Just published: "Human #languages trade off #complexity against #efficiency" in PLOS Complex Systems doi.org/10.1371/jour... Main findings: Langs that are (information-theoretically) more complex are more efficient. Larger speaker communities tend to use more efficient langs. 1/ - ThreadSky

sascha-wolfer.bsky.social • 27 days ago

Just published: "Human #languages trade off #complexity against #efficiency" in PLOS Complex Systems
https://doi.org/10.1371/journal.pcsy.0000032

Main findings: Langs that are (information-theoretically) more complex are more efficient. Larger speaker communities tend to use more efficient langs. 1/

Comments

datenwolf.bsky.social•27 days ago

Very interesting work.

Here's a hypothesis I've held for some time: In languages that employ whole sets of articles occupying the same grammar function, where the particular choice of specific article used depending on its associated noun actually acts as a form of forward error correction.

sascha-wolfer.bsky.social•27 days ago

Interesting! I don't think it relates directly to the article (unless there are specific consequences for our complexity metric) but anyhow: how you would test for such a forward error correction hypothesis. Also: Which characteristics of the noun determining article usage do you have in mind?

datenwolf.bsky.social•26 days ago

One thing that could server as forward error correction would be things like rhyming schemes, or certain "natural feeling" sentence rhythms, that avoid short term repetition of phrases but frame text into a predictable pattern.

datenwolf.bsky.social•26 days ago

Another forward error correction could be the grammatical gender and declination of articles in language such as German, Italian or French (der Löffel, die Gabel, das Messer) and how they change depending on the grammatical case of the associated noun.

datenwolf.bsky.social•26 days ago

As for how to test: By presenting a cohort of subjects with texts crafted from "bland" to "complex" in their breadth of vocabulary, varied choice of equivalent grammar (e.g. possessive declaration using "of" vs. "…'s") and various degrees of erasure from those texts.

datenwolf.bsky.social•26 days ago

For written text delete and/or alter increasing amounts of characters. For oral communication insert audible dropouts or overlay with noise. Measure what's the erasure threshold is, at which the semantics are no longer conveyed consistently.

sascha-wolfer.bsky.social•27 days ago

We trained 4 types of lang models (LMs: PPM, PAQ, LSTM, Transformers) on a corpus of 3 bn words across >6500 docs in >2k langs. Entropy rate distributions over LMs were surprisingly consistent. So, in this context, the choice of LM does not have a big impact on cross-linguistic studies. 2/

sascha-wolfer.bsky.social•27 days ago

What DOES make a different is language: if one lang is harder to process for an LM than another, this relationship holds across other LMs, text types, and even across symbolic levels (chars, words, BPE). But why would some langs evolve to be more complex, given the increased processing effort? 3/

sascha-wolfer.bsky.social•27 days ago

Our answer: it's a trade-off. Langs with higher complexity tend to produce shorter texts, i.e. are more efficient in communication. That means that a complex language might offer more options to convey the same idea using fewer symbols. We also show that this is not a trivial relationship. 4/

sascha-wolfer.bsky.social•27 days ago

So, is there any relationship to language-external variables? We found that larger communities tend to use more complex, more efficient langs. We speculate that this might have to do with the importance of written communication in larger societies which creates pressure favoring shorter messages. 5/

sascha-wolfer.bsky.social•27 days ago

This pressure might arise from resource costs for production and transmission – paper, storage, bandwidth... Keep in mind that we're studying complexity from the perspective of information theory. It would be interesting to see how these results align with other ways of measuring lang complexity. /6

Comments

Posting Rules

Reply