Human natural language metalanguage

Question

I was thinking about how a controlled grammar of English can be used as a programming language because it’s fully parsible.

The idea of doing this for other languages, such as Sanskrit, brought me to that fundamental questions of linguistics, which is that it is interesting how different the grammars of English and Sanskrit probably are, and with a little streamlining, presumably any human language can be used as a programming language. But we don’t know the extent of how many different languages and grammatical features of languages there are in the world. It opens the weird but intriguing question that maybe by discovering a new language via field work, we would “discover” a new syntactic feature of language that had never occurred to us before. Sort of related is when you learn a new language and are really surprised by a grammatical feature, for example, in Turkish, when you say “A means B”, you have two corresponding verbs to sort of tag the relationship: “A demek B demektir” (I think).

So it leads to the obvious question of if the grammar of human languages themselves seem to have infinite possible variation (not the number of sentences in a language, but the number of unique grammatical structures across all language), surely there is some set of axioms for how these grammar structure evolve and develop? Or if that’s assuming too much, at the very least, the more reasonable question of what are the limits of human language?

I find the second question way smarter than the first because of how empirical it is. It does not require chasing red herring theories into dead ends. It doesn’t require anything theorizing at all. It’s just plug-and-chug “observation - confirmation”. Sort of like using numerical approximations to get a closer idea of information about the distribution of prime numbers without perfectly proving anything yet, maybe a profoundly simple yet effective way to understand language is to extremely intensively keep trying to empirically generate simple hypotheses about what human language surely can not be like, and just keep doing that till we actually find a pretty stable boundary that holds up against any empirical evidence - found in the wild, or even a constructed language taught to children to see how or if they can assimilate it.

I was wondering if pursuing linguistics as a subtractive process is something any researchers have already attempted?

Answer 1

In 1984, I created an Expert System, named XTRAN (TM), whose domain of expertise is computer languages, data, and text. XTRAN parses language content to XTRAN Internal Representation, known as XIR (TM), which is a language meta-model I have created to represent language content across languages with a single representation (i.e., it is language-independent).

XTRAN's rules language can then be used to automate manipulation of the XIR resulting from parsing language content -- including analyzing it, transforming it, and translating it to a different language with the same functionality. XTRAN can also automate manipulation of data and unstructured text.

After creating a few computer language parsers for XTRAN, I automated the process by creating a parsing engine. It takes as its input Backus-Nauer Form (BNF), a meta-language for describing unambiguous grammars. Since the traditional syntax of BNF itself is pretty user-hostile, I use a functionally equivalent but friendlier syntax I call XTRAN BNF, or XBNF (TM). XTRAN parses the XBNF it is given via its rules language, stores it in an internal representation specific to it, and then executes it in order to parse the described language. That has allowed me to add computer languages quickly to XTRAN's repertoire. The parsing engine features "fastback" -- the ability to snap back to the last appropriate node in the syntax tree when it hits a dead end. That makes its parsing process very fast, even for complex grammars.

In the case of language content transformation and translation, the XIR representing language content must be externalized as text source code for human use. I call this process "rendering" of the XIR. I have also created a rendering engine, which provides many styling controls in the rendering process. The rendering engine is also fed XBNF; since all language ambiguity was resolved at parse time, rendering XBNF is simpler than parsing XBNF.

Recently, I wondered if there exists a human language that is sufficiently contstrained in terms of ambiguity to lend itself to XTRAN's method of parsing, manipulation, and rendering. I am exploring Simplified Technical English (STE), a constrained forn of English that has been adopted by a variety of European organizations for use in creating requirements, specifications, and documentation. I am exploring whether STE is amenable to my parsing and rendering engines' way of dealing with language content. If so, XTRAN's rules language can automate manipulation of that language.

Another hat I wear is that of Expert Witness, and I have observed that legal documents are similarly constrained, in order to avoid ambiguity. So I am also exploring the possibility of applying XTRAN's parsing and rendering engines to them.

Infornation about all of this is at WWW.XTRAN-LLC.com. The parsing and rendering engines are discussed at WWW.XTRAN-LLC.com/xtran.html#parse-gen. I'll be happy to answer any questions, and I welcome comments. I'm at [email protected].

Answer 2

we would “discover” a new syntactic feature of language that had never occurred to us before.

I think this is the core interesting question, which should be approached from a metatheoretical perspective, which contradicts the desideratum of plug-and-chug observation - confirmation. In a sense, you have discovered the main current trend in investigating language, which is to assume the fewest a priori theoretical devices and try to devise a way to talk about language structures without any theoretical preconceptions. In practice, though, linguists always have to assume certain principles. The very idea of a "structure" is a theoretical device that has to be presupposed and cannot be discovered from the raw data.

A "word" is a kind of linguistic structure. In English, "snake" is a word, and "snack" is a different word. You cannot set a machine to blindly computing over a corpus of raw data and have it discover that English has a word "snake" and a distinct word "snack". Now let's look at a different kind of structure – a multi-word utterance. That would include "snake", "a snake", "the snake", "snakes", "big snakes"... "I saw the big snakes"... "I saw the man carrying the big snakes"... The problem here is to define "a structure", and to say what constitutes an example of "Structure 12,975" versus "Structure 6,568,547" (we'll obviously run into a problem if we try to enumerate all of the utterances of a language and call each of them a "Structure". The moral of the story here is that in order to sort data into types, you have to have a theory of possible types – you need a theory. Data sorting always requires theorizing.

I'd say that you are correct in your interest in simple hypotheses, which brings us to the question of the logic of language acquisition. A child will not devise a crazy complicated hypothesis and then later reject it because it doesn't work empirically. The long-standing (inadequate) theory of acquisition that linguists usually adhere to is based on a sort of nutty idea that a child is a massive hypothesis-generator who gathers up all of the possible hypotheses and then picks the "best" one (let's just say "simplest", except that there is usually no unique simplest hypothesis). A better approach is the more inductive view that a child entertains hypotheses for which there is sufficient prima facie evidence, and does not ever conjecture that the language "could be" more complicated than there is evidence for. This is an "incremental" quasi-Bayesian view that there are initial hypotheses, then revisions of hypotheses. But still, the enterprise of looking at "possible" analyses requires some theory of what is "possible", as opposed to simply "imaginable, if you assume no knowledge whatsoever.

Answer 3

There are constructed languages that are designed to be parsable, the original was Loglan and it spawned some forks, e.g., Lojban and gua\spi (shortly mentioned in the Wikipedia article on Lojban).

I am not aware of a controlled language based on English that is parsable.

Answer 4

This is an interesting discussion of this question, from Pullum:

How Many Possible Human Languages Are There? Geoffrey K. Pullum Linguistic Inquiry Vol. 14, No. 3 (Summer, 1983), pp. 447-467 (21 pages) Published By: The MIT Press https://www.jstor.org/stable/4178339