpus. The TF-IDF is used in particular in the classification of spam in email softwares.
A web based-interface shows this algorithm through animations allowing to understand the different steps of text classification. How does a TF-IDF-based program read a text? How does it transform words into numbers?
Concept, code, animation: Sarah Garcin
The Book of Tomorrow in a Bag of Words
by Algolit
The bag-of-words model is a simplifying representation of text used in natural language processing. In this model, a text is represented as a collection of its unique words, disregarding grammar, punctuation and even word order. The model transforms the text into a unique list of words and how many times they're used in the text, or quite literally a bag of words.
This heavy reduction of language was the big shock when beginning to machine learn. Bag of words is often used as a baseline, on which the new model has to perform better. It can understand the subject of a text by recognizing the most frequent or important words. Often it is used to measure the similarities of texts by comparing their bags of words.
For this work the article 'Le Livre de Demain' by engineer G. Vander Haeghen, published in 1907 in the 'Bulletin de l'Institut International de Bibliographie' of Mundaneum, has been literally reduced to a bag of words. You can buy your bag at the reception of Mundaneum for 2€.
Concept: An Mertens
Growing a tree
by Algolit
Parts-of-Speech is a category of words that we learn at school: noun, verb, adjective, adverb, pronoun, preposition, conjunction, interjection, and sometimes numeral, article, or determiner.
In natural language processing there exist many writings that allow sentences to be parsed. This means that the algorithm can determine the part-of-speech of each word in a sentence. 'Growing a tree' uses this techniques to define all nouns in a specific sentence. Each noun is then replaced by its definition. This allows the sentence to grow autonomously and infinitely. The recipe of 'Growing a tree' is inspired by Oulipo's constraint of 'Littérature Définitionnelle', invented by Marcel Benabou in 1966. In a given phrase, one replaces every significant element (noun, adjective, verb, adverb) by one of its definitions in a given dictionary ; one reiterates the operation on the newly received phrase, and again.
The dictionary of definitions used in this work is Wordnet. Wordnet is a combination of a dictionary and a thesaurus that can be read by machines. Following Wikipedia it was created in the Cognitive Science Laboratory of Princeton University starting in 1985. The project was initially funded by the U.S. Office of Naval Research and later also by other U.S. government agencies including the DARPA, the National Science Foundation, the Disruptive Technology Office (formerly the Advanced Research and Development Activity), and REFLEX.
Concept, code: An Mertens
Interface: Gijs de Heij
Recipe: Marcel Benabou (Oulipo)
Technique: Wordnet
Contextual stories about Readers
Naive Bayes, Support Vector Machines or Linear Regression are called classical machine learning algorithms. They perform well when learning with small datasets. But they often require complex Readers. The task the Readers do, is also called feature engineering. This means that a human needs to spend time on a deep exploratory data analysis of the dataset.
Features can be the frequency of words or letters, but also syntactical elements like nouns, adjectives, or verbs. The most significant features for the task to be solved, must be carefully selected and passed over to the classical machine learning algorithm. This process marks the difference with Neural Networks. When using a neural network, there is no need for feature engineering. Humans can pass the data directly to the network and achieve fairly good performance right off the bat. This saves a lot of time, energy, and money.
The downside of collaborating with Neural Networks is that you need a lot more data to train your prediction model. Think of 1GB or more of pure text files. To give you a reference, 1 A4, a text file of 5000 characters only weighs 5 KB. You would need 8.589.934 pages. More data also requires more access to useful datasets and more, much more processing power.
Contents
1 Character n-gram for authorship recognition
1.1 Reference
2 A history of n-grams
3 God in Google Books
4 Grammatical features taken from Twitter influence the stock market
4.1 Reference
5 Bag of words
Character n-gram for authorship recognition
Imagine... you've been working for a company for more than ten years. You have been writing tons of emails, papers, internal notes and reports on very different topics and in very different genres. All your writings, as well as those of your colleagues, are safely backed-up on the servers of the company.
One day, you fall in love with a colleague. After some time you realize this human is rather mad and hysterical and also very dependent on you. The day you decide to break up, your now-ex creates a plan to kill you. They succeed. This is unfortunate. A suicide letter in your name is left next to your corpse. Because of emotional problems, it says, you decided to end your life. Your best friends don't believe it. They decide to take the case to court. And there, based on the texts you and others have produced over ten years, a machine learning model reveals that the suicide letter was written by someone else.
How does a machine analyse texts in order to identify you? The most robust feature for authorship recognition is delivered by the character n-gram technique. It is used in cases with a variety of thematics and genres of the writing. When using character n-grams, texts are considered as sequences of characters. Let's consider the character trigram. All the overlapping sequences of three characters are isolated. For example, the character 3-grams of 'Suicide', would be, “Sui,” uic”, “ici”, “cid” etc. Character n-gram features are very simple, they're language independent and they're tolerant to noise. Furthermore, spelling mistakes do not jeopardize the technique.
Patterns found with character n-grams focus on stylistic choices that are unconsciously made by the author. The patterns remain stable over the full length of the text, which is important for authorship recognition. Other types of experiments could include measuring the length of words or sentences, the vocabulary richness, the frequencies of function words; even syntax or semantics-related measurements.
This means not only your physical fingerprint is unique, but also the way you compose your thoughts!
The same n-gram technique discovered that The Cuckoo’s Calling, a novel by Robert Galbraith, was actually written by... J. K. Rowling!
Reference
Paper: On the Robustness of Authorship Attribution Based on Character N-gram Features, Efstathios Stamatatos, in Journal of Law & Policy, Volume 21, Issue 2, 2013.
News article: https://www.scientificamerican.com/article/how-a-computer-program-helped-show-jk-rowling-write-a-cuckoos-calling/
A history of n-grams
The n-gram algorithm can be traced back to the work of Claude Shannon in information theory. In the paper, 'A mathematical theory of communication', published in 1948, Claude Shannon performed the first instance of an n-gram-based model for natural language. He posed the question: given a sequence of letters, what is the likelihood of the next letter?
If you listen to the following excerpt, can you tell who it was written by? Shakespeare or an n-gram piece of code?
SEBASTIAN:
Do I stand till the break off.
BIRON:
Hide thy head.
VENTIDIUS:
He purposeth to Athens: whither, with the vow
I made to handle you.
FALSTAFF:
My good knave.
You may have guessed, considering the topic of this story, that an n-gram algorithm generated this text. The model is trained on the compiled works of Shakespeare. While more recent algorithms, such as the recursive neural networks of the CharNN, are becoming famous for their performance, n-grams still execute a lot of NLP tasks. They are used in statistical machine translation, speech recognition, spelling correction, entity detection, information extraction, ...
God in Google Books
In 2006, Google created a dataset of n-grams from their digitized book collection and released it online. Recently they also created an N-gram viewer.
This allowed for many socio-linguistic investigations of questionable reliability. For example, in October 2018, the New York Times Magazine published an opinion article titled It’s Getting Harder to Talk About God. The author, Jonathan Merritt, had analysed the mention of the word 'God' in Google's dataset using the N-gram viewer. He concluded that there was a decline in the word's usage since the 20th Century. Google's corpus contains texts from the 16th Century leading up to the 21st. However, what the author missed out on, was the growing popularity of scientific journals around the beginning of the 20th Century. This new genre that was not mentioning the word God, shifted the dataset. If the scientific literature was taken out of the corpus, the frequency of the word 'God' would again flow like a gentle ripple from a distant wave.
Grammatical features taken from Twitter influence the stock market
The boundaries between academic disciplines are becoming blurred. Economics research mixed with psychology, social science, cognitive and emotional concepts gives rise to a new economics subfield, called 'behavioral economics'. This means that researchers start to explain an economical behavior based on factors other than the economy only. Both economy and public opinion can influence or be influenced by each other. A lot of research is done on how to use public opinion to predict financial changes, like stock price changes.
Public opinion is estimated from sources of large amounts of public data, like tweets or news. To some extent, Twitter can be more accurate than news in terms of representing public opinion because most accounts are personal: the source of a tweet could be an ordinary person, rather than a journalist who works for a certain organization. And there are around 6,000 tweets authored per second, so a lot of opinions to sift through.
Experimental studies using machinic data analysis show that the changes in stock prices can be predicted by looking at public opinion, to some degree. There are multiple papers that analyze sentiments in news to predict stock trends by labeling them as either “Down” or “Up”. Most of the researchers used neural networks or pretrained word embeddings.
A paper by Haikuan Liu of the Australian National University states that the tense of verbs used in tweets can be an indicator of intensive financial behaviors. His idea was inspired by the fact that the tense of text data is used as part of feature engineering to detect early stages of depression.
Reference
Paper: Grammatical Feature Extraction and Analysis of Tweet Text: An Application towards Predicting Stock Trends, Haikuan Liu, Research School of Computer Science (RSCS), College of Engineering and Computer Science (CECS), The Australian National University (ANU)
Bag of words
In natural language processing, 'bag of words' is considered to be an unsophisticated model. It strips text of its context and dismantles it into a collection of unique words. These words are then counted. In the previous sentences, for example, 'words' is mentioned three times, but this is not necessarily an indicator of the text's focus.
The first appearance of the expression 'bag of words' seems to go back to 1954. Zellig Harris, an influential linguist, published a paper called "Distributional Structure". In the section called "Meaning as a function of distribution", he says "for language is not merely a bag of words but a tool with particular properties which have been fashioned in the course of its use. The linguist's work is precisely to discover these properties, whether for descriptive analysis or for the synthesis of quasi-linguistic systems."
Learners
Learners are the algorithms that distinguish machine learning practices from other types of practices. They are pattern finders, capable of crawling through data and generating some kind of specific 'grammar'. Learners are based on statistical techniques. Some need a large amount of training data in order to function, others can work with a small annotated set. Some perform well in classification tasks, like spam identification, others are better at predicting numbers, like temperatures, distances, stock market values, and so on.
The terminology of machine learning is not yet fully established. Depending on the field, statistics, computer science or the humanities, different terms are used. Learners are also called classifiers. When we talk about Learners, we talk about the interwoven functions that have the capacity to generate other functions, evaluate and readjust them to fit the data. They are good at understanding and revealing patterns. But they don't always distinguish well which of the patterns should be repeated.
In software packages, it is not always possible to distinguish the characteristic elements of the classifiers, because they are hidden in underlying modules or libraries. Programmers can invoke them using a single line of code. For this exhibition, we have therefore developed three table games that show the learning process of simple, but frequently used classifiers and their evaluators, in detail.
Works
Naive Bayes game
In machine learning Naive Bayes methods are simple probabilistic classifiers that are widely applied for spam filtering and deciding whether a text is positive or negative.
They require a small amount of training data to estimate the necessary parameters. They can be extremely fast compared to more sophisticated methods. They are difficult to generalise, this means, that they perform on very specific tasks, demanding to be trained with the same style of data that will be used to work with afterwards.
This game allows you to play along the rules of Naive Bayes. While manually executing the code, you create your own playful model that 'just works'. A little caution is needed: because you only train it with 6 sentences - instead of minimum 2000 - it is not representative at all!
Perceptron game
Neural Networks are the hew hype. They are everywhere, in your search engine, in your translation software, in the ranking of your social media feeds. The basic element of the Neural Network is the Perceptron algorithm. Perceptron is a single layer neural network. A stack of Perceptrons is called a Neural Network.
In this game you experience the specific talents of machines and humans. While we get quickly bored and tend to optimize repetitive tasks, machines are fond of repetitive tasks and execute them without any complaint. And they can calculate really really fast. This game takes 30 minutes to play, while a computer does exactly the same job in a few seconds.
Linear Regression game
Linear Regression is one of the most well known and well understood algorithms in statistics and machine learning. It has been around for almost 200 years. It is an attractive model because the representation is so simple. In statistics, linear regression is is a statistical method that allows to summarize and study relationships between two continuous (quantitative) variables.
By playing this game you will realize that as a player you have a lot of decisions to make. You will experience what it means to create a coherent dataset, to decide what is in and what is not in. If all goes well, you will feel the urge to change your data in order to obtain better results. This is part of the art of approximation that is at the basis of all machine learning practises.
Traité de documentation
Traité de Documentation. Algorithmic poem.
by Rémi Forte, designer-researcher at the l’Atelier national de recherche typographique, Nancy, France
serigraphy on paper, 60 × 80 cm, 25 ex., 2019
This poem, reproduced in the form of a poster, is an algorithmic and poetic re-reading of Paul Otlet's Traité de documentation. It is the result of an algorithm based on the mysterious rules of the human intuition. It is applied to a fragment taken from Paul Otlet's book and is intended to be representative of his bibliological practice. The algorithm splits the text, words and punctuation marks are counted and reordered into a list. In each line, the elements combine and exhaust the syntax of the selected fragment. Paul Otlet's language remains perceptible but exacerbated to the point of absurdity. For the reader, the systematization of the text is disconcerting and his reading habits are disrupted. Built according to a mathematical equation, the typographical composition of the poster is just as systematic as the poem. However, friction occurs occasionally; loop after loop, the lines extend to bite on the neighbouring column. Overlays are created and words are hidden by others. These telescopic handlers draw alternative reading paths.
Contextual stories about Learners
Contents
1 Naive Bayes & Viagra
1.1 Reference
2 Naive Bayes & Enigma
3 A story on sweet peas
3.1 References
4 Perceptron
5 BERT
5.1 References
Naive Bayes & Viagra
Naive Bayes is a famous learner that performs well with little data. We apply it all the time. Christian & Griffiths state in their book, 'Algorithms to Live by', that 'our days are full of small data'. Imagine for example you're standing at a bus stop in a foreign city. The other person who is standing there, has been waiting for 7 minutes. What do you do? Do you decide to wait? And if yes, for how long? When will you initiate other options? Another example. Imagine a friend asking advice on a relationship. He's been together with his new partner for 1 month. Should he invite the partner to join him at a family wedding?
Having preexisting beliefs is crucial for Naive Bayes to work. The basic idea is that you calculate the probabilities based on prior knowledge and given a specific situation.
The theorem was formulated during the 1740s by reverend and amateur mathematician Thomas Bayes. He dedicated his life to solving the question of how to win the lottery. But Bayes' rule was only made famous and known as it is today by the mathematician Pierre Simon Laplace in France a bit later in the same century. For a long time after La Place's death, the theory sunk to oblivion until it was dug out again during the Second World War in an effort to break the Enigma code.
Most people today have come in contact with Naive Bayes through their email spam folders. Naive Bayes is a widely used algorithm for spam detection. It is by coincidence that Viagra, the erectile dysfunction drug, was approved by the US Food & Drug Administration in 1997, around the same time as about 10 million users worldwide had made free web mail accounts. The selling companies were among the first to make use of email as a medium for advertising: it was an intimate space, at the time reserved for private communication, for an intimate product. In 2001, the first SpamAssasin programme relying on Naive Bayes was uploaded to SourceForge, cutting down on guerilla email marketing.
Reference
Machine Learners, by Adrian MacKenzie, The MIT Press, Cambridge, US, November 2017.
Naive Bayes & Enigma
This story about Naive Bayes is taken from the book: 'The theory that would not die', written by Sharon Bertsch McGrayne. Amongst other things, she describes how Naive Bayes was soon forgotten after the death of Pierre Simon Laplace, its inventor. The mathematician was said to have failed to credit the works of others. Therefore, he suffered widely circulated charges against his reputation. Only after 150 years the accusation was refuted.
Fast forward to 1939, when Bayes' rule was still virtually taboo, dead and buried in the field of statistics. When France was occupied in 1940 by Germany, who controlled Europe's factories and farms, Winston Churchill's biggest worry was the U-boat peril. The U-boat operations were tightly controlled by German headquarters in France. Each submarine received orders as coded radio messages long after it was out into the Atlantic. The messages were encrypted by word scrambling machines, called Enigma machines. Enigma looked like a complicated typewriter. It was invented by the German firm Scherbius & Ritter after the First World War, when the need for message encoding machines had become painfully obvious.
Interestingly, and luckily for Naive Bayes and the world, at that time, the British government and educational systems saw applied mathematics and statistics as largely irrelevant to practical problem solving. So the British agency charged with cracking German military codes mainly hired men with linguistic skills. Statistical data was seen as bothersome because of its detail-oriented nature. So wartime data was often analyzed not by statisticians, but by biologists, physicists, and theoretical mathematicians. None of them knew that the Bayes rule was considered to be unscientific in the field of statistics. Their ignorance proved fortunate.
It was the now famous Alan Turing, a mathematician, computer scientist, logician, cryptoanalyst, philosopher and theoretical biologist, who used Bayes' rules probabilities system to design the 'bombe'. This was a high-speed electromechanical machine for testing every possible arrangement that an Enigma machine would produce. In order to crack the naval codes of the U-boats, Turing simplified the 'bombe' system using Baysian methods. It turned the UK headquarters into a code-breaking factory. The story is well illustrated in 'The Imitation Game', a film by Morten Tyldum in 2014.
A story on sweet peas
Throughout history, some models were invented by people with ideologies that are not to our liking. The idea of regression stems from Sir Francis Galton, an influential 19th Century scientist. He spent his life studying the problem of heredity – understanding how strongly the characteristics of one generation of living beings manifested in the following generation. He established the field of eugenics, and defined it as ‘the study of agencies under social control that may improve or impair the racial qualities of future generations, either physically or mentally.’ On Wikipedia, Galton is a prime example of scientific racism.
Galton initially approached the problem of heredity by examining characteristics of the sweet pea plant. He chose this plant because the species can self-fertilize. Daughter plants inherit genetic variations from mother plants without a contribution from a second parent. This characteristic eliminates having to deal with multiple sources.
Galton's research was appreciated by many intellectuals of his time. In 1869, in 'Hereditary Genius', Galton claimed that genius is mainly a matter of ancestry and he believed that there was a biological explanation for social inequality across races. Galton even influenced his half-cousin Charles Darwin of his ideas. After reading Galton's paper, Darwin stated, "You have made a convert of an opponent in one sense for I have always maintained that, excepting fools, men did not differ much in intellect, only in zeal and hard work." Luckily, the modern study of heredity managed to eliminate the myth of racially-based genetic difference, something Galton tried so hard to maintain.
Galton's major contribution to the field was linear regression analysis, laying the groundwork for much of modern statistics. While we engage with the field of machine learning, Algolit tries not to forget that ordering systems hold power, and that this power has not always been used to the benefit of everyone. Machine learning has inherited many aspects of statistical research, some less agreeable than others. We need to be attentive, because these world views do seep into the algorithmic models that create new orders.
References
http://galton.org/letters/darwin/correspondence.htm
https://www.tandfonline.com/doi/full/10.1080/10691898.2001.11910537
http://www.paramoulipist.be/?p=1693
Perceptron
We find ourselves in a moment in time in which neural networks are sparking a lot of attention. But they have been in the spotlight before. The study of neural networks goes back to the 1940s, when the first neuron metaphor emerged. The neuron is not the only biological reference in the field of machine learning - think of the word corpus or training. The artificial neuron was constructed in strong connection to its biological counterpart.
Psychologist Frank Rosenblatt was inspired by fellow psychologist Donald Hebb's work on the role of neurons in human learning. Hebb stated that "cells that fire together wire together." His theory now lies at the basis of associative human learning, but also unsupervised neural network learning. It moved Rosenblatt to expand on the idea of the artificial neuron.
In 1962, he created the Perceptron. The perceptron is a model that learns through the weighting of inputs. It was set aside by the next generation of researchers, because it can only handle binary classification. This means that the data has to be clearly separable, as for example, men and women, black and white. It is clear that this type of data is very rare in the real world. When the so-called first AI winter arrived in the 70s and the funding decreased, the Perceptron was also neglected. For 10 years it stayed dormant. When Spring settled at the end of the 80s, a new generation of researchers picked it up again and used it to construct neural networks. These contain multiple layers of perceptrons. That is how neural networks saw the light. One could say that the current machine learning season is particularly warm, but it takes another Winter to know a Summer.
BERT
Some online articles say the year 2018 marked a turning point for the field of Natural Language Processing. A series of deep-learning models achieved state-of-the-art results on tasks like question answering or sentiment classification. Google’s BERT algorithm entered the machine learning competitions of last year as a sort of “one model to rule them all.” It showed a superior performance over a wide variety of tasks.
BERT is pre-trained; its weights are learned in advance through two unsupervised tasks. This means BERT doesn’t need to be trained from scratch for each new task. You only have to finetune its weights. This also means that a programmer wanting to use BERT, does not know any longer what parameters BERT is tuned to, nor what data it has seen to learn its performances.
BERT stands for Bidirectional Encoder Representations from Transformers. This means that BERT allows for bidirectional training. The model learns the context of a word based on all of its surroundings, left and right of a word. As such, it can differentiate between 'I accessed the bank account' and 'I accessed the bank of the river'.
Some facts:
BERT_large, with 345 million parameters, is the largest model of its kind. It is demonstrably superior on small-scale tasks to BERT_base, which uses the same architecture with “only” 110 million parameters.
to run BERT you need to use TPU's. These are the Google's CPU's especially engineered for TensorFLow, the deep learning platform. TPU's renting rates range from 8$/h till 394$/h. Algolit doesn't want to work with off-the-shelf-packages, we are interested in opening the blackbox. In that case, BERT asks for quite some savings in order to be used.
References
https://ai.googleblog.com/2018/11/open-sourcing-bert-state-of-art-pre.html
https://towardsdatascience.com/deconstructing-bert-distilling-6-patterns-from-100-million-parameters-b49113672f77
Sources
Reference in New Issue Copy Permalink