Matt said: The overview claims to be able to convert natural language sentences into Cycl assertions, and to convert questions to Cycl queries. So I wonder why the knowledge base is still not being built this way. And I wonder why there is no public demo of the interface, and no papers giving verifiable experimental results.
>From my employment at Cycorp, 1999-2006, I can answer your questions, and also >say why I am pursuing a natural language approach with Texai, extending the >OpenCyc ontology. First understand that Doug Lenat is a mathematician, not a >linguist nor cognitive psychologist. The Cyc Project began, in the 1980's, >hand-entering knowledge in a predicate calculus schema that was intentionally >agnostic with respect to natural languages. To this day, increasingly >sophisticated tools have enabled Cycorp's ontologists to more precisely and >more rapidly enter knowledge and perform queries, than they can by relying on >rather incomplete and relatively poorly performing English language tools. >For example, I witnessed some very long parsing times for input sentences over >15 words. In contrast, new concepts can be defined and positioned in the >existing ontology using point and click, non-NL screens very rapidly if the >ontologist is fully prepared with respect to what they want to accomplish - for example using the "Create Similar" tool. Addressing the public demo, I believe that first Cycorp does not want to expend the considerable effort to create and maintain a publicly accessible NL interface of high quality, when there are so many other areas of Cyc that sponsors are paying for that need attention. Secondly, I believe that because Cyc is proprietary, it precludes a public NL interface that enables the proprietary content to be extracted. Furthermore, the effort required to create some reasonable, but small example public partition would be prevented by my first observation. During my tenure, Cycorp had no fewer than three PhD computational linguists continuously employed on NL interfaces. But I believe their progress has been blunted by the need to maintain a large number of legacy parsers, all of which have some dead-end (i.e. not cognitively plausible) characteristics. Furthermore, the Cyc NL parsing and NL generation systems are at least two completely different bodies of code. Therefore Cyc is not capable of understanding all of what it can say and vice versa. Moreover, Cyc's NL system, similar to its other long-standing code components, is quite large, and sadly demanding that the great majority of the developer's time is spent maintaining, fixing, migrating, rewriting and tailoring the existing code rather than adding new functionality. In previous posts here and on my blog, I have described the Texai system as an English dialog system to achieve AGI via bootstrapping a small code base. By extending OpenCyc's ontology, and in particular biasing it towards the semantics of English language constructions, I hope to avoid some of the problems I saw at Cycorp. Matt also said: It seems to me the main limitation is that the language model has to be described formally in Cycl, as a lexicon and rules for parsing and disambiguation.t seems to me the main limitation is that the language model has to be described formally in Cycl, as a lexicon and rules for parsing and disambiguation. Agreed. For the Texai language model, I employ Fluid Construction Grammar as the encoding, and Double-R Theory for its grammatical constructs. Although I currently hand-write these in an symbolic-expression external format, they are serialized into RDF from corresponding Java rule objects, and stored in the Texai KB as any other assertion. My plan is to task the dialog system first to interact with its mentors to acquire new vocabulary (e.g. mappings from word senses to Cyc concepts, argument mappings to event roles, etc.) and new grammar constructions (e.g. "on the table" as a phase can have as one of its senses an instance of cyc:Negotiating). Those wanting to know more about Cyc should attend my Cyc tutorial at AGI-09. Or you can download OpenCyc whose first release I lobbied for, and then created with John DeOliveira while at Cycorp. -Steve Stephen L. Reed Artificial Intelligence Researcher http://texai.org/blog http://texai.org 3008 Oak Crest Ave. Austin, Texas, USA 78704 512.791.7860 ----- Original Message ---- From: Matt Mahoney <[EMAIL PROTECTED]> To: [email protected] Sent: Sunday, September 28, 2008 8:38:36 PM Subject: Re: [agi] universal logical form for natural language --- On Sun, 9/28/08, Ben Goertzel <[EMAIL PROTECTED]> wrote: >FYI, Cyc has a natural language front end and a lot of folks have been >>working on it for the last 5+ years... It still needs work. I found this undated (2004 or later) white paper which is apparently not linked from cyc.com. http://www.cyc.com/doc/white_papers/KRAQ2005.pdf And also this overview. http://www.cyc.com/cyc/cycrandd/areasofrandd_dir/nlu The overview claims to be able to convert natural language sentences into Cycl assertions, and to convert questions to Cycl queries. So I wonder why the knowledge base is still not being built this way. And I wonder why there is no public demo of the interface, and no papers giving verifiable experimental results. It seems to me the main limitation is that the language model has to be described formally in Cycl, as a lexicon and rules for parsing and disambiguation. There seems to be no mechanism for learning natural language by example. For example, if Cyc receives a sentence it cannot parse, or is ambiguous, or has a word not in its vocabulary or used in a different way, then there is no mechanism to update the model, which is something humans easily do. Given the complexity of English, I think this is a serious limitation with no easy solution. -- Matt Mahoney, [EMAIL PROTECTED] ------------------------------------------- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?&; Powered by Listbox: http://www.listbox.com ------------------------------------------- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
