Peter Kirby

See here:

http://adios.tau.ac.il/

I decided to run the algorithm (ADIOS-lite, the freeware) on First John (Robinson's Byzantine transliteration). These are some of the computer-generated sentences I got.

EAN EIPWMEN OTI AMARTIAN OUK EXOMEN EAUTOUS PLANWMEN KAI H ALHQEIA OUK ESTIN EN AUTW

TIS ESTIN O YEUSTHS EI MH O ARNOUMENOS OTI IHSOUS OUK ESTIN O XRISTOS OUTOS ESTIN O ANTIXRISTOS O ARNOUMENOS TON PATERA KAI TON UION

TIS ESTIN O NIKWN TON KOSMON EI MH O PISTEUWN OTI IHSOUS ESTIN O UIOS TOU QEOU

Some phrases here translate better than others, such as "and the truth is not in him" (mismatched preposition/object though), "that Jesus is not the Christ," and "that Jesus is the Son of God".

What do you think? Do you think that this could be useful, at least with a better or more extensive training corpus, or for some applications?

For comparison, here are some of the sentences genereated from their own training corpus of English:

* Cindy thinks that George believes that to read is tough #
* that the bird is tough to read bothers the horse #
* Pam thinks that Jim thinks that to read is tough #
* Cindy believes that George thinks that to please is tough #
* that the horse is easy to please disturbs George #
* Cindy thinks that Jim thinks that to read is easy #
* Beth thinks that Pam believes that to read is tough #
* Cindy thinks that Cindy believes that to read is easy #
* Cindy believes that Joe believes that to please is easy #
* that the cow is eager to read worries the bird #
* Cindy thinks that Joe believes that to read is easy #
* that Cindy is easy to please bothers the dog #
* Cindy thinks that Beth believes that to read is easy #
* that Joe is tough to read annoys the horse #
* that the rabbit is eager to please bothers the dog #
* Cindy thinks that Cindy thinks that to read is easy #
* Cindy thinks that Pam thinks that to please is tough #
* Cindy thinks that George thinks that to read is easy #

The vocabulary here seems to have been deliberately limited so that the algorithm could pick up on the syntactical structures. No attempt of any semantic processing is undertaken.

I suggest that its use, if any, is in helping linguists in their work of finding regularities of language, and in helping those who process texts by giving them another tool for breaking down syntax.

kind thoughts,
Peter Kirby

jonesg

Have you ever considered Godel's theorum,
it would indicate a computer cannot have an idea.

Peter Kirby

Gödel's Incompleteness Theorem is the most widely misunderstood and misused truth in mathematics.

kind thoughts,
Peter Kirby

copernicus

I think that the effort is commendable, but not of much practical value. Being able to recognize and generate syntactic configurations of words is not the same as being able to use them to communicate. The ultimate question is not what phrase structure to assign to a given string of words but WHY you assign that structure. How does that syntactic configuration contribute to the representation of a running discourse? In fact, what is lacking from virtually all linguistic processing models is some coherent representation of a running discourse. Until we begin to tackle that problem, language processing programs are going to be useful only in very limited applications--where the nature of the discourse is largely predictable. Natural language processing can do a good job of extracting information from text, but the techniques are not really scalable to full natural language understanding systems.

I don't think that such a program can be of much use to lingusits, who can produce much more sophisticated syntactic analyses of unknown languages far more quickly. After all, the program in question knows nothing about language typologies, for example. It just uses some simple context free analytical techniques to recognize and generate patterns. On the other hand, it might help computational linguists scale up grammars that are used in so-called deep linguistic processing programs. There seems to be some renewed interest in syntactic parsers, as the statistical methods that have dominated the field in the last decade seem to have plateaued. Right now, efforts seem directed at hybrid systems that can use statistical analyses to enhance syntactic scale-up. I think that the ADIOS program is a decent contribution to that effort.