Showing posts with label narrative networks. Show all posts
Showing posts with label narrative networks. Show all posts

Wednesday, August 3, 2022

Collective Chronopathy


Twitter reveals dynamics of stories surrounding Trump's presidency
Dec 2021, phys.org

Just here for the words:
  • Story turbulence - how quickly a new story fizzled-out as new stories popped
  • Narrative control - how often someone else retweets your tweet shows how much you control the story
  • Collective chronopathy - the rate at which a population’s stories for a subject seem to change over time

via Public Library of Science and University of Vermont, Burlington: Dodds PS, Minot JR, Arnold MV, Alshaabi T, Adams JL, Reagan AJ, et al. (2021) Computational timeline reconstruction of the stories surrounding Trump: Story turbulence, narrative control, and collective chronopathy. PLoS ONE 16(12): e0260592. doi.org/10.1371/journal.pone.0260592

Post Script:
Northern English verbal mannerisms being lost
Aug 2021, phys.org

Image credit: Silicone oil droplets, Thiévenaz and Sauret UC Santa Barbara, 2022. A droplet of silicone oil pinches off from fluids with different concentrations of 140 µm particles: (A) is pure liquid, (B) is 2% concentration, (C) is 20%, and (D) is 50%. Credit: Thiévenaz and Sauret

Friday, July 9, 2021

Writing Robots 3

OpenAI (competitor to DeepMind) created a text generator that basically wipes the need for human writers. Or to be a bit less hyperbolic, it makes it all but impossible for you to discriminate between human writing and robot writing. 

It's called GPT, for "Generative Pre-trained Transformer". An older version, GPT-2, was so good that it wasn't released for fear that it would be used for malicious purposes (i.e., fake news). Now we're on GPT-3, which will be released, and is 100 times more powerful. But no need for alarm, OpenAI intends to "prevent misuse by limiting access to approved customers and use cases" (though it's now been exclusively licensed to Microsoft).

For comparison purposes, this is a much earlier version of an artificial writing program -- The Policeman's Beard Is Half-Constructed (1984).

At last, I wonder, what does this mean for the blogosphere, of which Network Address has been a part  for over 10 years now. Will it be completely diluted by an entire universe of fake blogs? And then I realize, who cares, nobody reads this blog except for robots anyway (i.e., spiderbots).

AI tool summarizes lengthy papers in a sentence
Jan 2021, phys.org
Semantic Scholar is notable for achieving the greatest compression rate of all summarizing tools; powered by AI and used for scientific research. With its new summarization feature, it surveys massive numbers of scientific research papers and reduces them to one-sentence summaries. 
It began as an AI-fueled dungeon game
May 2021, Wired via Ars Technica

AKA "GPT-3 text generator video game generates sex scenes involving children"

You know the story by now. Just ask Tay, the Microsoft Chatbot who should have been called the n*gg**bot because it's entire lexicon defaulted to saying the n-word over and over, with a little bit of Holocaust denial sprinkled in for good measure. (See Ars reporting on that.)

Also though, complaining about your "8-year old laptop" while playing this game will get you shadowbanned. Who knew robots were that sensitive.

A college kid created a fake, AI-generated blog -- It reached #1 on Hacker News
Dec 2020, MIT Technology Review

And this is exactly what they thought would happen when its release was withheld:
The lab gave the algorithm to select researchers who applied for a private beta, with the goal of gathering their feedback and commercializing the technology by the end of the year.

Porr submitted an application. He filled out a form with a simple questionnaire about his intended use. But he also didn’t wait around. After reaching out to several members of the Berkeley AI community, he quickly found a PhD student who already had access. Once the graduate student agreed to collaborate, Porr wrote a small script for him to run. It gave GPT-3 the headline and introduction for a blog post and had it spit out several completed versions. Porr’s first post (the one that charted on Hacker News), and every post after, was copy-and-pasted from one of the outputs with little to no editing.

The trick to generating content without the need for much editing was understanding GPT-3’s strengths and weaknesses. “It's quite good at making pretty language, and it's not very good at being logical and rational,” says Porr. So he picked a popular blog category that doesn’t require rigorous logic: productivity and self-help. [hilarious]

Porr says his experiment also shows a more mundane but still troubling alternative: people could use the tool to generate a lot of clickbait content. “It's possible that there's gonna just be a flood of mediocre blog content because now the barrier to entry is so easy,” he says. “I think the value of online content is going to be reduced a lot.”
Post Script:
It doesn't stop with writing, it can also generate imagery:
New module for OpenAI GPT-3 creates unique images from text
Jan 2021, phys.org
A team of researchers at OpenAI, a San Francisco artificial intelligence development company, has added a new module to its GPT-3 autoregressive language model. Called DALL·E, the module excerpts text with multiple characteristics, analyzes it and then draws a picture based on what it believes was described

The system is able to create images by using a corpus of information consisting of internet pages. Each part of the text is researched in an attempt to learn what it should look like. For the previous example, it would search for and analyze thousands of pictures of dogs. Then it would study cats, and what their claws look like, and then birds and their tails. Then, it combines the results into several graphic images to give users a variety of results.

via OpenAI: DALL·E: Creating Images from Text: openai.com/blog/dall-e/

Thursday, January 14, 2021

Exposure Science


Fox News hosts have measurable effect on COVID cases, study finds
Apr 2020, Ars Technica

The study comes from a group of researchers led by the University of Chicago's Leonardo Bursztyn and uses survey data gathered in April from 1,045 regular viewers of Fox News (aged 55 and over) to examine the timing of behavioral changes in response to the virus—when people began to cancel travel, isolate, increase the frequency of hand-washing, and so on.

Survey participants who preferred watching the alarmed newscaster began changing their behavior on average three days earlier than other Fox News viewers. Meanwhile, participants who preferred the more languid newscaster acted much later — five days after other Fox News viewers, and eight days after alarmed viewers.

I think the key word here is "alarm"?

And the above unrelated image is not a tree being struck by lightning, it's a synthetic light painting.
 

Precipitating the Noosphere


How conspiracy theories emerge—and how their storylines fall apart
Jun 2020, phys.org

^Came for the image: "Researchers produced a graphic representation of the narratives they analyzed, with layers for major subplots of each story, and lines connecting the key people, places and institutions within and among those layers." Credit: UCLA College, UCLA Samueli School of Engineering

Stay for the story: They studied Bridgegate NJ (real conspiracy) and Pizzagate DC (fake conspiracy). And the AI can figure out which is which.

All stories are a narrative network, the elements being the characters, places and things in the story, and the network emerging from the relationships between these elements.

If you tell the AI enough stories, it will detect patterns in the networks.

"One of the characteristics of a conspiracy theory narrative framework is that it is easily 'disconnected,'" said Timothy Tangherlini, one of the paper's lead authors, a professor in the UCLA Scandinavian section whose scholarship focuses on folklore, legend and popular culture. "If you take out one of the characters or story elements of a conspiracy theory, the connections between the other elements of the story fall apart."

Real stories can still make sense even if you remove one of the elements. Something about the myriad connections between the elements, because they're real people and places, all connected in other ways not intrinsic to the story. Robustness you might call it when talking about telecommunications networks; or redundancy.

Pizzagate? You remove Wikileaks from the network and it falls apart. (People used this massive data dump to link all kinds of elements in creative ways, building vast and complex stories. But the data was too big, and had no connection to each other beyond their being in the same repository. 

Also the timeframe, and for similar reasons, is a good point of contrast. Real stories build over time, because again, all the elements are connected in myriad ways with each other for reasons that have nothing to do with any one particular story. Fake stories appear out of nowhere, and tend to evaporate just as quickly. 

Timothy R. Tangherlini et al. An automated pipeline for the discovery of conspiracy and conspiracy theory narrative frameworks: Bridgegate, Pizzagate and storytelling on the web, PLOS ONE (2020). DOI: 10.1371/journal.pone.0233879. http://dx.doi.org/10.1371/journal.pone.0233879

Same article but from The Conversation, Nov 2020.


Coronavirus - Harmful lies spread easily due to lack of UK law
Jul 2020, BBC News

I won't even bother to link here the other recent report about how malicious social interference dramatically (and wholly undetected?) altered public opinion in the Brexit election. I think that's general public knowledge by now.

From the article: Despite investing in measures, the UK Digital, Culture, Media and Sport Committee says tech firms can not be left to self-regulate. They also said that social media firms' advertising-focused business models had encouraged the spread of misinformation and allowed "bad actors" to make money from emotional content, regardless of the truth. By bad actors, their report identifies state actors, including Russia, China and Iran; the Islamic State group; far-right groups in the US and the UK; and scammers.


Tracking misinformation campaigns in real-time is possible, study shows
Aug 2020, phys.org

Metadata is key. 

There's so much good information just in this summary alone, so I copying much of it:

Using data from past misinformation (troll) campaigns from China, Russia, and Venezuela waged against the United States before and after the 2016 election, combined with posts to Twitter and Reddit and the hyperlinks or URLs they included, using a "postURL pair," which is simply a post with a hyperlink.

8,000 accounts and 7.2 million posts from late 2015 through 2019.

They reinforced their model with a baseline of a rich dataset of politically engaged and average user posts collected over many years by NYU's Center for Social Media and Politics (CSMaP).

They teased out data for timing, word count, whether the mentioned URL domain is a news website, and most importantly "metacontent," for example, whether a URL was in the top 25 political domains shared by other trolls.

Trolls referring to other trolls and not new sources: "Both trolls and normal people often include local news URLs in their posts, but the trolls tended to mention different users in such posts, probably because they are trying to draw their audience's attention in a new direction. Metacontent lets the algorithm find such anomalies."
-Jacob N. Shapiro, professor of politics and international affairs at the Princeton School of Public and International Affairs

Across almost all of the 463 different tests, it was clear which posts were and were not part of an influence operation, meaning that content-based features can indeed help find coordinated influence campaigns on social media.

Venezuelan trolls only retweeted certain people and topics, making them easy to detect. Russian and Chinese trolls were better at making their content look organic, but they, too, could be found. In early 2016, for example, Russian trolls quite often linked to far-right URLs, which was unusual given the other aspects of their posts, and, in early 2017, they linked to political websites in odd ways. Overall, Russian troll activity became harder to find as time went on. It is possible that investigative groups or others caught on to the false information, flagging the posts and forcing trolls to change their tactics or approach, though Russians also appear to have produced less in 2018 than in previous years.

via Princeton University, New York University and New Jersey Institute of Technology: M. Alizadeh el al., "Content-based features predict social media influence operations," Science Advances (2020). doi:10.1126/sciadv.abb5824. https://advances.sciencemag.org/lookup/doi/10.1126/sciadv.abb5824

Nikon Small World 2020 - Single Neuron - Nadia Efimova

A novel strategy for quickly identifying twitter trolls
Aug 2020, phys.org

50 tweets is all it takes. 

Using sociolinguistic "troll-specific restrictions" on words and word pairs, the signal in the noise becomes clear. Granted, they would have to update their restrictions depending on what cultural phenomena they're using in their campaigns, but it works in real time. 

Also, they checked their model using a library of Russian troll tweets; because this exists.

Monakhov S (2020) Early detection of internet trolls: Introducing an algorithm based on word pairs / single words multiple repetition ratio. PLoS ONE 15(8): e0236832. doi.org/10.1371/journal.pone.0236832


Calls to city 311 lines can predict opioid overdose hotspots
Nov 2020, phys.org

This is just old school predictive analytics.

"Complaints to the city about issues like streetlight repair, abandoned vehicles and code violations reflect disorder and distress that are also linked to opioid use," said Yuchen Li, lead author of the study and doctoral candidate in geography at The Ohio State University.

via Ohio State: Scientific Reports (2020). DOI: 10.1038/s41598-020-76685-z


New algorithm signals a possible disease resurgence
Sep 2020, phys.org

The algorithm monitors public health data to detect statistical patterns associated with impending outbreaks to predict the reemergence of existing infectious diseases.

Changes in vaccination rates, or even changes in climate, can predict the emergence of an epidemic.

They used 10,000 sets of simulated case reports covering a period of 10 years. When tested on a 2004-5 Mumps outbreak, the model predicted it four years on advance. Looking at Pertussis, it could predict 100% at the state level 1990-2000. It works for vector-borne disease as well (ticks, mosquitoes).

Tobias S. Brett et al. Dynamical footprints enable detection of disease emergence, PLOS Biology (2020). DOI: 10.1371/journal.pbio.3000697


How Facebook, Twitter, YouTube, and Reddit are handling the election
Nov 2020, Ars Technica

"In addition to its stated policies, Facebook is reportedly standing ready to implement a slew of policies it has used in the past for managing election content in "at-risk" countries such as Sri Lanka and Myanmar. These tools would include limiting the rate at which content beginning to go viral can travel, as well as tweaking the newsfeed to change what type of content users see."

So we can stop the virus from spreading all along.

Sunday, August 16, 2020

Sociothermodynamics and Predictive Analytics

 

Investigating dynamics of democratic elections using physics theory

Feb 2020, phys.org

 

Sometimes, physics theories and models can be used to study seemingly unrelated phenomena, such as social behaviors or social dynamics.

 

While human beings are not exactly similar to individual physical particles, theories or techniques that physicists use to analyze "behavioral patterns" in atoms or electrons may aid the general understanding of large-scale social behaviors, as long as these behaviors do not depend on small-scale details.

 

Based on this idea, some researchers use physics theories to investigate social behaviors that take place during democratic elections, for example.

 

The very idea of a quantifiable social network was unknown to the average person prior to the year 2000. (Check out Barabasi's Links for a reminder of what the world was like at that time.) Now, the idea is clear as day, with the trademark "nodes-and-links" emblazoned into the retina of any 21st century citizen. We all understand what a social network is, and being that a social network is the marriage of sociology and network science, we can all understand these models and their implications.

 

Below is a list of articles related to sociothermodynamics, although I tend to use this term as a synonym for predictive analytics as well, so there are other articles about using all kinds of big data (like semantic analysis) to make sense out of the sloppy sociobiological mess that we spread all over the planet. And speaking of spreading all over, the last article uses biological data from slime molds to describe human social behavior.

 

 

The physics that drives periodic economic downturns

Mar 2020, phys.org

 

There exist universal mechanisms that give rise to laws governing the growth of economics.

 

The way spilled milk spreads across the floor can explain why economic downturns regularly occur, as a natural feature of physics, rooted in the time-dependent movement of spreading over an area. Growth of the innumerable spreading phenomena over time follows the shape of an "S curve" otherwise known as the sigmoid function.

 

A bottle of milk spilled on the floor will have a small initial footprint, followed by a rapid finger-shaped expansion across the kitchen's tiles, followed by a final phase of slow creep.

 

This same history of slow, fast, slow can be seen in chemical reactions, population growth, the adoption of new technology and even the spreading of new ideas.

-Adrian Bejan et al, Energy theory of periodic economic growth, International Journal of Energy Research (2020). DOI: 10.1002/er.5267

 

 

Secularism and tolerance of minority groups predicts future prosperity of countries

Feb 2020, phys.org

 

Changes in culture generally come before any improvements in wealth, education and democracy, rather than the other way around.

 

Promoting democracy, whether through economic exchange or regime change, will only succeed if combined with promoting openness and tolerance of minority groups.

 

(In the 20th Century) places which had the greatest improvement also tended to have pre-existing secular and tolerant cultures.

 

 

Study of 62 countries finds people react similarly to everyday situations

June 2020, phys.org

 

The world is a much more similar and unified place than we once thought.

 

[Because at the right scale, we are all just particles banging around on the surface of the planet.]

 

This project is unprecedented. Very few international studies look at relationships between more than two countries, let alone 62," Lee, a doctoral researcher in the lab of UCR Distinguished Professor David Funder, and the lead author of the paper.

 

 

Facebook users change their language before an emergency hospital visit - study

Mar 2020, phys.org

 

Researchers from the Perelman School of Medicine at the University of Pennsylvania and Stony Brook University's Computer Science Department compared patients' Facebook posts to their medical records, which showed that a shift to more formal language and/or descriptions of physical pain, among other changes, reliably preceded hospital visits.

 

 

Wikipedia visits to disease outbreak pages show impact of news media on public attention

Mar 2020, phys.org

 

During the 2016 Zika outbreak, news exposure appears to have had a far bigger impact than local disease risk on the number of times people visited Zika-related Wikipedia pages in the U.S.

 

Previous research has explored how public opinion responds to media exposure during an emerging outbreak, but has mostly relied on surveys rather than observational data.

 

The analysis showed that Zika-related Wikipedia page view counts during the outbreak were highly synchronized with mentions of the virus in web and national TV news at both the national and state level.

 

 

What do 'Bohemian Rhapsody,' 'Macbeth,' and a list of Facebook friends all have in common?

June 2020, phys.org

 

I almost forgot to mention -- network science is rapidly evolving into a useful tool for understanding our world, and there's a possibility that it can go even further than physics in modeling human dynamics; I saw earlier this year a that network science could do a better job than physics of describing what happens inside the sun. [The link to that study is probably on this weblog somewhere].

 

New research published in Nature Physics uses tools from network science to explain how complex communication networks can efficiently convey large amounts of information to the human brain. Conducted by postdoc Christopher Lynn, graduate students Ari Kahn and Lia Papadopoulos, and professor Danielle S. Bassett, the study found that different types of networks, including those found in works of literature, musical pieces, and social connections, have a similar underlying structure that allows them to share information rapidly and efficiently.

 

The researchers evaluated 40 real-world communication networks to see what features were crucial for communicating information. They looked at works of English literature, including the canon of Shakespeare and Jane Austen's "Pride and Prejudice," along with musical pieces such as Mozart's Sonata No 11 and Queen's "Bohemian Rhapsody." They also studied networks of social relationships, including co-authorship networks in science and Facebook friend connections.

 

After looking at this diverse group of networks, the researchers found that the large-scale structure of a network was essential to that network's ability to convey information. What was surprising was just how similar this structure was across the different networks, whether the network was representing noun transitions in a work of literature or melodic progressions in a piece of music.

 

What makes these networks both information-rich and efficient is a balance between two key network features known as "community" structure and "heterogeneous" structure. Community structure occurs when nodes clump together and form clusters that evoke related concepts. Saying the word "dog" might bring to mind "ball," "Frisbee," or "bone," for example. Such community structure helps make networks more efficient because a person can anticipate what word or idea might come next.

 

But if a person can anticipate what comes next, there won't be much information conveyed because information is directly related to surprise. To provide information, networks have to have a "heterogeneous" mixture of both well-connected and sparsely connected nodes. Take the works of Shakespeare as an example. While "the" and "and" are used 28,944 and 27,317 times, respectively, there are also 12,493 word forms that only occur once. "At a hub like, 'the,' you can't anticipate where you are about to go," says Lynn. "It turns out that these hub nodes are really important for generating surprisal or, equivalently, information."

 

 

Evolution selects for 'loners' that hang back from collective behavior-at least in slime molds

Mar 2020, phys.org

 

Testing whether loners are random or a predictable quantity, possibly subject to natural or cultural selection:

 

Target system: cellular slime mold Dictyostelium discoideum

 

Evolution could indeed select for loner behavior in slime molds. Loners are both an ecological and an evolutionary insurance plan, a way to diversify a genetic portfolio to ensure the survival of the social, collective behavior:

 

Slime molds, when they are threatened by starvation, the tiny amoebae coalesce into slug-like creatures that then aggregate into a large, swaying tower that grows upward with a burgeoning slimy top—until that top sticks to an unwitting passing insect, the starvation-resistant spores hitchhiking out into the world, while all the individuals making up the base and stalk die.

 

But what caught Tarnita's eye were the slime mold loners, the amoebae that resist the biochemical call to form the tower.

 

Here and there, some scattered cells on the plate just didn't seem to react at all to this aggregation process.

 

They tested the loners to see if they were flawed in some way, but they couldn't find anything wrong with them.

 

It could make sense for some fraction of the slime mold population to remain behind in order to take advantage of any resources that might return in the environment while the rest of the cells are aggregating.

 

30% chose the loner life over collective action.

 

The proportion of solitary cells in the social amoeba Dictyostelium discoideum is not simply determined by each cell individually tossing a coin. It results instead from interactions between the [organism] and the environment.

 

The decision not to become part of the collective is, in fact, taken collectively. All the cells kind of talk to each other chemically: 'Oh, you're going? I guess I'm staying.' There's communication involved in becoming a loner.

-"Eco-evolutionary significance of 'loners'" by Fernando W. Rossine, Ricardo Martinez-Garcia, Allyson E. Sgro, Thomas Gregor and Corina E. Tarnita, appears in the Mar. 18 issue of the journal PLoS Biology.

 

Image source: Fernando Fornies Gracia, infrared photography

 

Other Infrared Photos

Saturday, July 11, 2020

Writing Robots 2

OpenAI releases powerful text generator
June 2020, phys.org

For scale and context:
Those concerns led OpenAI in February 2019 to take the unusual step of declining to release the early version, GPT-2, citing fears potential misuse could be dangerous. 
GPT-3, about 100 times more powerful than GPT-2, has performed admirably in tests, according to the OpenAI report. It tackled reading compression exercises requiring filling in word blanks, tackling "on-the-fly reasoning," and generating compositions up to 500 words.

Partially Related Article:
Your brain shows if you are lonely or not
June 2020, phys.org
The closer participants felt to someone, the more similarly their brain represented them throughout the social brain.
I mean if this doesn't say something about identity -- You are a collective of those closest to you, not a person so much as a multi-person chimera.

Friday, July 10, 2020

Neutral Networks


The image you see here is a partial map of the internet, circa 2005. It's a colorful organic explosion of lifewebs. It's the closest thing to seeing the noosphere that we have ever come. It's beautiful, and I think we need to see it once in a while, but it's totally unrelated to this post.

Study finds no media bias when it comes to story selection
Apr 2020, phys.org

Hassell's work on the study began in 2017 and embraced a number of research-gathering methods, including a survey of 6,000 journalists, election returns, Twitter data and a novel correspondence experiment design.

"There is an institutional set of norms that dictate newsworthiness," he said. "Lots of things drive those norms, but reporters' personal, ideological preferences about what they view as valuable is not one of them."

Notes:
Hans J. G. Hassell et al. There is no liberal media bias in which news stories political journalists choose to cover, Science Advances (2020). DOI: 10.1126/sciadv.aay9344