[Lingtyp] Expletive derivational negation (revisited)

Wed Aug 28 19:05:23 UTC 2024

Hi Rainer,

To have semantics without reference to meaning, i.e. based only on form, we need to establish the stage at which form and meaning are in a one-to-one relationship because at this stage it does not matter whether we work with meaning or with form. Since working with meaning is more difficult to control than working with form, ChatGPT works with form. As we know from linguistics, we disambiguate through addition of form, e.g. cut is ambiguous between a noun and a verb, but by saying a cut and to cut, we immediately solve the problem. The same strategy is used in ChatGPT, only ChatGPT works with very long sequences of forms (tokens or subword units; ChatGPT uses subword tokenization). Since a machine does not literally need to understand language (cf. image recognition: a computer does not really see), for language processing it is enough if the computer "understands" and produces existing long sequences of forms, which is easy to achieve by observing co-occurrences of forms statistically. This is how ChatGPT "understands" and generates long sequences of forms which, though produced without reference to meaning, appear meaningful text for us.

Is this how a human being processes language? I do not know – there is no research on this issue, as far as I know. But I could imagine that a human being memorizes long sequences of forms, which is the parallel to what is going on in ChatGPT.

Two of the computer scientists behind ChatGPT (Ilya Sutskever and Dario Amodei) have confessed many times that they do not know why their algorithms work for language processing. They seem to believe that this is because of the large amount of data used and therefore they regularly enlarge the amount of the data.  In other words,  the above is my personal view on why ChatGPT can generate language. And I think that linguists can contribute to LLMs optimization by helping computer scientists establish the optimal length of the sequence necessary for language generation that is exclusively form-based.

Please do not hesitate to respond to this message if any questions arise.

Greetings from Vienna,
Gauss:AI
https://sites.google.com/view/stelamanova/gaussai<https://sites.google.com/view/stelamanova/gaussai?authuser=0>

________________________________
From: Rainer Feer <rainer_feer at sil.org>
Sent: Wednesday, August 28, 2024 11:59 AM
To: manova.stela at gmail.com <manova.stela at gmail.com>
Subject: WG: [Lingtyp] Expletive derivational negation (revisited)

Dear Stela,

I would be thrilled hearing your explanation how one can generate language without considering semantics. That didn’t even occur to me. Hope others request that as well :)

Best wishes

Rainer

Von: Lingtyp <lingtyp-bounces at listserv.linguistlist.org> Im Auftrag von Stela Manova via Lingtyp
Gesendet: Dienstag, 27. August 2024 16:02
An: lingtyp at listserv.linguistlist.org; Salminen, Jutta <jutta.salminen at uni-greifswald.de>
Betreff: Re: [Lingtyp] Expletive derivational negation (revisited)

Dear lingtyp readers,

Now that the thread “Expletive derivational negation” has been continued, I am adding a few sentences from the perspective of a Large Language Model such as ChatGPT.

ChatGPT operates with subword units (NOT with WORDS as many linguists seem to believe). Expletive morphemes such as those discussed in this thread are not an issue because ChatGPT does not associate morphemes (subword units) with meaning (which is similar to how a-morphous, word-based and paradigm-based morphology treat morphemes in linguistic theory: semantics is assigned at the level of the word, morphemes are just markings without meaning). If lingtyp readers are interested in how ChatGPT, which does not consider semantics at all, produces meaningful language, I can post an explanation.

Now, since ChatGPT can understand and generate a significant number of languages, it seems that the only way to achieve a universal linguistic analysis (= understanding and generation of all languages) is by treating morphemes as phonemes, i.e. as building blocks that create meaning but do not have semantics of their own.

I also think that we need the term ‘subword unit’ in linguistics, to refer to units without meaning as well as to units which, for some reason, cannot be clearly related to (a specific) meaning, see about -er below. I am against the use of 'morphome' for that purpose because 'morphome' has been linked to inflectional morphology and seems to require a root or a stem to be involved. ‘Morpheme’ will remain for the designation of the morphological unit that relates meaning and form. The adoption of 'subword unit' will help us avoid misleading definitions such as the following ones from Haspelmath & Sims (2010), Understanding Morphology:

  *   morpheme: the smallest meaningful part of a linguistic expression that can be identified by segmentation; a frequently occurring subtype of morphological pattern. (p. 335)
  *   exponent: when a morphological pattern (e.g. -ed) expresses an inflectional feature value (e.g. past tense), it is the exponent of that feature value. (p. 328)

In relation to these definitions, the following questions arise: What is a "subtype of morphological pattern" + "frequently occurring"? Does a morpheme have a specific meaning, or can it be just a pattern, see about -er below? What is the difference between 'morpheme' and 'exponent'? Can derivational morphemes have exponents?

>From psycholinguistics, we know that for the human brain work -- work-er and corn -- corn-er are morphologically related, though corner is not derived from corn. Then, if morphology is pattern-based, as Haspelmath and Sims seem to claim and psycholinguistics to evidence, we should add to this pattern clean -- clean-er, too. I therefore suggest pieces of structure such as -er to be called ‘subword units’.

Greetings from Vienna,

Gauss:AI

https://sites.google.com/view/stelamanova/gaussai<https://sites.google.com/view/stelamanova/gaussai?authuser=0>

________________________________

From: Lingtyp <lingtyp-bounces at listserv.linguistlist.org<mailto:lingtyp-bounces at listserv.linguistlist.org>> on behalf of Salminen, Jutta via Lingtyp <lingtyp at listserv.linguistlist.org<mailto:lingtyp at listserv.linguistlist.org>>
Sent: Monday, August 26, 2024 2:50 PM
To: lingtyp at listserv.linguistlist.org<mailto:lingtyp at listserv.linguistlist.org> <lingtyp at listserv.linguistlist.org<mailto:lingtyp at listserv.linguistlist.org>>
Subject: [Lingtyp] Expletive derivational negation (revisited)

Dear all (who recently discussed the phenomenon of derivational expletive negation),

I was informed about this conversation last week and will now add my contribution to it with a little delay. I couldn’t respond directly to the thread, since I just joined the list, but I shortly cite here the relevant questions and conclusions that I will comment from my perspective:

Bastian Persohn (Fri Aug 16 11:55) “– – Their closest relatives are probably found in instances like Un-fall ‚accident‘ < Fall ‚case‘, i.e. ‚the undesirable case‘ or Un-tier ‚monster‘, lit ‚un-animal‘. What all these have in common is a negative element, albeit in the subjective rather than the material domain.”

Tim Zingler (Fri Aug 16 12:15): „I like the idea that the function has shifted as part of a subjectification (?) process. Does that happen with negators cross-linguistically?”

My case is not literally parallel to the lexical negators and lexemes discussed in the thread so far, but there are similarities on the general level.

In my dissertation (Salminen 2020<https://helda.helsinki.fi/items/013d254d-e401-4ffe-ae09-372e60d6ffb6>), I studied the Finnish verb epäillä ‘doubt; suspect, suppose,’ which – as the translations reveal – can have basically two opposite meanings ‘to think that something is, or is not the case.’ But besides the material negation or the lack of it (to apply the term from B. Persohn), undesirability plays a key role in motivating the connection and the diachronic shift from the negation-inclining to the affirmation-inclining meaning of this verb: very often epäillä is used to either doubt something desirable or suspect something undesirable, which both share the tone of undesirability. (I keep the description of the variation here very short and simplified; it can be read in English in a nutshell in my 2018 paper<https://www.jbe-platform.com/content/journals/10.1075/fol.15030.sal>, and in more detail in Finnish in other parts of my dissertation.)
What is perhaps most interesting for the current discussion, is the fact that the verb epä-illä is derived from a negator: epä is a present participial form of the Finnish negative auxiliary, historically also used as a 3SG form of it (cf. ei). (The same morpheme also functions as a prefix, e. g. epämukava ‘unpleasant,’ but in the verb epäillä functions as a derivational base. The possible chain of derivation is: epä > evätä ‘refuse, decline’ > epäillä (frequentative).)

Since, a notable part of this verb’s use in present day Finnish lacks the interpretation of material negation and denotes “only” undesirability or some other aspect of evaluative negativity, it can be seen as one (complex) example of a diachronic change, where a negator developed more subjective meaning(s) of negativity (instead or besides the negation proper).

Lastly, all the mentioned meanings are still available for epäillä, and they are dependent on the clausal, sentential and wider context of the verb.

Best regards,

Jutta

_________________
Dr. Jutta Salminen
Postdoktorand / postdoc-tutkija, FT
Lektorin für Finnisch / suomen lehtori

Institut für Fennistik und Skandinavistik
Ernst-Lohmeyer-Platz 3
17489 Greifswald
jutta.salminen at uni-greifswald.de<mailto:jutta.salminen at uni-greifswald.de>
Tel.: +49 (0)3834 420 3601

www.uni-greifswald.de/fennistik<http://www.uni-greifswald.de/fennistik>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listserv.linguistlist.org/pipermail/lingtyp/attachments/20240828/b2efeee6/attachment.htm>