|
This is very interesting. This semester I am (B.A. student) writing a course paper on philosophy of language. Kind of taking a stance against Searle's Chinese Room thought experiment, I want to argue that a sufficiently developed "program" loaded into a rather "idle" hardware, could indeed perform mental processes that we humans are performing [1]. As part of linguistics debate, I also argue that natural language is a tool for "thinking"; everything we can think of, we can find expressions for (some people might be better or worse), but surely we cannot "think" of something for which we have no expressions for. At the end, everything falls into "something". In a way, our human language a perfect tool for describing this ambigious physical world which we experience through our senses and try to make sense of. Thinking through these ideas with natural language, no coincidence that I tried a bit to find out; what should AI really look like? I mean its computer design? Around this time, I arrived at the following thoughts, and this paper really strengthened my reasoning: I think that this "program" should be a combination of first order logic for reasoning tasks, and neural networks for any problem on which NNs are infamously good at, and as an answer to the holy grail of "talking computers" question, our "program" should have a "bridge" between formal logic reasoning tasks VS natural language, whose ambiguity makes it difficult. I studied a bit of computer linguistics in my bachelor recently, and "function(ambiguity) = first order logic" seems possible to me if we employ a number assumptions and linguistic tools. If 'it' converts natural language into formal 'action' statements and/or first order logic statements etc, then 'its' intentions could be inspired by human speech. In that regard, I agree with the paper about that it should be capable of "formal proofs", which I interpret as first order logic in its essence. I was searching "first order logic python or logic programming with python" on google, and to my surprise all the titles say "AI programming with Python", as it turns out that first order logic programming is a whole ass programming paradigm and researchers since the 60's have been developing languages like LISP or Scheme to precisely do that. I discovered a python package called minikanren, which as I understood adds LISP-like logic programming capacity to Python. It is a bit complicated but I am trying to understand it. I believe with current NLP tools in python and a logic checker introduced with kanren, we can kind of easily write a program that understand natural human speech? Just bear with me here: Let's pick some hard examples. And this is straight from wikipedia: https://en.wikipedia.org/wiki/John_Searle#Speech_acts According to Searle, the sentences...
Sam smokes habitually.
Does Sam smoke habitually?
Sam, smoke habitually!
Would that Sam smoked habitually!
... each indicate the same propositional content (Sam smoking habitually) but differ in the illocutionary force indicated (respectively, a statement, a question, a command and an expression of desire)
Philosophers of mind and language has identified many such illustrative example of natural speech. In the example above, we can easily 'parse' this sentence to its grammar constituents (i.e. part of speech tagging, dependency graph etc.) with simple Python tools. Then using a small bit of magic of transformative grammar rules [2], we can extract the initial sentence (sam smokes habitually), now, we can also seperate this atomic statement from its illocutionary vector, having gotten this knowledge acknowledged by our 'program'.Some immediate concerns come to my mind is that it is difficult for a program to still 'grasp' what it means to say 'Sam smokes habitually'. Even this is a tremendously difficult problem, it seems. Albeit it is a simple answer, I must simply say that we can add a background knowledge for basically everything we can. For example Sam could be anybody or even a dog. But if Sam smokes then he must be a human, because smoking entails certain conditions, which we can describe many, since readers are also humans, we can skip it. Since we know what smoking entails, and since Sam smokes, we can then conclude that "Sam is a man and Sam indeed smokes", and we can further extract the information that "sam actually smokes habitually", which a "program" which also understand it with its "entailments", that he smokes every now and then. Such that, if someone ever asks our 'program', would you expect Sam to smoke right now, and if the 'program' has recently observed that Sam has smoked, 'it' might give out answer such as "No, but he may in a bit", having also a pre-disposition that "if ask(user, yes_no_question) and if (answer == no) then; say('but' + answer_for_when_yes)" such that it would additionally say ".. but he may in a bit", rather than a cold resounding "No." answer. I believe these are the stuff happening in our brains and we could try to simulate them with some bold assumptions and see what happens? We basically have mental images in our heads but they only gain their true meaning for us when we give names to them - so without language I don't believe we are much further than animals, and I hope to believe that this is more or less what mainstream thinks anyways. But I guess it is never easy to be sure of how our brains work. I am curios what other's think. I am still very young and discovering all these early works people have been doing. It is so surprising that most of it are super new, and it makes it so much more exciting too. It is a shame that symbolic AI did not work in 60s, I guess they just did not have compute power to calculate the complexity of reality. But with current computation power, all ambigious problems seems to be dominated, from vision to seemingly "sound" language production (i.e chatgpt). So don't you guys think that we should indeed have a paradigm shift back to a symbolic programming empowered with Neural Networks? [1] I can't say whether that hardware + program would actually be conscious, since I can't define it myself. But to me, all human thought processes seems to be reducible to very certain first order logic statements. [2] Transformative Grammars: quite well known linguistic exercise, in which you convert a sentence to something else without changing its meaning, for example: Dog eats food == Food is eaten by Dog - each sentence has similar part of speech tags except their dependency graph is different, etc. etc. Chomsky and others gaves us all those rules |