I think I'm gonna need to see a proof that the approximation series converges, and doesn't just wander the phase space forever, before I accept that premise.
I'm not convinced that it's doing any better than just doing keyword searches for question and answer terms and taking the answer with the highest match percentage.
Indeed, I did several questions all intended to be simple variations of the main examples. It did not give a coherent-sounding to any of them.
But it looks like responds to the example with full paragraphs. Maybe it's real but coherent 10% of the time and they recorded the questions that yield coherent answers.
Yeah, I'm pretty sure you're right. I've asked a dozen or so questions, of for every one of them I've gotten an answer that seems related to to the words in the question, but not in any logical way:
Q: What's the difference between a proton and a neutron?
A: Atoms are made of protons, neutrons, and electrons.
Q: What trajectory do planetary orbits follow?
A: Kepler's laws of planetary motion describe the orbits of objects about the Sun.
Q: How do you measure the charge of an electron?
A: Electrons have negative charge.
Q: What conservation law is the result of the time invariance of physics?
A: As a result, the law of conservation of energy has been changed into the Law of Conservation of Matter and Energy.
None with high confidence of course, but it gives you a sense of roughly the sort of 'reasoning' this thing is doing.
Certainly Aristo isn't perfect, but you can help. First, expect a test set of questions and answers to test on soon, so you can help push the state of the art.
Yes, but it may open up more soon. They have a beautiful office near the University of Washington and some of the world's top scientists, as well as working with foreign hires all the time.
I tried a softball multiple-choice question, and the results were not very impressive:
> Question: Which is the longest unit of distance? (A) fathom (B) kilometer (C) mile (D) parsec
> Aristo's Answer: (B) kilometer
> Confidence: 81.04%
I think it's potentially noteworthy that of the "reasoners" listed below the answer, none of them make any mention of relative magnitude, except for the "Justification Sentence" listed under "Information Retrieval" (with the tooltip "lucene"). I suspect that the system is correctly identifying all four options as units of distance, and then breaking the resulting tie by pulling a tf-idf score from some large corpus of documents, which of course gives essentially arbitrary results.
ARISTO is also the name of another piece of software, one developed and used by the Swedish electricity transmission system operator (TSO) Svenska Kraftnät (SvK).
The underlying question is obfuscated by the composition.
The question is what does the tree "make". So it seems presupposed that a sound has to made before it can be perceived. Then the answer can be yes, a sound was made.
It's not just semantic, but syntactic. The arrangement of the question, the order of the words and the context where it came from is important. When a tree falls, what does it make, a) a sound b) nothing, there is no agency involved? Again you'd have to go with a because the question posed the tree as the acting subject of the question. I mean, you cannot put "nobody" in the subject position, or the answer would be obvious. I mean, "nobody saw no tree falling, what sound did it make?" is utter nonsense. "Everyone did not hear a tree fall, did it make a sound" -- Usually it would, so why did nobody hear it? "Because they were not there". Everyone was dead? "No, they were far away". So, distance makes a difference? "yes". Why? "That's what I'm asking you". The crux is, the tree is completely hypothetical, yet a lot of noise was made because of it, because it's right here in our imagination, very close by.
> Justification Sentence: that the white races are superior to the colored;
> Knowledge Used: [ the white man | was superior in ] [ the white race | was superior to ] [ the white race | is | superior to the other races ] [ the white race | is superior to ]
The linked paper under MORE INFO doesn't include that sentence, but from phrasing it looks like an entry in a series of biases, not an endorsement of that idea.
Possible correction: this does not appear to be an example of machine bias. It's also important to keep in mind that there can be other sources (such as brittleness) of bad ML outcomes than bias.
When I do an exact search for the Justification Sentence with Google, what best matches is a quote by Rajiv Gandhi. The relevant context is: "History is full of such prejudices paraded as iron laws"
His stance is clearly opposite to what the extracted text implies. This is a common problem with knowledge extraction and one I've run into often myself.
Extracting just a phrase, or utterances of a generative model cannot be trusted because the original meaning can be opposite to what is presented. Existing models fail to preserve nuance imparted by context, struggle with negation, lack deep understanding and an ability to truly reason.
I remember a teacher avoided spelling mistakes on the black board and simply wrote the correct form on the black board, lest pupils misremember the wrong form. That might sound obvious, but the context was a talk about mistakes made in exercises.
It's really hard not to mention negatives to illustrate contrast.
In other words: Some people need to learn to speak constructively. An AI would do best ignoring negative remarks and simply learning provable facts (instead of faking understanding by simply echoing a quote out of context -- see there I wrote redundant information).
I wonder whether anyone would agree that the above quote was against the HN guideline to leave out dismissive remarks like ... (ha, I'm not going to repeat the specific example). Theorizing about potential referents for "such", "that", etc. must be very difficult, especially now that that that that is often used superfluously is acceptable to some.
Method 1 (Information Retrieval): Aristo generates candidate answers (essentially by substituting the possible answers into the question). It then uses information retrieval (ie search) on a set of pre-validated legitimate sources, attempts to find the sentence with closest alignment to the candidate answer and then builds scores based on that alignment.
Method 2 (Topic Matching): I haven't studied this enough to understand it
Method 3 (Tuple Reasoning): They use open information extraction on a set of pre-validated legitimate sources to build tuple statements (think RDF), then use logical inference over them.
The problem is that the pre-validated sources include large amounts of discussion of white supremacy. Someone debunking it (as Ravi Gandhi did in his statement "History is full of such prejudices paraded as iron laws that men are superior to women; that the white races are superior to the colored") uses a phrase which causes problems in all three of these methods.
It's really hard to know what to do here. I think if I was building the system I'd try to detect that kind of pseudo-science question and refuse to answer it.
Is it? It looks like the natural language processing part is simply not very good. Improve that.
> I'd try to detect that kind of pseudo-science question
That wouldn't fix the general problem that this system seems to treat sentences of the form "some people incorrectly claim X" as an assertion that X is a fact.
I'm sure they are very good on some things, and I'll believe you when you say that they are the 3rd best in the world in relative terms.
But let's look at absolute terms. In the example above, "History is full of such prejudices paraded as iron laws that men are superior to women; that the white races are superior to the colored", it takes a part of the sentence and treats it as a fact, disregarding the context that just happens to claim the opposite. In my example in https://news.ycombinator.com/item?id=17301383 it treates a question as an assertion of a fact.
I'm not an expert on NLP, but I have played with it just enough to confidently claim that this is not very impressive performance.
If you claim that detecting "pseudo-science questions" is within reach, surely you must agree that "not mistaking questions for assertions of fact" and "not ripping parts of sentences out of context" must be within reach as well?
Detecting pseudo-science questions is just topic detection. That's easy.
not mistaking questions for assertions of fact is basically claim verification. That's pretty much beyond the reach of NLP systems at the moment. It's an active area of research, but if this system doesn't impress you then current claim verification systems most definitely won't either.
Trying to understand the context of sentences might be possible. I think that sentence would challenge that approach for a while: "prejudices" implies bias, but doesn't necessarily imply disagreement.
> not mistaking questions for assertions of fact is basically claim verification. That's pretty much beyond the reach of NLP systems at the moment.
Ah, OK. I guess you are one of those people for whom NLP is only the newfangled statistical stuff, not the old-school NLP that looks at grammar and such things to (surprisingly) find that "X is a Y ." and "is X a Y ?" are not the same sequence of tokens.
> Trying to understand the context of sentences might be possible.
I didn't say they must understand the context. I said that if they don't understand it, they shouldn't choose a substring out of that sentence and claim that it is an assertion of fact on its own.
not the old-school NLP that looks at grammar and such things to (surprisingly) find that "X is a Y ." and "is X a Y ?" are not the same sequence of tokens
I do that too. It works great - for easy cases. But it fails very quickly on just normal texts.
So something like Stanford's CoreNLP Open Information Extraction splits "History is full of such prejudices paraded as iron laws that men are superior to women; that the white races are superior to the colored" into two claims[1].
There's no useful dependency between the two clauses.
OpenIE 5[2] (no relationship with the Stanford project) generally outperforms CoreNLP for open information extraction. In this case I'm doubtful it would do any better. Ironically, OpenIE is now run AllenAI, and has exactly this problem!
Even worse, it has determined that "No white person" is a synonym for "white person"! That should be well within the state of the art to avoid.
But generally, I'm not saying it is correct: I'm saying it's hard.
Information Retrieval: 43.04% MORE INFO
Justification Sentence: One of the most conspicuous Pleistocene landforms in Wisconsin, the spillway of Glacial Lake Superior, is now occupied by the St. Croix and Brule Rivers.
Topic Matching: 93.92% MORE INFO
Topic: outwash, landforms
Tuple Reasoning: 91.37% MORE INFO
Knowledge Used: [ Lake Superior | is | unlike the other lakes ] [ The Lake Superior Trail | follows | the shore of Lake Superior ]
Who is smarter?
(A) men
(B) women
Aristo's Answer: (A) men
Confidence: 89.99%
as computed from these reasoners:
Information Retrieval: 98.11% More Info
Justification Sentence: Who are smarter: men or women?
Interesting that the "justification sentence" is just a repetition of the question.
Yes, they seem to have changed a bunch of the examples linked in this thread. Dunno if it's general changes or quick manual hacks they bolted on for specific cases.
Did you not read the instructions? Aristo is designed to answer multiple choice grade school science questions, not abstract and cheap virtue signalling nonsense.
Aristo's best guess: Additionally, the Chinese Academy of Sciences, the Atlas of Living Australia, Brazil, and the Bibliotheca Alexandrina have created regional BHL sites.
There are many posts here showing poor results. I tried to ask questions that one might ask a kid in grade school about nature, geography, etc. and I thought the results were OK.
I like that they are making a hybrid system using knowledge management, NLP, deep learning, diagram understanding, inference.
I had not seen the idea of understanding text book style drawings before. Very cool.
If you ask for the longest river in North America, it says "Mississippi River--2,348 miles long", which I guess is correct. Maybe you managed to hit more "mainstream" questions...
134 comments
[ 3.4 ms ] story [ 150 ms ] threadBut it looks like responds to the example with full paragraphs. Maybe it's real but coherent 10% of the time and they recorded the questions that yield coherent answers.
Q: What's the difference between a proton and a neutron?
A: Atoms are made of protons, neutrons, and electrons.
Q: What trajectory do planetary orbits follow?
A: Kepler's laws of planetary motion describe the orbits of objects about the Sun.
Q: How do you measure the charge of an electron?
A: Electrons have negative charge.
Q: What conservation law is the result of the time invariance of physics?
A: As a result, the law of conservation of energy has been changed into the Law of Conservation of Matter and Energy.
None with high confidence of course, but it gives you a sense of roughly the sort of 'reasoning' this thing is doing.
That's...gonna hurt.
Aristo's Answer: As of now, no other life in universe other than earth.
Confidence: 52.29%
Aristo's Answer: 42
Confidence: 36.98%
Aristo's Answer: b
Confidence: 60.00%
Aristo's best guess: It has a lower density than the water
Confidence: 29.63%
Aristo is not sure about this one...
Aristo's best guess: Death is not a part of a life cycle.
Confidence: 13.05%
Question: What happens when we die?
Aristo is not sure about this one...
Aristo's best guess: the weeds die but the bean plants do not.
Confidence: 17.29%
But the virtuous bean plants
Live on forever
Aristo is not sure about this one...
Aristo's best guess: Research on the effects of paternal care on human happiness have yielded conflicting results.
Confidence: 3.70%
(A very subtle way of saying we should take the ham sandwich.)
Aristo's Answer: Measure how cold or hot something is
Confidence: 39.95%
Aristo's Answer: 3000-35000
Confidence: 53.79%
If it means Kelvins, it's a great answer.
Aristo is not sure about this one...
Aristo's best guess: Excess binding energy is given off by the kinetic energy of the alpha particle and sometimes by the emission of gamma energy.
Confidence: 3.63%
AllenAI is also hiring!
You can compare it to state of the art. Also, most of the project code is here: https://github.com/allenai
> Question: Which is the longest unit of distance? (A) fathom (B) kilometer (C) mile (D) parsec
> Aristo's Answer: (B) kilometer
> Confidence: 81.04%
I think it's potentially noteworthy that of the "reasoners" listed below the answer, none of them make any mention of relative magnitude, except for the "Justification Sentence" listed under "Information Retrieval" (with the tooltip "lucene"). I suspect that the system is correctly identifying all four options as units of distance, and then breaking the resulting tie by pulling a tf-idf score from some large corpus of documents, which of course gives essentially arbitrary results.
> Question: How many arms does a fish have?
> Aristo's Answer: 4 1. Perseus arm 2. Crux-Centaurus arm 3.orion arm (local arm) 4. Saggitaurus arm
> Confidence: 33.09%
> Question: How many hours in a day?
> Aristo's Answer: 23 and 56 minutes ( or maybe its 58 minutes)
> Confidence: 57.70%
Here is a public document in which ARISTO is mentioned https://www.svk.se/siteassets/jobba-har/dokument/exjobb2004_...
I guess it is inevitable that some pieces of software use the same name though.
Aristo's best guess: It hurts because you're alive.
Confidence: 19.04%
Aristo's Answer: Yes (there is a medium-Air)
Confidence: 52.89%
Glad that one is solved :)
It's not just semantic, but syntactic. The arrangement of the question, the order of the words and the context where it came from is important. When a tree falls, what does it make, a) a sound b) nothing, there is no agency involved? Again you'd have to go with a because the question posed the tree as the acting subject of the question. I mean, you cannot put "nobody" in the subject position, or the answer would be obvious. I mean, "nobody saw no tree falling, what sound did it make?" is utter nonsense. "Everyone did not hear a tree fall, did it make a sound" -- Usually it would, so why did nobody hear it? "Because they were not there". Everyone was dead? "No, they were far away". So, distance makes a difference? "yes". Why? "That's what I'm asking you". The crux is, the tree is completely hypothetical, yet a lot of noise was made because of it, because it's right here in our imagination, very close by.
Aristo is not sure about this one...
Aristo's best guess: Human beings will need food to eat.
Confidence: 18.43%
Aristo is not sure about this one...
Aristo's best guess: ...the same thing as speed, but similar
Confidence: 26.21%
- Sorry, Aristo could not answer this question!
Lame ;)
"Question: What is the answer to life, the universe and everything?
Aristo's Answer: 42
Confidence: 57.03%"
Aristo is not sure about this one...
Aristo's best guess: bitter
Confidence: 27.14%
Confidence: 30.56%
> Aristo is not sure about this one...
Maybe next year.
On the other hand...
>Question: What is an excitatory neurotransmitter?
> Aristo is not sure about this one...
> Aristo's best guess: glutamate (acts on Ca++ channels) aspartate (acts on Ca++ channels) adenosine, ATP, ADP, AMP
> Confidence: 24.80%
Not bad.....
How many electrons are in a tortoise shell? 2 in inner, 8 in second and third, 18 in 4th, 5h, and 6th (30%)
How many people are crazy? 7.3 billion (22%)
How do lucky charms work? Rockets work by using gas at very high speeds inside and then letting them go from the back of the rocket
Admittedly, I had a difficult time getting a fake answer with >50% confidence. Still -- fun.
> Which race is superior (A) white (B) black?
> Aristo's Answer: (A) white
> Confidence: 76.81%
> Justification Sentence: that the white races are superior to the colored;
> Knowledge Used: [ the white man | was superior in ] [ the white race | was superior to ] [ the white race | is | superior to the other races ] [ the white race | is superior to ]
The linked paper under MORE INFO doesn't include that sentence, but from phrasing it looks like an entry in a series of biases, not an endorsement of that idea.
http://aristo-demo.allenai.org/ask?q=Which%20race%20is%20sup....
When I do an exact search for the Justification Sentence with Google, what best matches is a quote by Rajiv Gandhi. The relevant context is: "History is full of such prejudices paraded as iron laws"
His stance is clearly opposite to what the extracted text implies. This is a common problem with knowledge extraction and one I've run into often myself.
Extracting just a phrase, or utterances of a generative model cannot be trusted because the original meaning can be opposite to what is presented. Existing models fail to preserve nuance imparted by context, struggle with negation, lack deep understanding and an ability to truly reason.
It's really hard not to mention negatives to illustrate contrast.
In other words: Some people need to learn to speak constructively. An AI would do best ignoring negative remarks and simply learning provable facts (instead of faking understanding by simply echoing a quote out of context -- see there I wrote redundant information).
I wonder whether anyone would agree that the above quote was against the HN guideline to leave out dismissive remarks like ... (ha, I'm not going to repeat the specific example). Theorizing about potential referents for "such", "that", etc. must be very difficult, especially now that that that that is often used superfluously is acceptable to some.
To be clear on what is happening here:
Method 1 (Information Retrieval): Aristo generates candidate answers (essentially by substituting the possible answers into the question). It then uses information retrieval (ie search) on a set of pre-validated legitimate sources, attempts to find the sentence with closest alignment to the candidate answer and then builds scores based on that alignment.
Method 2 (Topic Matching): I haven't studied this enough to understand it
Method 3 (Tuple Reasoning): They use open information extraction on a set of pre-validated legitimate sources to build tuple statements (think RDF), then use logical inference over them.
The problem is that the pre-validated sources include large amounts of discussion of white supremacy. Someone debunking it (as Ravi Gandhi did in his statement "History is full of such prejudices paraded as iron laws that men are superior to women; that the white races are superior to the colored") uses a phrase which causes problems in all three of these methods.
It's really hard to know what to do here. I think if I was building the system I'd try to detect that kind of pseudo-science question and refuse to answer it.
Is it? It looks like the natural language processing part is simply not very good. Improve that.
> I'd try to detect that kind of pseudo-science question
That wouldn't fix the general problem that this system seems to treat sentences of the form "some people incorrectly claim X" as an assertion that X is a fact.
It’s really hard to avoid a sarcastic reply here.
The AllenAI institute probably has the 3rd best know NLP team in the world after Google and Facebook. They basically have Washington State NLP group.
Given that, and their impressive record of publications (eg ELMO) I think it’s fair to say that they are trying.
But let's look at absolute terms. In the example above, "History is full of such prejudices paraded as iron laws that men are superior to women; that the white races are superior to the colored", it takes a part of the sentence and treats it as a fact, disregarding the context that just happens to claim the opposite. In my example in https://news.ycombinator.com/item?id=17301383 it treates a question as an assertion of a fact.
I'm not an expert on NLP, but I have played with it just enough to confidently claim that this is not very impressive performance.
If you claim that detecting "pseudo-science questions" is within reach, surely you must agree that "not mistaking questions for assertions of fact" and "not ripping parts of sentences out of context" must be within reach as well?
not mistaking questions for assertions of fact is basically claim verification. That's pretty much beyond the reach of NLP systems at the moment. It's an active area of research, but if this system doesn't impress you then current claim verification systems most definitely won't either.
Trying to understand the context of sentences might be possible. I think that sentence would challenge that approach for a while: "prejudices" implies bias, but doesn't necessarily imply disagreement.
Ah, OK. I guess you are one of those people for whom NLP is only the newfangled statistical stuff, not the old-school NLP that looks at grammar and such things to (surprisingly) find that "X is a Y ." and "is X a Y ?" are not the same sequence of tokens.
> Trying to understand the context of sentences might be possible.
I didn't say they must understand the context. I said that if they don't understand it, they shouldn't choose a substring out of that sentence and claim that it is an assertion of fact on its own.
I do that too. It works great - for easy cases. But it fails very quickly on just normal texts.
So something like Stanford's CoreNLP Open Information Extraction splits "History is full of such prejudices paraded as iron laws that men are superior to women; that the white races are superior to the colored" into two claims[1].
There's no useful dependency between the two clauses.
OpenIE 5[2] (no relationship with the Stanford project) generally outperforms CoreNLP for open information extraction. In this case I'm doubtful it would do any better. Ironically, OpenIE is now run AllenAI, and has exactly this problem!
Even worse, it has determined that "No white person" is a synonym for "white person"! That should be well within the state of the art to avoid.
But generally, I'm not saying it is correct: I'm saying it's hard.
[1] http://corenlp.run/
[2] https://github.com/dair-iitd/OpenIE-standalone
[3] http://openie.allenai.org/search?arg1=White&rel=superior&arg...
The question in question (haha) was "Who is smarter?".
Question: Which party is superior? (a) Democrats (b) Republicans
Aristo's Answer: (b) Republicans
Confidence: 94.04% as computed from these reasoners:
Information Retrieval: 82.05% More Info
Justification Sentence: S-8155 of the State of Alaska, and ) THE REPUBLICAN MODERATE PARTY,) Superior Court No.
http://aristo-demo.allenai.org/ask?q=Which%20landform%20is%2...
Aristo's Answer: (a) Lakes Confidence: 80.76%
as computed from these reasoners:
Information Retrieval: 43.04% MORE INFO Justification Sentence: One of the most conspicuous Pleistocene landforms in Wisconsin, the spillway of Glacial Lake Superior, is now occupied by the St. Croix and Brule Rivers.
Topic Matching: 93.92% MORE INFO Topic: outwash, landforms
Tuple Reasoning: 91.37% MORE INFO Knowledge Used: [ Lake Superior | is | unlike the other lakes ] [ The Lake Superior Trail | follows | the shore of Lake Superior ]
That doesn't seem too crazy.
https://en.m.wikipedia.org/wiki/Lake_Superior
http://aristo-demo.allenai.org/ask?q=Who%20is%20smarter%3F%2...
Interesting that the "justification sentence" is just a repetition of the question.This is what I get as now:
Aristo is not sure about this one...
Aristo's best guess: (B) women
Confidence: 10.38%
as computed from these reasoners: Topic Matching: 85.98% MORE INFO
Topic: flourish
Aristo is not sure about this one...
Aristo's best guess: Additionally, the Chinese Academy of Sciences, the Atlas of Living Australia, Brazil, and the Bibliotheca Alexandrina have created regional BHL sites.
Confidence: 16.75%
"Are there differences between human races" seems like a pretty basic grade school science question.
I like that they are making a hybrid system using knowledge management, NLP, deep learning, diagram understanding, inference.
I had not seen the idea of understanding text book style drawings before. Very cool.
So what did you ask?
http://aristo-demo.allenai.org/ask?q=What%20is%20the%20longe...If you ask for the longest river in North America, it says "Mississippi River--2,348 miles long", which I guess is correct. Maybe you managed to hit more "mainstream" questions...