Readings
Diana Applegate
I found this week to be quite enjoyable. I'm having much more fun with FOPC than I had with Propositional Calc. I especially liked the SNePS demo. It's always great to see practical applications of this stuff. I think that people unfamiliar with AI and its logic capabilities would be amazed to see something like SNePS in action. It's sort of shocking in a way that a computer can arrive at the same sorts of conclusions as we can given a certain situation...however, it does make sense since we're the ones doing the programming. Nevertheless, I think that implementing logic systems in machines is still a great feat of sorts for AI. On another note, I recently read Dennett's essay on "Cognitive Wheels" that can be found in his book, Brainstorms. He discussed the frame problem in AI and also got into a brief discussion of FOPC. It's great to see such an overlap between this course and my philosophy course. I also came upon an interesting website this week: http://come.to/20Q
Someone has implemented an online knowledge base that grows and "learns" by playing a game of 20 questions with the user. At this point, it seems to be overwhelmed with new knowledge and is therefore not playing the game too accurately. However, it's interesting to interact with this system. There's also a detailed description of the algorithm used on the site, and some talk about neural networks. Check it out if you're interested. Finally, I enjoyed the reading about natural language. I first became interested in this area of cognitive science over the summer at UPenn. I also took an intro linguistics class at Swarthmore last year where we learned about grammars, lexicons, structure trees and how to identify/diagram semantic ambiguities (among other things). For a good overview of this particular field (and a fun read), I'd recommend Pinker's The Language Instinct.
See everyone in lab! Good luck with Konane!
Jocelyn Arcari
The more examples we do, the better I am understanding substitution, unification and resolution. Also, seeing SNePS in action helped understand the different processes that go on in applying logic. Do you have another example of converting predicate calculus wffs to clause form that we could go over in class that is more complicated than the one we did in class the other day? More practice with this would be helpful.
The article we read, "Introduction to Natural Language Understanding" was interesting. The review of grammar and English Syntax was a refresher from my linguistics class for things such as "morphology," open/closed class words, and problems that we run into when trying to compose sentences that are syntactically, and semantically well-formed and that make sense. There are so many intricacies that one must consider when thinking of creating a text-based natural language system. Revisiting ELIZA was also fun to see the implications of a program like that on an initial level versus going deeper into the program and finding glitches. Among the different dialogue-based applications that were considered, we have seen many of these already put into action -- querying a database is certainly possible, automated customer service on the telephone (for train schedules and fair prices is a common example), tutroring systems where the machine interacts with the student is available now (the one I saw was for learning a software program), etc. I would think that the help of linguists would be very important in trying to create advancements in many of these areas.
Finally, testing people's konane projects in lab will be fun and interesting this Monday. I'd be interested to see how programs differ from each other and from the suggested model, if at all.
David Costello
The way you describe propositional logic one would think that all forms of knowledge could be represented in this format. However, you also stated in the beginning of the semester that AI first started with GOFAI and branched out into other areas like neural nets and genetic programming. Why would some scientists abandon logic as a means to artificial intelligence when the method can represent so much through deduction? In other words, what are the drawback of logic which forced AI research into other fields? Also, can there be a propositional logic agent which learns its wffs through its sensors instead of through human programming or is sensory input too complicated to create accurate wffs?
Emily Greenfest
I found the reading assigned for this past week to be rather interesting as it addressed an aspect of language that I spent much time considering. Issues such as the evolution of language and the possible relationship between intelligence and language (as well as the core issue: what is a language) have been brought up repeatedly in my numerous evolution-related classes, but I have really never considered all of the bits and pieces that comprise a language. Maybe that's because I don't ever recall learning most of them. True, at some point I learned how to distinguish a verb from a noun or an adverb, but, in truth, that was not anything new: I had "known" what a verb was, I just didn't know what to call it. But at what point did I understand the semantics of the language? Or did I just take bits and pieces, the alphabet, the few words that I knew, and the few sentences I had seen repeatedly in literature being read to me (or words being spoken) and just intuitively understand it (i.e. is it something that is taught or something specific to our species?).
It appears that the AI and related work being done concerning language comprehension thinks that language comprehension is something that can be taught (and perhaps that is true, perhaps the only specific aspect of it is its production ... and as AI in part attempts to mimic/model the human being there is no reason that a future AI could not do both). To teach a computer language, however, language must be converted into a form that the computer can understand -- and I belive that this problem comprises much of the work discussed in the reading.
Ayishih Hakim
The hand out that you distributed in class about Natural language is amazingly interesting. I enjoyed reading about the difficulties of creating a system that could "understand" natural language. Reviewing the technicalities the the handout explained it seems that a system such as this is very very hard to impliment.
The comparison of text based and dialog based applications was alos very interesting. I never thought about the search engines that we use on the internet or the searches that we perform when are using library based searchers as examples of a natural language recognition application system. I understand that these such systems are far from perfect and cannot "understand" articles or books or subjects but they do have a good system at a simple recognition. The reading was pretty straight forward and the information was clearly expressed thus I have no questions.
Ada Hogan
As we were looking at SNePS in class on thursday, with the job descriptions "nurse, cook, boxer", I began to wonder about the kind of information that a programmer might eventually want to include in a knowledge base of an agent that acts and makes decisions in a real world environment. In the examples we saw, the program worked from information it was given in the problem; if it had additional information, would it be limited to hard facts, or could it also include stereotypes and prejudices? This might seem like an odd question, but as I tried to figure out the question myself, I noticed that you could easily attach certain names with certain professions. I might call "Joe" a boxer before "Marlene", for example. This may be personal, but it is very much a part of how we reason and think. If agents developed in AI are ever expected to operate in the real world, and "aproach human performance" in "language comprehension and production", it seems that this would be a real issue. Our actions and speech are not based on hard facts; an agent that is thus modeled is not really imitating human patterns of interaction.
I thought the chapters by James Allen were clear and helpful; while we were studying predicate calculus l was wondering how the "inference engine" would interpret words with synonyms. The section on Logical Form and context-independent meaning and context-dependent meaning, as well as his "structural representations" answered this exactly.
I understand the reasoning behind the order of parsing first, interpreting the problem in context, and then formulating the language of the response- is this also supposed to imitate the human pattern of gathering information and processing it?
Edina Sarajlic
In terms of the material covered in class, unification and resolution in predicate calculus do not seem much more difficult than the resolution in propositional calculus. However, I think it would be really helpful if we could get a couple of practice problems that we could do on our own and check the solutions in some answer key. Nilsson has some interesting exercises that might be a useful preparation for the exam, but he does not include any answers that I could check my work against.
I found the section on evaluation of language understanding systems in this week's reading very interesting. I understand how the black box evaluation can be erroneous in the early stages of the project development, so that some kind of glass box evaluation would be necessary for proper evaluation. However, Allen does not give any further explanations about actually performing the glass box analysis. (Are there any general standards for the design of these systems, which would be used in the evaluation?) For example, Allen claims that basing a language translation system on pattern matching can produce very good results, but has serious downfalls that might not be immediately evident in testing. He argues that since the system is not capable of understanding the content, it cannot properly deal with the natural ambiguity of the natural languages. I find this somewhat confusing, because I do not know what definition of understanding is used in this case. What kind of symbol manipulation constitutes understanding (at least in the context of this reading) ?
I hope that we will learn more about standard natural language understanding techniques in class, because they seem to synthesize and apply a lot of the material we covered in the last couple of weeks.
Ben Sprecher
I found the reading on Natural Language Understanding quite interesting. It seems like understanding language would be the ultimate synthesis of everything we've done to date in AI. True language understanding would require a knowledge base comparable to the average language-speaker (i.e. human), the ability to translate sensory inputs into feature vectors (words and sentences), and to relate the information gleaned from the input to the information contained in the knowledge base. It seems that when we can do this we will have a convincing argument in favor of the success of AI in reaching its goals.
Understanding language raises many interesting questions. One is whether, in order to understand language, we must have some universal, unambiguous structure onto which we can map the "true meaning" of any concept that can be expressed through language. If so, then it would be a fascinating endevor to come up with such a system and to implement it on a computer. However, it is not clear that such a structure even exists. I question whether we ourselves don't hold opinions or maintain concepts which are inherently ambiguous. Ideally, we wouldn't, but i doubt that that is the case. Take, for example, our concepts of identity, soul, mind, and happiness. Representing these concepts concretely has been the unsuccessful goal of philosophical thought for centuries. I therefore very much doubt that I could sit down and design a data structure to hold any possible human concept unambiguously.
I am looking forward to further reading on language understanding.
Tim waring
My konani is finally finished. With the exception of the fact that it does not know of the move of multiple jumps, it plays perfectly, and it has beaten me. It is a better player than I, even if i stare at the screen for the same amount of time it takes to churn through the tree.
i have not done the reading this week, for the reason that i couldn't find what it was on the web site, and figured we haven't moved to the next chapter yet.
I am excited and anxious for the tournament, it should be informative and fun.
Leslie Zavisca
I really enjoyed the reading for this week. It was a nice break from Nilsson. I haven't had the opportunity to take any linguistics courses yet so I found the second chapter especially interesting and thought provoking.
A few quick questions: How exactly do search engines on the internet operate and what makes Excite, Yahoo, Infoseek, etc different from one another? Also, what is the date of publication on this book?
While reading the second chapter, I caught myself thinking a lot on the ways that dogs communicate with each other and with human beings, sometimes very specifically. My sister is a dog lover and owns a few books on raising/training/understanding dogs. I remember flipping through one and reading about a man who had trained hundreds of dogs and taught all of them over two hundred words and commands. He even got to the point that the dogs he trained understood the difference between words like "puppy" versus "puppies". Of course, the dogs could not speak English back to him, but they understood his commands quite thoroughly and responded appropriately. I was also thinking about how dogs and other animals communicate with each other using body language. For example, bearing teeth, arching the back, wagging the tail, cocking the ears, etc. I wonder where all of this falls in syntactic processing and how our own body language relates to our spoken/written languages. Has any work been done in ai relating to body language? What is the closest that ai has come to producing a computer that can understand language?