where are there good analyses of the text-to-meaning problem? i.e. given a body of text in an unknown language (and pretty much nothing else), figure out what it means?
a version of the problem comes up in machine learning: some people think LLMs in effect solve this problem, others (e.g. bender and koller) think it can't be solved. (plausibly you'd need some constraints to solve it; one question is how weak those constraints can be.)
the text-to-meaning problem is related to the problem of "radical interpretation" in philosophy (roughly, interpreting speech of an unknown speaker from scratch) but the former is even more radical, because it doesn't allow the interpreter to use knowledge of the environment.
i'm interested in analyses from any perspective -- including linguistics, machine learning, and philosophy, but not limited to these.
@jplennon630 a classic! but a slightly different problem. https://x.com/davidchalmers42/...
@davidchalmers42 Here ya go. First, on symbol emergence and meaning (this was preliminary; a more developed take is under review): https://link.springer.com/chap... ...and on empathy: https://ieeexplore.ieee.org/do... ...and finally cognition, with associated proofs / experimental evidence: https://www.techrxiv.org/artic...
@davidchalmers42 In my previous life, I wrote about the emergence of meaning in language models in the context of Hutto & Myin (hard problem of content) Skyrms (signalling games) http://avant.edu.pl/wp-content...
@davidchalmers42 Meaning cannot be represented independent of the agent doing the understanding. Meaning is a set of pre-existing and newly formed models that fit the input. Models that, when applied make predictions that are both (mostly) self-consistent and (mostly) track future inputs.
@davidchalmers42 You can’t. What does this mean? ξξ→ζ ξζ→ξ ζξ→ξ ζζ→ξ There are, at least, three possible answers, of which only one is correct. Meaning exists in the behavior of things (turn the volume on your TV off) and text is static.
@davidchalmers42 I don't know anything about this but I'm thinking about what I'd do and it's fun.
@davidchalmers42 “what it means” seems ill defined if given “pretty much nothing else”.
@davidchalmers42 If we assume sensory input or other modalities then the notion of "grounding" (i.e. associating text with a sensory domain) is relevant here. See @ChrisGPotts https://youtu.be/7b2_3dDTKMc?t... take on this topic:
@davidchalmers42 You may find this paper about understanding scientific explanations in an unknown language interesting. Here "understanding" means becoming able to make predictions and discover new explanations for previously unseen phenomena. Paper: https://arxiv.org/abs/2201.102... 1/2 continues...
@davidchalmers42 Is there any reason to think the text-to-meaning problem is different from the more general sensory input-to-meaning problem?
@davidchalmers42 The answer depends on the form(s) of life whose language games were being "played". Absent that information, how is a solution even conceivable? Language isn't a picture of / isomorphic to reality - the same text can have utterly different meanings in contexts that aren't given
@davidchalmers42 From a physics perspective, "Meaning" stands for averaging of EM waves. Text has an average of Charge density. The two are related and can be calculated using Jafomenko's equations: https://en.m.wikipedia.org/wik... (the text is orthogonal to its meaning)
@davidchalmers42 Here is an attempt: we always observe the world through filters. For example we see the reflection of light on surfaces and then build models of the world from these observations. The entity emitting the text is just a complex reflection device the world is reflected through.
@davidchalmers42 What are the restrictions we have? How we should understand «context»? I got you: the problem is, in fact, similar to the radical interpretation issue.
@davidchalmers42 If the language is human then it's solvable I guess cause we have the same abstractions in our heads and we could exhaustively search all the combinations , but if it's alien even then probably possible unless they r 4d beings or something
@davidchalmers42 Possibly relevant: https://arxiv.org/abs/1711.000...
@davidchalmers42 This is relevant: https://arxiv.org/abs/2208.029...
@davidchalmers42 En tiedä sellaisista
@davidchalmers42 There are tools that create summaries and tools that disguise meaning with obtuse terminology - jargonification. In terms of physics, they take a Fourier limited text and degrade its https://www.rp-photonics.com/m... I have a quality data set that I need help interpreting -- hook me up.
@davidchalmers42 It's possible to build a translation machine that vectorizes word co-occurences in one language and compares this to a similarly created vector from another language. It presupposes a somewhat similar underlying distribution however, so you couldn't translate alien languages.
@davidchalmers42 @nescio13 It seems to me that within a decade or so, Chat GPT will be able to use the available variables, the input, to produce several alternatives for us to choose from, but getting to a more-or-less definitive version will inevitably require a Rosetta Stone.
@davidchalmers42 LLMs need a situated body for that? Lots of sensors to learn the difference between body and world.
@davidchalmers42 I know it's a popular team stage problem at International Olympiad in Linguistics
@davidchalmers42 There are none as this is a Nearly Decidable problem. It's Undecidable—tho decidable in special cases for *reasons* bound up in, eg, Pike's -emic and -etic distinction or the four-letter PCP alignment case: morphe has adjoints it can tame (suppressed *for* a Decision)— Or not!
@davidchalmers42 Interesting related paper: "Unsupervised Machine Translation Using Monolingual Corpora Only" https://openreview.net/forum?i...
@davidchalmers42 @nescio13 NOTHING else? you mean you don't even know what species prduced it? on what planet? whether it was language at all? surely u need SOME information. do ALL human languages have common features that can be used to crack them?
@davidchalmers42 Presumably one can design a text to be interpretable, as has been proposed in detail for transmission to extraterrestrials. Start with counting, then maths, then identifiable physical constants & laws, then you’re starting to be able to talk about the environment