This comment reminds me of another Shirky essay [1]. He says:
> [This is] the pattern for descriptions of the Semantic Web. First, take some well-known problem. Next, misconstrue it so that the hard part is made to seem trivial and the trivial part hard. Finally, congratulate yourself for solving the trivial part.
The difficult part is figuring out exactly what is and what isn't true. If we had a perfect map of reality it would be easy to detect fake news. In fact we wouldn't need any news at all or journalists or anyone else except the people who build the ontology.
[1] http://www.shirky.com/writings/herecomeseverybody/semantic_s...
If that is indeed your argument, it is remarkably similar to the argument "Perfect security is impossible, this system is worthless." I think it is demonstrably obvious that there is a related parameter in both cases, which is the cost of overcoming the system is greater than the gain from overcoming said system.
As an example of how that argument fails in practice, consider textbooks. When I was working on this problem full time the organization had a relationship with a major publisher of textbooks. Consider the system which can construct an ontology and knowledge graph based on the contents of a text book. It may not be 'true' in the Aristotle sense, but it is an effective tool for evaluating whether or not a paper or essay correlates that ontology and knowledge graph or doesn't. In such a system there is value in evaluating the extent with which the knowledge in the textbook has been communicated to the person writing the essay.
Being able to automate the analysis has other benefits as well which extend beyond flagging documents which disagree a previous established canon.
It does amuse me how much this framing of problems happens. Both in the annoying cases of arm chair quarterbacks, and in the cases of folks that take on problems previously thought of as too difficult.
As a disclaimer, we tried to get this model off the ground in YC summer 2017 batch but were rejected after the phone interview. I did not assist in the research; only an effort to publicize.
I whole-heartedly believe this is our best bet at combatting "fake news" online for the moment. Taking the lessons of "The Most Human Human: What Artificial Intelligence Teaches Us About Being Alive" [2], we might tease out credibility instead of fact or fiction to help guide our biased/opinionated labels.
From the abstract:
"In other words, the language used by millions of people on Twitter has considerable information about an event’s credibility. For example, hedge words and positive emotion words are associated with lower credibility"
[1]: https://credcount.com/whitepaper.pdf
[2]: https://www.amazon.com/Most-Human-Artificial-Intelligence-Te...
On top of that what might be considered fake news by some doesn't have to be by others because it might have utility.
And lats but not least. Combatting fake news would mean that you would be able to cover all contextual interpretations of a given statement which we simply don't have access to.
So from my perspective you can't battle fake news in any meaningful ways which won't end up doing more harm than good as "fake news" is not really fake in some objectively defined way.
Do you have an example of this in mind?
Ten years ago, I would have said that this would not happen within 20 years, but now I am much less certain, due to a couple of profound breakthroughs: word embeddings (Glove etc), and their famous linear analogies, and sense disambiguation.
These have not only been achieved in the emergent "look how amazingly bizarre" sense, but have over the last few years gotten a much more rigorous footing: notably the papers:
[1] Linear Algebraic Structure of Word Senses, with Applications to Polysemy (RECOMMEND reading both first and last version, as some things are more lucid in the first version, and some in the last. The main takeaway is that the word vectors are linear combinations of a low number of sense vectors, and these sense vectors can be recovered by sparse coding)
[2] Towards Understanding Linear Word Analogies (Which finally explains the "king" + "woman" - "man" = "queen" phenommena)
Now consider physics for example, running text and equations, then consider
* an "overscored n": is this 1) normal vector or 2) an antineutron or 3) ... ?
* a greek beta symbol: is this 1) v/c as in relativity 2) kT as in thermmal or statistical physics or 3) ... ?
* an e symbol: is this the exponent function, or the electron or ... ?
With today's tools it should already be possible to make this: type a LaTeX formmula, or perhaps even a picture of a formula, then mouse-over an ambiguous symbol to see the sense of the symbol, or alternatively generate a wikipedia style list: "where ... is ..."
It should be possible to do this for math as well, and pretty soon software should be able to formalize quasi-formal mathematics books, papers, derivations, proofs ... into MetaMath, and the beautiful part? MetaMath contains the ground truth upon verifying so it can keep training the formalizer while formalizing without supervision! The verifier will become the supervisor!
Of course MetaMath will not tell you if your postulates and axioms are correct, but it will check if your reasoning is sound. The axioms or postulates will be what characterize the belief systems, and any rammifications or inconsistencies will be linked to in the "belief package" managers so that whenever you encounter a formalized claim you can then reply with how the claim does not follow in your own belief system, and it will hence become netiquette to not just share vague claims, but formalized ones and not just formalized ones but accompanying a reference to the claims proof and in which belief systems they hold, it will become much easier to identify exactly where people who disagree with you differ in assumptions or axioms, and hence easier to collaborate on identifying inconsistensies in the endless stream of forked belief systems to avoid being filtered by the reader's or platform's access to gossiped inconsistency proofs...
A corollary is that the semi-crackpot semi-true realpolitik interpretations from the lower regions of society will force the higher regions of society to either formalize explicit injustice and oppression, or else democractically cede the definition of the boundary between just and unjust to the populace, so that we can have a formalized egalitarian society...
Just like the printing press democratized communication, machine learning and formalization will democratize interpretation.
You might be able to say that fake news doesn't align with true facts that aren't part of the facts being reported as news. That is you can tell that "the color of the sky unexpectedly changed from green to yellow yesterday" is fake news, but you can't use true facts to tell that ""the color of the sky unexpectedly changed from blue to yellow yesterday" is fake news. (I mean, yes, you might still use true facts about atmospheric composition, light scattering, and the spectrum of solar output, but you can't use "the sky is blue".)
If you get 5 streams of bits and four of them correlate with some epsilon number of new bits, and the fifth has 10x epsilon new bits, one of things is likely true. Either the odd stream out is corrupted or the stream has changed.
Apply that to news stories where the ontology (O) and knowledge graph (KG) are combined to "error correct" the content of the story. "Error bits" would be defined as relationships that don't match what the ontology says can be true, or facts that are different than those facts already in the knowledge base. You can flag the odd news story as lacking alignment with the existing KG&O, further you can quantify that lack of alignment fairly precisely.
So if the news story has 0 bits of information it correlates very strongly with other existing stories. If it has many bits of information it contrasts with other existing stories. At some number of stories on the same topic you can begin to reason about which are more likely 'true' and which are more likely 'false.'
To suggest a specific example, imagine a person living inside WWII-era Germany correlating a propaganda news article with every other news article they have read, all of which are also propaganda. Correlation-based news epistemology isn't even based on a prevailing social consensus, it's based on the consensus of broadcasters.
If we had a system like this that worked, we could at least leverage it to be able to quantify the coherence of a much larger set of inputs and with many controlled corpora at once.
It also fails when there are few streams to compare, as when a new story first breaks. Or when there is coordination of actors.
My suggestion is not that it is fail safe or perfect, only that is a useful application of an ontology which disputes the original author's thesis that they are overrated.
All of these things are hallmarks of fake news: it is claimed to be new (hence "news",) is not published very widely (you can't compare a aquariusrising.biz article on Green Space Aliens to a CNN article on the same, because there isn't a CNN article on that), and it's usually coordinated between copycat opportunists. Even worse, none of these things are distinguishable from a grassroots report unless you know the true news a-priori. This better not be a mechanism to ban everything that isn't on CNN.com.
If your ontology is probabilistic and multi-dimensional, it's much more likely to represent truth.
I actually think ontologies are a trap for smart people. For people who like to systematize, they have an allure. Ontologies give you the illusion of putting the chaotic world into order. Even phrases like "true facts" - in reality, facts are hard to come by and change over time, and the closer you look at them the more woolly they get - predicates are far too abstract and simplistic, and when you dig in, the fractal messiness overwhelms.
It's a hard problem.
I’ve followed this topic off and on for 20 years, and I’ve yet to see a big success. I’ve only ever seen this cycle:
1) Someone discovers ontology and rdf,
2) They implement some obvious examples,
3) They get hyped and tell everyone,
4) They try to generalize their examples to real-world problems,
5) They get silent and disappear.
Fake news is a great example of how sloppy categorization can end up being. Fake news is mostly about creating clickbaity stories which will make people share and thus increase advertising revenue, yet today most people think about it as some sort of political propaganda.
Reality is more nuanced than our categories allow us to explore and often times they are limiting our ability to think about the world around us.
Clays point sees to be that the value of information is not in categorizing it but allowing for new relationships to be made, which is exactly what ontology can often hinder.
Someone with an agenda in the symbolic vs statistic AI question could take this parallel as an example of how close ontology-based AI approaches are to the way humans think. And then someone with the opposite agenda would point out that the example is all hypothetical.
Nice try, but this is missing the whole point. This is the crux of the matter and the problem with the category of "fake news", and precisely why this ontology is garbage.
The fact that you can read about the same topic in several (more often opposing) ways is already a reason to red-flag the piece in the first place. This isn’t readily available through any of the current newspapers. Everyone has an opinion.
I would happily pay for an aggregator that simply ejects facts (stated facts, not truths) from the stories. Then combines that with similar other pieces.
There is already the AP news ticker, or Reuters. Bloomberg has the same thing for financial matters. If you read only that, you will die of boredom.
Would you otherwise prefer to read Russian propaganda speaking of German authorities policing Germans of Russian heritage? Because that's exactly what they are currently doing.
0 facts, amazing opinionated narrative in multi-paragraph form. Pictures on top of that. But that definitely sounds more fun. Moreover - that sort of narrative is way closer to the general sentiment of the working classes who are generally predisposed towards learning about how "the Man" is taking advantage of them in the first place.
The only narrative is the selection of what facts to report. Without that you get lost in the deluge and get equally uninformed.
When done correctly, ontology combined with knowledge graphs can make an impractical problem much more practical. Consider for the moment the problem of 'fake news'. What characterizes 'fake news' is that it has facts or concepts that are not in alignment with "true" facts (authoritative). And without getting into the debate of what is and isn't true, you can take a document, a knowledge base populated with trusted facts, and a topic ontology and score that document for alignment with the existing data set. Whether you use that for checking Wikipedia articles or flagging fake news its a very powerful combination.