[Attempto] Broader the range of texts that can be mapped to ACE

[Pierre-Alexandre Voye]

I would like to replace the debate by introducing a thing Jean-Marc forgot
to specify (we discussed about it) : A machine learning algorithms would
only be an assistant and never a full automatic way to translate natural
English to ACE.

Because I'm a true lazy person, I often use google translate to translate
french text into english. But I also know that the automatic translation
isn't very good and misinterpretation isn't far, even common with this tool.
So I always carefully review the text produced by the software, so it make
the big part of the job, I just have to fix its misinterpretation.

I think the tool Jean-Marc proposes would work in the same way : a tool
which make the major exhausting work which would of course have to be
totally reviewed.
Of course this tool would produce 50% faulty sentences, but 50% of the job
would be done. And has I know now how it is difficult to express what you
want in ACE grammar, it would be a great help.

Because it's a machine learning system, the system would be trained more
and more, so it would become better by the time.


2013/1/28 Norbert E. Fuchs <fuchs at ifi.uzh.ch>

>
> On 27 Jan 2013, at 23:53 , Simon Spero <sesuncedu at gmail.com> wrote:
>
> > On Sun, Jan 27, 2013 at 12:15 PM, Norbert E. Fuchs <fuchs at ifi.uzh.ch>
> wrote:
> >
> > In my view, natural English and controlled English are conceptually
> different languages that need a human translator.
>
> This sentence – seen in isolation – could be misunderstood. Perhaps I
> should have been more careful when I wrote it.
>
> In the context of Jean-Marc's message
>
> > ... Would that make sense to train a text machine learning algorithm
> witt texts in natural english and their translation in ACE? ...
>
> and the rest of my message the sentence was meant as
>
> In my view, natural English and controlled English are conceptually
> different languages and you need a human translator for the translation
> from natural English to controlled English.
>
> > ... I would claim that any controlled version C of a natural language L
> must be a subset of L,  such that every sentence in C is a sentence of L;
>
> I would write "grammatical subset" since the semantics may differ. The
> sentence "A man sees a girl with a red skirt." does have different meanings
> in English and in ACE.
>
> However, to make the grammatical rules of a controlled language
> consistent, you may end up violating the pragmatics of the respective
> natural language. An example is ACE's rule "Phrasal particles and those
> prepositions that introduce a complement of a transitive verb, must be
> hyphenated to the verb." Thus both "John turns-off the light." and "There
> is a light. John turns-off it." are accepted by the ACE parser. The first
> example obeys English pragmatics, the second not.
>
> > that every sentence in C has a single parse,
>
> Though ACE texts have a single parse, there are other controlled subsets
> of English that allow more than one parse.
>
> > and that this parse must be a possible parsing of the sentence in L.
>
> Sure.
>
> > ... A designer of a controlled version of a natural language is
> constrained by the grammar of the natural language; they  are not free to
> introduce new grammatical constructs, etc.,
>
> This depends on the application you want to use the controlled language
> for. In one instance we considered extending ACE by URLs and name-spaces
> which people might disagree as belonging to the grammar of English.
>
> > should recognize/generate as natural a set of sentences as possible, but
> must avoid ambiguity.
>
> Again, there are controlled languages that allow for ambiguity.
>
> > ... texts in ACE are understandable by speakers of English;
>
> Grammatically yes, semantically not necessarily as the above "red skirt"
> example shows.
>
> > however, these speakers need some training or guidance to produce
> correct ACE.
>
> We replaced the grammatical training by guidance in the predictive ACE
> editor and in AceWiki, though some training of the semantics is still
> needed.
>
> Regards.
>
>    --- nef
>
>
>
> _______________________________________________
> attempto mailing list
> attempto at lists.ifi.uzh.ch
> https://lists.ifi.uzh.ch/listinfo/attempto
>



-- 
---------------------
https://twitter.com/#!/ontologiae/
http://linuxfr.org/users/montaigne
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.ifi.uzh.ch/pipermail/attempto/attachments/20130128/1ef7b5ea/attachment.html>