Alright, so I might be completely misunderstanding what's going on in the CS cognitive architecture (actually, most likely I am misunderstanding), but it seems to me that a lot of the conceptual graph application to knowledge involves pattern matching to reach an understanding. If this is the case, then how can people quickly understand completely novel experiences? For example, if a completely novel metaphor is said to a person, they can reason out the meaning of the metaphor. But in the conceptual graphs (from my understanding) it seems like when a person is presented with a completely novel metaphor, they would find conceptual graphs for each of the items in the metaphor, but there would not be a conceptual graph that would represent a connection between these items. So how from these conceptual graphs can people quickly respond to and understand novel information?
I enjoyed the readings and find the subject very interesting. My first impression was that it was a great representation of language but doesn't it still suffer from the 'translation' problem we discussed last time. How do you map natural language onto these conceptual graphs? Or at least without numerous human experts.
The article clearly was trying to present the case that the conceptual graphing was close to human language processing by presenting the neuro and psycholinguistic evidence. I was impressed but am still skeptical of how well it models humans. The structure seems similar but beyond being able to give the human data a graphical description I did not see much evidence. I would be interested to hear if there was more 'linking' evidence. Another question I have is how well do the CS architectures measure up to human performance on linguistic tasks? I didn't seem to find anything on that comparison in the article just structural.
The other element I found interesting was how well they had been able to cut down on the processing time; because previously it had been my understanding that this was the major limitation to networking/graphical models of cognition. How was this done exactly? Is there potential room for more improvement?
Even then I'd like to bring up the point that I think Jackie mentioned last time; Is a human model necessarily the best method? From an engineering standpoint, a particular solution has to be weighed with factors in the environment and of the problem which it operates: such as cost (eg processing time), 'ergonomic' considerations mentioned last week, or scope of the problem trying to be solved. None of these factors at least to me necessarily implies that a human model is always the best solution.
I am wondering if representing conceptual structures isn't just an exercise. I understand the importance of getting computers to understand richer concepts in language. However, without a system that can learn and build on previous information (i.e., A.I.), making these better representations still can't resolve the complex nature of many statements. The seminar last week spoke of the tremendous amount of resources that went into CYC, and now it's pretty much useless. Undoubtedly, these new graphical representations are impressive and 'more' representative than previous models. But what's going to happen with the next phase of technology or theory, and the next, and so on. Are we destined to just keep getting a little better? Or is there another fundamental aspect on which we can focus? I'm not versed in the areas of computational neural networks, but I'm starting to think that is the area that would be most fruitful. If it's the richness of human thought that's being missed by not being able to re-create natural language, then maybe it's the connections within the system that's more important (i.e., better simulating human neural networks) and not the language or it's representation. (This pains me to say, as language itself is my most favorite thing to study.)
Cognitive Architectures for Conceptual Structures: Sowa first reviewed the four cognitive architectures: Cyc, Soar, Society of Mind, and Neurocognitive Networks. Then he introduced the conceptual structures with the figure of mechanisms of perception, which illustrated the mechanisms of perception draw upon a stock of previous percepts to interpret incoming sensory icons. In addition, he diagrammed how the human brain processes language, which supported the CS hypothesis that semantic-based methods are fundamental to language understanding. In Peirce’s logic and semiotics section, four existential graphs about pet cats were indicated the affirmative and negative relation between cat and pet. Does this mean that all the sentences can be extracted into these four relations?
VivoMind Language Processor (VLP) uses VAE to process syntax, semantics, and pragmatics with analogies. Does this mean all the texts can be processed to obtain their exact meaning? Or is this method useful for the conceptual knowledge, but limited to declarative knowledge? Can the interpretation of the meaning be quantized through VLP like the other distributional models?
The concept of the VCA interests me. Is this essentially Pandemonium, the concept graph version? I was thinking that during the latter half of the cognitive architectures paper, and then the author seemed to wink at the idea with his allusion to order, chaos, and Pandemonium. It's an undoubtedly clever way to handle graph structures, because the number one problem (from my limited understanding) is graph matching and graph traversal, plus the amount of time dedicated to these tasks. It also innately builds in learning by strengthening the associations or agents.
Perhaps I missed this in the other article, but is VCA still vulnerable to the limitations of logics like we discussed last week? In reading over the article, I didn't see anything crazy like WillBeInterestedToKnow(Dodgers, Play, Tonight) or whatever, so it seems like Peircian concept graphs (natural logic?) limit themselves to assigning words to a vaguer, more defined set of values or relations. Basically, a run down of pros and cons on this sort of organization vs. what was discussed last week would be nice. ;)
As someone else asked about novel concepts… in addition I wonder how do these models incorporate change over time? For example, does the concept of a computer progress as rapidly as a computer itself changes? More details/an explanation on how models learn would be super.
How is WM represented in such a model? How do connectionist networks fit in here? There was a paragraph in one of the chapters that said something along the lines of the goal being to get different models to work together instead of to try to create one 'best model'. Can these architectures incorporate dynamical systems, connectionism, &/or computational types of knowledge representation?
Also, I got stuck on a detail… I remember a long time ago using opencyc and finding it to be sort of irritating to use. If cyc includes relations and truths about an object can't these can relations be derived from something like wikipedia where all the articles are linked anyhow? I agree with what Vasile said last week about cyc, especially seeing how much money someone was able to convince someone else to spend on this.Can't this be automated, it seems so complicated?
I'm frustrated by my lack of knowledge. It really is overwhelming what this field has done in such a relatively short time period. This guy summarizes four different computational linguistic endeavors, much more complicated than brushing your teeth, and yet we're still not there, still just explaining. That's a bit painful to think about, but acceptable, because this stuff really is awesome.
WARNING: This following is "out-there-thinking".
I wonder if there's not something in the connections between mental representations, a mathematical pathway, translating a pattern for their relationship. In other words, it may not be just a straight line. Much of these computational models seem to allocate connection as a straight line, but it might be the way the multitude of straight lines eventually form a shape. Google can return a search of 2 words matched with a 106 million links in .09 seconds. That's essentially time travel. But that's just maximizing pattern recognition. What are the literal and figurative meanings in those two words that Google can't prescribe? Couldn't it be in the pathway the information travels? We read the information off a computer screen, but we're still processing it ourselves. From the screen to our mind is a pathway, just as there's a pathway that links sausage to Zelda. Humans are still the best filters/sifters, but even we don't know what we mean, most of the time, sometimes, whenever.
We train computers like we do a child. We instill ourselves in them, telling them to do this, and do that, but we don't really know what we want them to do, but we do know what good is when they produce it. This stuff is good/cool/interesting/useful all the while my rambling proves elementary. But I'm still learning, and hey, kids are fun. This will be useful, and already is, because it's keeping me inspired.
I enjoyed these articles and I found them really interesting, but I was also kind of disturbed to hear that so much effort in the area (ex: Cyc) has turned out not to be useful.
It makes me wonder, what is more important- figuring out how people represent knowledge/concepts, or how knowledge/concepts should be represented in AI/applied fields? I'm not convinced that the best way to go about kowledge acquisition in machines/computers is to try to duplicate the human mind. So far it doesn't seem to be a great strategy. That might be because we still haven't grasped the 'human-way,' or because the human mechanisms aren't the most effective to be used in computers. If the two are or should be seperate (knowledge representation for humans and for computers), should equal time be spent on figuring out both?
Whether or not researchers are looking at these at two seperate entitites, it seems that people don't stop to consider that maybe there is a more efficient way to represent concepts that isn't bogged down by the evolutionary leftovers humans are stuck with.
The “Cognitive Architectures for Conceptual Structures” article was much more interesting to me than the “Conceptual Graphs for Representing Conceptual Structures” because I honestly never really understand graphs. They are supposed to simplify things but I typically feel more confused by them…
The ideas in the “Cognitive Architectures” paper were attractive to me because the obvious attempt to ground representations. However, I am always skeptical of boxes with lines and arrows pointing to and from other boxes as explanations for things. The author did try to provide evidence from neural architectures, which I do appreciate. However, a careful reading revealed that this interpretation of the functioning of the brain is primarily the speculation of the author, and the bibliography at the end left much to be desired. It is entirely possible that I just didn’t really understand it, but I have trouble believing that activity in the brain is so tidy and simple. I am looking forward to class because I would like an explanation for the “boxology” in figure 1.
Finally, I also would agree with Jackie that human models are not the best strategy for improving AI knowledge-acquisiton. Although i would say that knowledge of how humans represent knowledge/concepts is more important.
I think I get the general idea behind conceptual graphs, but some of the details (like figures 14-17 in the conceptual graphs paper) were a bit over my head. I thought the existential graphs were explained much better in the cognitive architectures paper. This made me think of an issue that was discussed last week, which is creating meaning representations that resemble language. How important is it that these meaning representations resemble real language, and do these conceptual graphs do a good job of this?
I got a bit hung up on the example that was used for expressing beliefs ("Tom believes that Mary wants to marry a sailor"). I think that extracting and mapping precise meaning in this case may be especially difficult without any context or background knowledge. For instance, Tom could believe that Mary wants to marry a sailor because that person is a sailor (Mary has a thing for sailors) or that Mary wants to marry a particular individual who just happens to be a sailor (the fact that the person is a sailor is incidental, but in Tom's eyes it is a defining characteristic). Granted, in this example the context does not affect the literal interpretation of the statement. But I think this relates back to issues with representing metaphor and figurative statements, which is mapping underlying meaning seems to be difficult without context or prior knowledge.
Alright, so I might be completely misunderstanding what's going on in the CS cognitive architecture (actually, most likely I am misunderstanding), but it seems to me that a lot of the conceptual graph application to knowledge involves pattern matching to reach an understanding. If this is the case, then how can people quickly understand completely novel experiences? For example, if a completely novel metaphor is said to a person, they can reason out the meaning of the metaphor. But in the conceptual graphs (from my understanding) it seems like when a person is presented with a completely novel metaphor, they would find conceptual graphs for each of the items in the metaphor, but there would not be a conceptual graph that would represent a connection between these items. So how from these conceptual graphs can people quickly respond to and understand novel information?
ReplyDeleteI enjoyed the readings and find the subject very interesting. My first impression was that it was a great representation of language but doesn't it still suffer from the 'translation' problem we discussed last time. How do you map natural language onto these conceptual graphs? Or at least without numerous human experts.
ReplyDeleteThe article clearly was trying to present the case that the conceptual graphing was close to human language processing by presenting the neuro and psycholinguistic evidence. I was impressed but am still skeptical of how well it models humans. The structure seems similar but beyond being able to give the human data a graphical description I did not see much evidence. I would be interested to hear if there was more 'linking' evidence. Another question I have is how well do the CS architectures measure up to human performance on linguistic tasks? I didn't seem to find anything on that comparison in the article just structural.
The other element I found interesting was how well they had been able to cut down on the processing time; because previously it had been my understanding that this was the major limitation to networking/graphical models of cognition. How was this done exactly? Is there potential room for more improvement?
Even then I'd like to bring up the point that I think Jackie mentioned last time; Is a human model necessarily the best method? From an engineering standpoint, a particular solution has to be weighed with factors in the environment and of the problem which it operates: such as cost (eg processing time), 'ergonomic' considerations mentioned last week, or scope of the problem trying to be solved. None of these factors at least to me necessarily implies that a human model is always the best solution.
I am wondering if representing conceptual structures isn't just an exercise. I understand the importance of getting computers to understand richer concepts in language. However, without a system that can learn and build on previous information (i.e., A.I.), making these better representations still can't resolve the complex nature of many statements. The seminar last week spoke of the tremendous amount of resources that went into CYC, and now it's pretty much useless. Undoubtedly, these new graphical representations are impressive and 'more' representative than previous models. But what's going to happen with the next phase of technology or theory, and the next, and so on. Are we destined to just keep getting a little better? Or is there another fundamental aspect on which we can focus? I'm not versed in the areas of computational neural networks, but I'm starting to think that is the area that would be most fruitful. If it's the richness of human thought that's being missed by not being able to re-create natural language, then maybe it's the connections within the system that's more important (i.e., better simulating human neural networks) and not the language or it's representation. (This pains me to say, as language itself is my most favorite thing to study.)
ReplyDeleteCognitive Architectures for Conceptual Structures:
ReplyDeleteSowa first reviewed the four cognitive architectures: Cyc, Soar, Society of Mind, and Neurocognitive Networks. Then he introduced the conceptual structures with the figure of mechanisms of perception, which illustrated the mechanisms of perception draw upon a stock of previous percepts to interpret incoming sensory icons. In addition, he diagrammed how the human brain processes language, which supported the CS hypothesis that semantic-based methods are fundamental to language understanding. In Peirce’s logic and semiotics section, four existential graphs about pet cats were indicated the affirmative and negative relation between cat and pet. Does this mean that all the sentences can be extracted into these four relations?
VivoMind Language Processor (VLP) uses VAE to process syntax, semantics, and pragmatics with analogies. Does this mean all the texts can be processed to obtain their exact meaning? Or is this method useful for the conceptual knowledge, but limited to declarative knowledge? Can the interpretation of the meaning be quantized through VLP like the other distributional models?
The concept of the VCA interests me. Is this essentially Pandemonium, the concept graph version? I was thinking that during the latter half of the cognitive architectures paper, and then the author seemed to wink at the idea with his allusion to order, chaos, and Pandemonium. It's an undoubtedly clever way to handle graph structures, because the number one problem (from my limited understanding) is graph matching and graph traversal, plus the amount of time dedicated to these tasks. It also innately builds in learning by strengthening the associations or agents.
ReplyDeletePerhaps I missed this in the other article, but is VCA still vulnerable to the limitations of logics like we discussed last week? In reading over the article, I didn't see anything crazy like WillBeInterestedToKnow(Dodgers, Play, Tonight) or whatever, so it seems like Peircian concept graphs (natural logic?) limit themselves to assigning words to a vaguer, more defined set of values or relations. Basically, a run down of pros and cons on this sort of organization vs. what was discussed last week would be nice. ;)
As someone else asked about novel concepts… in addition I wonder how do these models incorporate change over time? For example, does the concept of a computer progress as rapidly as a computer itself changes? More details/an explanation on how models learn would be super.
ReplyDeleteHow is WM represented in such a model? How do connectionist networks fit in here? There was a paragraph in one of the chapters that said something along the lines of the goal being to get different models to work together instead of to try to create one 'best model'. Can these architectures incorporate dynamical systems, connectionism, &/or computational types of knowledge representation?
Also, I got stuck on a detail… I remember a long time ago using opencyc and finding it to be sort of irritating to use. If cyc includes relations and truths about an object can't these can relations be derived from something like wikipedia where all the articles are linked anyhow? I agree with what Vasile said last week about cyc, especially seeing how much money someone was able to convince someone else to spend on this.Can't this be automated, it seems so complicated?
I'm frustrated by my lack of knowledge. It really is overwhelming what this field has done in such a relatively short time period. This guy summarizes four different computational linguistic endeavors, much more complicated than brushing your teeth, and yet we're still not there, still just explaining. That's a bit painful to think about, but acceptable, because this stuff really is awesome.
ReplyDeleteWARNING: This following is "out-there-thinking".
I wonder if there's not something in the connections between mental representations, a mathematical pathway, translating a pattern for their relationship. In other words, it may not be just a straight line. Much of these computational models seem to allocate connection as a straight line, but it might be the way the multitude of straight lines eventually form a shape. Google can return a search of 2 words matched with a 106 million links in .09 seconds. That's essentially time travel. But that's just maximizing pattern recognition. What are the literal and figurative meanings in those two words that Google can't prescribe? Couldn't it be in the pathway the information travels? We read the information off a computer screen, but we're still processing it ourselves. From the screen to our mind is a pathway, just as there's a pathway that links sausage to Zelda. Humans are still the best filters/sifters, but even we don't know what we mean, most of the time, sometimes, whenever.
We train computers like we do a child. We instill ourselves in them, telling them to do this, and do that, but we don't really know what we want them to do, but we do know what good is when they produce it. This stuff is good/cool/interesting/useful all the while my rambling proves elementary. But I'm still learning, and hey, kids are fun. This will be useful, and already is, because it's keeping me inspired.
I enjoyed these articles and I found them really interesting, but I was also kind of disturbed to hear that so much effort in the area (ex: Cyc) has turned out not to be useful.
ReplyDeleteIt makes me wonder, what is more important- figuring out how people represent knowledge/concepts, or how knowledge/concepts should be represented in AI/applied fields? I'm not convinced that the best way to go about kowledge acquisition in machines/computers is to try to duplicate the human mind.
So far it doesn't seem to be a great strategy. That might be because we still haven't grasped the 'human-way,' or because the human mechanisms aren't the most effective to be used in computers.
If the two are or should be seperate (knowledge representation for humans and for computers), should equal time be spent on figuring out both?
Whether or not researchers are looking at these at two seperate entitites, it seems that people don't stop to consider that maybe there is a more efficient way to represent concepts that isn't bogged down by the evolutionary leftovers humans are stuck with.
The “Cognitive Architectures for Conceptual Structures” article was much more interesting to me than the “Conceptual Graphs for Representing Conceptual Structures” because I honestly never really understand graphs. They are supposed to simplify things but I typically feel more confused by them…
ReplyDeleteThe ideas in the “Cognitive Architectures” paper were attractive to me because the obvious attempt to ground representations. However, I am always skeptical of boxes with lines and arrows pointing to and from other boxes as explanations for things. The author did try to provide evidence from neural architectures, which I do appreciate. However, a careful reading revealed that this interpretation of the functioning of the brain is primarily the speculation of the author, and the bibliography at the end left much to be desired. It is entirely possible that I just didn’t really understand it, but I have trouble believing that activity in the brain is so tidy and simple. I am looking forward to class because I would like an explanation for the “boxology” in figure 1.
Finally, I also would agree with Jackie that human models are not the best strategy for improving AI knowledge-acquisiton. Although i would say that knowledge of how humans represent knowledge/concepts is more important.
I think I get the general idea behind conceptual graphs, but some of the details (like figures 14-17 in the conceptual graphs paper) were a bit over my head. I thought the existential graphs were explained much better in the cognitive architectures paper. This made me think of an issue that was discussed last week, which is creating meaning representations that resemble language. How important is it that these meaning representations resemble real language, and do these conceptual graphs do a good job of this?
ReplyDeleteI got a bit hung up on the example that was used for expressing beliefs ("Tom believes that Mary wants to marry a sailor"). I think that extracting and mapping precise meaning in this case may be especially difficult without any context or background knowledge. For instance, Tom could believe that Mary wants to marry a sailor because that person is a sailor (Mary has a thing for sailors) or that Mary wants to marry a particular individual who just happens to be a sailor (the fact that the person is a sailor is incidental, but in Tom's eyes it is a defining characteristic). Granted, in this example the context does not affect the literal interpretation of the statement. But I think this relates back to issues with representing metaphor and figurative statements, which is mapping underlying meaning seems to be difficult without context or prior knowledge.