John,
Thanks for your further comments:
JS
> That's the ultimate goal of AI:
>
> PJ
> > Tala is not a CNL, because its goal (and to some extent, current design)
> > is to support unconstrained NL syntax and unconstrained semantics.
>
> But the goal of *supporting* unconstrained syntax and semantics,
> does not require that the notation(s) used to support it need to be
> unconstrained in the same way.
>
> For example, our VivoMind tools support unconstrained NL with conceptual
> graphs. The syntax of CGIF (Conceptual Graph Interchange Format) is
> tightly constrained by the ISO standard for Common Logic. The semantics
> (the definitions of the labels on those graphs) is not constrained by
> the CL standard. But there are various levels of constraints that can
> be imposed on a CG or collection of CGs [...]
Understood. This is what I meant by writing "many have viewed the problem as providing a mapping from NL syntax into some other language, or different set of computable forms, e.g. predicate calculus, conceptual graphs, frames, etc."
I do not claim the goal of supporting unconstrained NL syntax and semantics, *requires* that the notation(s) used be unconstrained in the same way. Most AI research on NL understanding has attempted to translate NL into a syntactically constrained formal language, and to support NL semantics within it. I would agree a lot of progress has been achieved with such approaches, and at least in theory, they could eventually achieve human-level understanding of natural language, and support human-level AI.
However, as you've noted, we are still very far from achieving an AI system that can understand natural language with anything approaching human-level intelligence, even matching a five-year old child. Since this is a major problem for achieving human-level AI, my thesis reconsiders McCarthy's 1955 proposal to develop an artificial language with properties corresponding to those of English.
To that end, I hypothesize that the goal of supporting unconstrained NL semantics *permits* using a notation with a syntax corresponding to unconstrained NL syntax. In effect, I suggest that natural language syntax is often the best way of (at least initially) representing natural language semantics.
I use the term "TalaMind" to refer to approaches that explore this hypothesis, and the other two hypotheses of the thesis, and to an architecture described for such systems.
In this approach, most reasoning about natural language semantics at least initially occurs using Tala conceptual expressions. This permits using other notations, such as conceptual graphs: I do not claim natural language syntax is always the best way of representing natural language semantics, nor of modeling information about the world in general. Yet natural language syntax seems to be the best first approximation, and gateway, to representing information about the world.
> PJ:
> > Your slides give the following definition:
> > CNL = A subset of a natural language that has a well-defined mapping
> > to and from a computable form.
>
JS:
> I tried to make that definition as loose as I could without being
> hopelessly vague. I consider CNLs to be similar to Wittgenstein's
> games: give examples and say "These and anything similar are CNLs."
> A restrictive definition would stifle innovation, not promote it.
>
Of course, I don't wish to stifle innovation, just to avoid confusion. The word "controlled" is embedded in the term CNL, which seems to imply a constrained representation of NL syntax / semantics. Hence, I objected to calling Tala a CNL.
If memory serves, the original use of the term CNL was more in accordance with the definition I gave: A constrained subset of NL syntax to support representation of a constrained semantic domain.
Rather than redefine the term CNL, I'd propose using Fred Thompson's term "sublanguage", to refer to a subset of a natural language used for a specific semantic domain.
> >> Why haven't these great ideas become the foundation for the Semantic
> >> Web? Should they be considered? If not, why not? If so, how?
> >
> > Perhaps one answer to the first question is that Berners-Lee (1998)
> > specifically disclaimed "A Semantic Web is not Artificial Intelligence".
>
> Yes, but... If you look at his DAML proposal (2000), he cites a wide
> range of AI systems and researchers. [...] The
> disputes were not between AI and non-AI methods. They were disputes
> between different AI factions.
>
> I suspect that the primary reason for Tim's statement was political:
> The non-AI members of the W3C had a bad impression of AI based on
> many years of unfulfilled promises (hype) from the AI community.
This seems likely.
I wonder to what extent Thompson's sublanguages could be supported, with the technologies developed for the Semantic Web. (?) His 1992 proposal for telephone-computers (to support a worldwide network of sublanguages) was published in parallel with the global public emergence of Berners-Lee's WWW. Perhaps by the time the Semantic Web was proposed in 1998, most people had forgotten about Thompson's proposal, if they were ever aware of it.
> > Regarding how to develop these ideas going forward, my recommendation
> > is that people work on developing systems according to the TalaMind
> > approach.
>
> If, as I suggested, you adopt a loose definition of CNL, that could
> include a large family of methods -- ranging from Ed Lowry's PROSE,
> Fred & Bozena Thompson's ASK, the VivoMind methods outlined above,
> to many other systems that have been developed in AI.
>
Yes, but this large family of methods would not correspond to my use of the term "TalaMind", to refer to the approach described above. The TalaMind approach appears to be different, for the reasons discussed above.
Phil