Aaron, On Mon, Apr 1, 2013 at 1:46 PM, Aaron Hosford <[email protected]> wrote:
> Question answering, Jeopardy, and psychiatric applications each have their >> own peculiar requirements. >> > > So far, no one has even suggested a representation that serves this >> disparate assortment of applications. Hence, either you must pick your >> application (as I did in my patent application), or START by proposing a >> representation that serves these disparate applications (which would be >> REALLY valuable if you can figure this out). > > >> In any event, I don't see any possibility of AGI pipe dreams ever >> becoming reality, without some sort of representation that adequately >> serves its claimed goals. Everyone in AGI seems to want to start at the >> front end (parsing) without knowing where they are going. I STRONGLY >> identifying the goal before planning a strategy to achieve it. > > > I have been preaching "representation first" for some time now, but no one > seems to be listening. > This may be a first for the AGI list - that two of us actually agree on something!!! On an earlier version of their English to Russian translator, the Russian Academy of Sciences people found a curious way around representation. I don't know if this is adaptable to AGI, but I thought it worthy of mentioning. They kept a file of over 200 sentence structures. Each new English sentence that came in was compared with them to find the one that fit the best. Then, a parallel Russian sentence structure was selected, and the words were translated one-by-one and placed into their designated places in the Russian sentence structure. This had problems so they moved to an ontological approach, but it served as a possible counterexample that representations were unavoidable. Nonetheless, I think we will be better off with a robust representation. I see a versatile cognitive representational scheme as being a fundamental > to the successful construction of AGI. (How can a system "think" a thought > if it can't represent its meaning internally?) I have a flexible internal > representation for my system which has been refined to handle vagueness, > incompleteness, ambiguity, multiple meanings, puns, contradictory > statements, misstatements, approximations, connotations, different > perspectives, etc., and a parser built around that representation. This > representation is designed from the start with the intention of > facilitating answering questions and otherwise interacting conversationally. > Answering questions is DIFFICULT, more difficult and generally less valuable than finding solutions to problems. The parser converts raw language into this internal representation, > preserving the vagueness, etc., inherent therein, but encoded in a directly > accessible way. The system is not (yet) intelligent, but I have most of the > foundation laid to begin work on that. I think that even without true > intelligence, the availability of an effective representation will make > human language accessible to experimentation with algorithms that > potentially have direct utility towards real world applications. So I fully > agree with your emphasis on representation, but I am the counter-example to > your statement that no one is doing this already. > GREAT. Have you written about your representation, so we can see what this is all about? This sounds REALLY interesting. Steve ================= > On Sun, Mar 31, 2013 at 3:55 PM, Steve Richfield < > [email protected]> wrote: > >> Jim, >> >> On Sun, Mar 31, 2013 at 7:40 AM, Jim Bromer <[email protected]> wrote: >> >>> Steve, >>> I just read the first message in this thread. Yes, successive layers of >>> speech may be necessary to determine what is being referenced. And of >>> course having other information from other IO modalities about the referent >>> would help. But first we have to figure out a way to do it with a computer. >>> >> >> ... whatever "it" is, which is VERY different for problem solving >> applications than for automatic language translation applications. In the >> first, it appears you just want to know which (if any) of many specific >> statements (or implications) of fact were made, and in the second, it >> appears you need to parse, translate the parse structure to the new >> language, and then drop in the translations of the words, after >> disambiguating sufficiently to identify which translation you want. >> >> Question answering, Jeopardy, and psychiatric applications each have >> their own peculiar requirements. >> >> So far, no one has even suggested a representation that serves this >> disparate assortment of applications. Hence, either you must pick your >> application (as I did in my patent application), or START by proposing a >> representation that serves these disparate applications (which would be >> REALLY valuable if you can figure this out). >> >> In any event, I don't see any possibility of AGI pipe dreams ever >> becoming reality, without some sort of representation that adequately >> serves its claimed goals. Everyone in AGI seems to want to start at the >> front end (parsing) without knowing where they are going. I STRONGLY >> identifying the goal before planning a strategy to achieve it. >> >> >>> I think the idea I have in mind can be examined with examples. When you >>> mentioned "payload" I had a pretty good idea about how you wanted to use >>> the word. However, when you started off by saying that computerized text >>> understanding is very similar to patent classification I realized that I >>> did not have a good idea about the methodology of understanding that you >>> were associating with the word. So I inhibited myself from making any >>> assumptions about the words that you were using until I had a chance to >>> reread what you were saying. >>> >>> As I wrote, my new theory of understanding is based on acquiring insight >>> into the specialization of the concepts that are being considered. So, >>> while your use of the term payload to refer to something that I might call >>> the semantic content or the intended meaning was not that different from >>> what I thought you meant, your underlying theory about discovering the >>> meaning was. Yes, if we don't understand something we need to speak about >>> it with different kinds of remarks (or study it in other ways). However, I >>> disagree that these are successive layers where all the details of >>> differentiation of speech can be found on some bottom-level subclass. I >>> just don't think it is that simple. My latest model of comprehension is >>> that the complexity of knowledge is only found in a growing awareness of >>> conceptual specialization (where concepts are either based on a group >>> of shared concepts or from a process of observation and conjecture tied to >>> some personal concepts. In other words we can build higher models by >>> communicating or by using our imaginations.) In my opinion, the basis of >>> differentiation or specialization will not be found in some bottom-level >>> subclass but in examining some construct of thought using different ways to >>> think about it. >>> >> >> My point (in regard to this) is that there is a large family of >> bottom-level constructs that affect this, some combination of which will >> guide differentiation, specialization, disambiguation, etc. By selectively >> triggering the consideration of higher-level analyses based on the presence >> of relevant bottom-level constructs, you have SOME chance of being able to >> understand a full blown language definition in real time on modern-day >> computers. Without this or some adequate substitute for it, you will fall >> into the same performance trap that has consumed all prior attempts. >> >>> >>> I now have a simplified model of what you were talking about in my >>> mind. I can generalize this. I see you believe that there is a hierarchy >>> of detail which would reveal the specialized meanings of words. I will >>> remember this about you and look for it in other ideas that you talk about >>> and I will look for in other people's ideas. (I disagree with the single >>> hierarchy, a general hierarchy of differentiation and the bottom-level full >>> of details. To my thinking these are metaphors which are standing in as >>> substitutes for effective methods.) >>> >> >> You wouldn't think a modern gigahertz computer could be trapped by lowly >> text, but when you understand this, you can see how methods must be crafted >> to fit the machines at hand. >> >>> >>> I did not have a substantial disagreement with most of the other things >>> you said in this, the first message of this thread. The fact that you >>> mentioned that an inability to precisely understand what someone said might >>> lead to a mistaken misinterpretation of ignorance was a confirmation of my >>> opinion that you can be open-minded and that you are able to use this >>> ability to discover a context of misunderstanding in ways that some people >>> are not. >>> >> >> More to the point in DrEliza.com, some of its analysis is to identify >> where the user is ignorant in important ways that can be remedied with a >> well-targeted explanation. >> >> >>> By being open minded one can see possibilities that closed-minded people >>> may miss. >>> >> >> Often all it takes is deeper domain-specific knowledge, part of which is >> understanding popular misconceptions. >> >> >>> This is an example of a conceptual specialization and it can be tied >>> into the meaning of language even when the language is not about the >>> subject of being open close minded. >>> >> >> YES. People presume that DrEliza.com is there to drill down into their >> illnesses, and it is. The drill-down process is often impaired by common >> misconceptions, that are MUCH easier to squarely address, than to leave in >> place and try to work around. >> >> Steve >> ============================= >> >> On Thu, Mar 28, 2013 at 12:27 AM, Steve Richfield < >> [email protected]> wrote >> >>> *Jim, et al,* >>> >>> *I'm starting a new thread with this...* >>> >>> It is my theory that computerized speech and text understanding has >>> eluded developers for the past ~40 years, because of a lack of a >>> fundamental understanding of the task, which turns out to be very similar >>> to patent classification. >>> >>> When classifying a patent, successive layers of sub-classification are >>> established, until only unique details distinguish one patent from another >>> in the bottom-level subclass. When reviewing the sub-classifications that a >>> particular patent is filed within, combined with the patent’s title, what >>> the patent is all about usually becomes apparent to anyone skilled in the >>> art. >>> >>> However, when a patent is filed into a different patent filing system, >>> e.g. filed in a different country where the sub-classifications may be >>> quite different, it may be possible that the claims overlap the claims of >>> other patents, and/or unclaimed disclosure would be patentable in a >>> different country. >>> >>> Similarly, when you speak or write, in your own mind, most of your words >>> are there to place a particular “payload” of information into its proper >>> context, much as patent disclosures place claims into the state of an art. >>> However, your listeners or readers may have a very different context in >>> which to file your words. They must pick and choose from your words in an >>> effort to place some of your words into their own context. What they end up >>> placing may not even be the “payload” you intended, but may be words you >>> only meant for placement. Where no placement seems possible, they might >>> simply ignore your words and file *you* as being ignorant or deranged. >>> >>> Many teachers have recorded a classroom presentation and transcribed the >>> recording, only to be quite surprised at what they actually said, which can >>> sometimes be the opposite of what they meant to say. Somehow the class >>> understood what they meant to say, even though their statement was quite >>> flawed. When you look at these situations, the placement words were >>> adequate, though imperfect, but the payload was okay. Indeed, where another >>> person’s world model is nearly identical to yours, very few placement words >>> are needed, and so these words are often omitted in casual speech. >>> >>> These omitted words fracture the structure of around half of all >>> sentences “in the wild”, rendering computerized parsing impossible. Major >>> projects, like the Russian Academy of Science’s Russian Translator project, >>> have wrestled with this challenge for more than a decade, with each new >>> approach producing a better result. The results are still far short of >>> human understanding due to the lack of a human-level domain context to >>> guide the identification and replacement of omitted words. >>> >>> As people speak or write to a computer, the computer must necessarily >>> have a *very* different point of view to even be useful. The computer >>> must be able to address issues that you can not successfully address >>> yourself, so its knowledge must necessarily exceed your own in its subject >>> domain. This leads to some curious conclusions: >>> >>> 1. Some of your placement words will probably be interpreted as >>> “statements of ignorance” by the computer and so be processed as valuable >>> payload to teach you. >>> >>> 2. Some of your placement words will probably refer to things outside >>> of the computer’s domain, and so must be ignored, other than being >>> recognized as non-understandable restrictions on the payload, that may >>> itself be impossible to isolate. >>> >>> 3. Some of your intended “payload” words must serve as placement, >>> especially for statements of ignorance. >>> >>> My invention seeks to intercept words written to other people who >>> presumably have substantial common domain knowledge. Further, the computer >>> seeks to compose human-appearing responses, despite its necessarily >>> different point of view and lack of original domain knowledge. While this >>> is simply not possible for the vast majority of writings, the computer can >>> simply ignore everything that it is unable to usefully respond to. >>> >>> If you speak a foreign language, especially if you don’t speak it well, >>> you will immediately recognize this situation as being all too common when >>> listening to others with greater language skills than your own speaking >>> among themselves. The best you can do is to quietly listen until some point >>> in the conversation when you understand enough of what they are saying, and >>> you have something useful to add to the conversation. >>> >>> Note the similarity to the advertising within the present Google Mail, >>> where they select advertisements based upon the content of email that is >>> being displayed. Had Google performed a deeper analysis they could probably >>> eliminate ~99% of the ads as not relating to users’ needs and greatly >>> improve the users’ experience, and customize the remaining 1% of the ads to >>> precisely target the users. >>> >>> That is very much the goal with my invention, where the computer knows >>> about certain products and solutions to common problems, etc., and scans >>> the vastness of the Internet to find people whose words have stated or >>> implied a need for things in the computer’s knowledge base, and have done >>> so in terms that the computer can “understand”. >>> Steve >>> >> *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> >>> <https://www.listbox.com/member/archive/rss/303/10443978-6f4c28ac> | >>> Modify <https://www.listbox.com/member/?&;> Your Subscription >>> <http://www.listbox.com> >>> >> >> >> >> -- >> Full employment can be had with the stoke of a pen. Simply institute a >> six hour workday. That will easily create enough new jobs to bring back >> full employment. >> >> *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> >> <https://www.listbox.com/member/archive/rss/303/23050605-2da819ff> | >> Modify <https://www.listbox.com/member/?&;> Your Subscription >> <http://www.listbox.com> >> > > *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> > <https://www.listbox.com/member/archive/rss/303/10443978-6f4c28ac> | > Modify<https://www.listbox.com/member/?&;>Your Subscription > <http://www.listbox.com> > -- Full employment can be had with the stoke of a pen. Simply institute a six hour workday. That will easily create enough new jobs to bring back full employment. ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
