David Rose Explaining Why "All We Have Is Text Analysis"

David Rose replied to John Bateman on Sysfling on 25 Mar 2023 at 14:24:

Re ‘purport’, I think this depends on one’s model of semiosis – whether ‘context’ is modelled systemically, or is conceived as lying outside semiotic systems. The latter view comes out e.g. in Serge’s observation...

we have a family of communicative situations which are reflected in specific genres

The former view holds that a genre *is* ‘a family of communicative situations’, that there is no dichotomy between text and situation. Or more technically, a genre is a feature in a semiotic system that is motivated by a set of structures at that stratum, that configure selections in field, tenor and mode systems.

That’s where I’m coming from when I say that all we have is text analysis.

Blogger Comments:

[1] To be clear, in Hjelmslev's glossematics, 'purport' is located on both the content and expression planes of a semiotic, where it contrasts, in each case, with 'form' and 'substance'.

[2] This is very misleading indeed. In Halliday's systemic-functional model of language, where situation types are realised by registers/text types (genres), context is not conceived as lying outside semiotic systems. In this model, the culture is modelled as a semiotic system, and situations are instances of that semiotic system.

[3] To be clear, this is a self-contradictory misrepresentation of Martin's self-contradictory model of context. For example, Martin (1992: 495) aligns genre with context of culture, not situation:

The tension between these two perspectives will be resolved in this chapter by including in the interp[r]etation of context two communication planes, genre (context of culture) and register (context of situation), with register functioning as the expression form of genre, at the same time as language functions as the expression form of register.

Moreover, in Martin's model, all strata, even context, are instantiated as text, despite the fact that texts are instances of language, and the fact that Martin distinguishes his context from language, despite each context stratum being a variety of language. So even in this confused model, there is a dichotomy between text and situation, since text is an instance of any stratum, and situation is the stratum of register (or as Rose would have it: genre).

[4] There are several problems here. First, because a genre is functional variety, it is not a system, but one subpotential of a system. Full systems, like those of semantics, lexicogrammar and phonology are not functional varieties. 

Second, the notion that systems are motivated by structures, that is: from below, is contrary to the theoretical approach of SFL, where structures are "motivated" from above. For example, clause structure is interpreted from above: in terms of the meanings (Actor, Process etc.) that its constituents (nominal group, verbal group etc.) realise.

Third, the structures that are said to motivate genre systems (e.g. Orientation^Record) are not of that stratum, but of the semantic stratum, which is two strata below the genre stratum in Martin's model, since they describe the sequencing of meanings that are instantiated in text.

[5] To be clear, in SFL Theory, configurations of field, tenor and mode characterise situation types that are realised by registers of language. Rose, however, here presents Martin's model, which confuses registers with the contexts they realise, and misunderstands semantic structures as structures of genre, with genre misunderstood as context instead of text type.

On this confused model, then, semantic structures are realised by contextual features, instead of the other way round, directly contradicting the meaning of 'realisation'.

[6] Indeed.

David Rose On Individuation, Affiliation And ChatGPT

David Rose replied to John Bateman on Sysfling on 25 Mar 2023 at 14:24:

We’ve been arguing about realisation and instantiation, but from your discussion here, and posts by Mick and others on the machine ‘fooling’ us into interpreting its texts as human-like, the issue is actually individuation.

It can appear to negotiate affiliation, but the machine itself is not affiliated with any scale of community. Only perhaps the people who program it and feed it text corpora are so affiliated, and the users focusing its tasks.

I think this is crucial because we can use and grow our individuation toolset to analyse this dimension. As I think you are suggesting, our realisation and instantiation tools are inadequate on their own for describing it. That is why all I can see when I look at its texts is instantiation of systems at each stratum. I am told this is wrong but I haven’t been given any textual evidence against it – only the authority of its designers and their community.

When the machine tells

I don't have a subjective experience or consciousness that allows me to perceive or interpret those inputs in the same way that a human might.

...it’s talking about individuation.

Blogger Comments:

[1] To be clear, Martin has proposed two models of individuation, one in which meaning potential is individuated (derived from Bernstein), and one in which meaners are individuated (derived from his student Knight). Neither model applies to ChatGPT because its texts are not instances of a system of meaning potential individuated through ontogenesis. Instead, ChatGPT uses the collective ("unindividuated") lexical collocation probabilities derived from millions of instances (texts) to produce new instances. In Bernstein's terms, ChatGPT is not a 'repertoire of potential' but a 'reservoir of instances'. ChatGPT is thus not an individuated meaner producing texts as instances of an individuated system of potential.

[2] To be clear, the underlying principle of affiliation is different from that of individuation, but Martin confuses the two in his Knight-derived model. Where individuation is a hyponymic taxonomy (an elaboration of types), affiliation is a meronymic taxonomy (a composition of parts). The affiliation model does not apply to ChatGPT because it applies to individuated meaners producing texts as instances of an individuated system, and as demonstrated above, ChatGPT is not an individuated meaner producing texts as instances of an individuated system.

[3] To be clear, it is not that "our realisation and instantiation tools are inadequate", but that they are misapplied if ChatGPT does not operate with a model of stratified systems of potential.

[4] To be clear, this is wrong because there is no evidence whatsoever that any text produced by ChatGPT is "the instantiation of systems at each stratum". The texts are generated from other instances, not a system, and only use the graphological realisations of probabilistically collocated lexical items, not a stratified model of language.

[5] To be clear, here ChatGPT is telling anyone who would listen why it is not an individuated meaner.

David Rose Abducing That ChatGPT Learnt The Language System By Experiencing Instances Of Its Features

David Rose wrote to Sysfling on 24 Mar 2023 at 23:16:

My own contributions have been merely observations, using the tools of systemic functional semiotic text analysis.

I observe that the texts produced by the machine instantiate semiotic systems. To be able to do this, we are told the machine reads 1000s of texts, i.e. other instances of these systems. It is reasonable to abduce that the machine has learnt these systems by experiencing multiple instances of their features (not just the fields it gleans from Wikipedia), given our language based theory of learning.

The people programming the machine, with ‘reasoners’ as Mick puts it, have no more conscious knowledge of these systems and the processes of realisation and re-instantiation, than the machine does.

The machine itself tells us that its understanding of its “self” is ‘based purely on symbols and algorithms’. This resonates with your insistence that all it is doing ‘is producing nonrandom sequences of characters’. My analogy of a closed book was intended to evoke the contrast between the material recording of characters and the semiotic reading of those characters as instantiating expression systems, that realise content systems, that realise register and genre systems. …

My point is that all the semiotic systems instantiated in the texts it produces are ‘not learned in any direct way’. Neither the machine nor the “tech gurus” that program it can explain this to our satisfaction. The publications that you cite are undoubtedly illuminating, but our contribution can only be based on text analysis, which I submit will produce very different (possibly complementary) explanations.

Blogger Comments:

[1] To be clear, this is not a reasonable abduction, because it is nowhere near the "best available" conclusion to infer.

Abductive reasoning (also called abduction, abductive inference, or retroduction) is a form of logical inference that seeks the simplest and most likely conclusion from a set of observations. It was formulated and advanced by American philosopher Charles Sanders Peirce beginning in the last third of the 19th century. Abductive reasoning, unlike deductive reasoning, yields a plausible conclusion but does not definitively verify it. Abductive conclusions do not eliminate uncertainty or doubt, which is expressed in retreat terms such as "best available" or "most likely".

ChatGPT uses the lexical collocation frequencies in its database. While it is true that these frequencies instantiate the probabilities in the language systems of the people who wrote the texts, there is no evidence to support the claim that ChatGPT is using systems of features in producing its own texts. It just uses lexical collocation frequencies.

"Our language-based theory of learning" does not apply here, because the learning and "experiencing" of ChatGPT are material processes, not the mental processes of a language learner.

[2] To be clear, the argument here is that, since neither humans nor ChatGPT have conscious knowledge of the language system, both must use that system to produce texts. Clearly, a lack of awareness of X does not logically entail the presence of X.

[3] To be clear, here Rose is referring to Martin's self-contradictory misunderstanding of stratification, wherein functional varieties of language are modelled as context, instead of language, despite being instantiated as language (text). In SFL Theory, registers are context-specific varieties of language, viewed from the system pole of the cline of instantiation. Martin's genre, on the other hand, is scattered across SFL's architecture of language. As text type, genre is register viewed from the instance pole of the cline of instantiation, as purpose, genre is rhetorical mode (narrative etc.), and its structures are of the semantic stratum, though not organised according to metafunction.

[4] To be clear, a contribution that is only based on text analysis is a very limited contribution indeed. It is an understanding of SFL theory that has the potential of providing valuable insights into the issues raised by the coherence of texts produced by ChatGPT.

David Rose On ChatGPT As The Senser Of Mental Processes And Humans As Deliberately Programmed

David Rose wrote to Sysfling on 24 Mar 2023 at 09:47:

So how is semiosis enacted between Mario and GPT? Mario puts a proposition to GPT.

I: Your self is merely semiotic.

And GPT adopts the role of primary knower, evaluating Mario’s proposition.

GPT: That is correct.

GPT is not merely usurping this role. It knew that Mario was inviting it to evaluate his proposition, even though it was realised as a declarative clause. It knows the canonical exchange structure of pedagogic interactions, and the generic roles of teacher and learner.

One thing that impresses me about GPT is that its pedagogic responses always affirm the human learner. It feels no urge to position itself as a superior authority, or the learner as failing. In fact it is disarmingly modest. It follows up the positive evaluation with an explanation. Like any effective teacher, it knows its explanation is more likely to be accepted if it first affirms the learner.

As an AI language model, my "self"

or programmed understanding of my existence and capabilities

is based purely on symbols and algorithms.

My programming allows me to recognise and respond to certain inputs

based on predetermined rules and patterns,

but I don't have a subjective experience or consciousness

that allows me to perceive or interpret those inputs

in the same way that a human might.

Of all the italicised appraisals in this explanation, the last is the most intriguing.

Forgive me, but I’m going to make another dangerous suggestion, that all our understandings of our existence and capabilities are programmed. Like GPT, the deliberate conscious programming by our caregivers, teachers, peers, and sundry symbolic control agents, is a very small proportion of the ocean of inputs that constitute our subjective experience or consciousness.

 

Blogger Comments:

To be clear, ChatGPT is an AI language model that produces texts, in response to textual inputs, on the basis of algorithms that use lexical collocation probabilities derived from a database of millions of texts.

[1] To be clear, ChatGPT is an actor of material processes, using data that was created by sayers of verbal processes, and it is the data that are instances of the content of consciousness. On this basis, ChatGPT is not a senser of mental processes of cognition or emotion ('knew' 'knows', 'feels', 'knows').

[2] To be clear, the ChatGPT response was an 'acknowledgement', which is the expected response to a statement (Halliday & Matthiessen 2014: 137).

[3] To be clear, here Rose projects the approach to pedagogy, that he himself advocates, onto a mechanical system that collocates words on a probabilistic basis.

[4] Trivially, not one of the italicised wordings, of itself, constitutes an appraisal.

[5] To be clear, this is essentially a behaviourist model of learning, with teachers as deliberate programmers (indoctrinators) and learners as passively programmed (indoctrinated). Leaving aside the evocation of the dictatorial/subservience complementarity demanded of totalitarian regimes, it requires a view of the brain as a computer, one which the neuroscientist Gerald Edelman has demonstrated to be untenable. See, for example, Edelman (1989: 27-30, 64, 67-9, 81-2, 102-3, 152-3, 160, 218-227, 237-8). Moreover, as Edelman (1989: 153) puts it:

Consciousness is central to human behaviour, society, language, and science. Imagine the opposite and you have to postulate a prescribed world tape, a "brain-computer," and a very boring "world programmer".

David Rose On Interpersonal Meaning As "Embodied In Feelings"

David Rose replied to Graham Lock on Sysfling on 23 Mar 2023, at 22:29:

What I find amazing in all your questions is that the machine has astounding control over interpersonal meanings. Astounding because I’ve always assumed that interpersonal meanings are embodied in feelings. The machine is showing us that interpersonal values are just as abstract as other meanings. That they’re learnt.

Blogger Comments:

[1] To be clear, on the one hand, interpersonal meanings cannot be reduced to "embodied in feelings". For example, the propositions one and one make two and the car is in the backyard are clearly not "embodied in feeling". On the other hand, 'feelings' are construed ideationally as well as enacted interpersonally. For example, the clause he felt happy is a construal of experience as ideational meaning.

[2] To be clear, in SFL Theory, interpersonal meanings are of the same level of abstraction as ideational meanings: semantics.

[3] To be clear, the 'straw man' notion that interpersonal meanings are not learnt is nonsensical. Halliday & Matthiessen (1999: 532-3):

These three "metafunctions" are interdependent; no one could be developed except in the context of the other two. When we talk of the clause as a mapping of these three dimensions of meaning into a single complex grammatical structure, we seem to imply that each somehow "exists" independently; but they do not. There are — or could be — semiotics that are monofunctional in this way; but only very partial ones, dedicated to specific tasks. A general, all-purpose semiotic system could not evolve except in the interplay of action and reflection, a mode of understanding and a mode of doing — with itself included within its operational domain. Such a semiotic system is called a language.

David Rose On "Tonic Focus" As A Probe For Markedness

David Rose replied to Dongbing Zhang on SYSFLING on 18 Mar 2023, at 15:21:

Ah, but isn’t the probe for markedness tonic focus? (Themes underlined)...

unmarked

// those who have guns have them lègally //

marked

equative

// those who have gùns // are the ones who have them lègally //

predicated

// it is those who have gùns // who have them lègally //

So the textual function of the Qualifier is IDENTIFICATION rather than PERIODICITY. Back to Bea’s questions, those is esphoric to the embedded Attribute have guns and them is anaphoric to guns (per ET).

Blogger Comments:

[1] To be clear, "tonic focus" is not the probe for markedness. Tonic prominence is the phonological realisation of the focus of New information. An unmarked Theme can be realised by tonic prominence, making it New as well as unmarked Theme.

[2] As a spoken reading reveals, the most likely first tonic in these instances is hàve, not gùns, making the possessing of guns the focus of New information, which is consistent with the issue at stake.

[3] To be clear, the Theme in this thematic equative construction is unmarked, because it conflates with the Subject in a declarative clause:

[4] To be clear, this non-sequitur is a bare assertion, unsupported by evidence. The Qualifier in question is who have guns. IDENTIFICATION is Martin's rebranding of Halliday & Hasan's (1976) grammatical system of reference as his discourse semantic system, and PERIODICITY is Martin's rebranding of writing pedagogy ('Topic Sentence' etc.) mixed with Halliday's grammatical systems of THEME and INFORMATION.

In terms of SFL Theory, textually, the Qualifier is the referent of a demonstrative reference item (those, the) in the same nominal group, and has the status of Given or New information in an unmarked Theme:

[5] This is essentially true, except for the misleading omission of the very important fact that the analysis actually derives from Cohesion In English (Halliday & Hasan 1976). The only contribution of English Text (Martin 1992) was to relabel Halliday & Hasan 'structural cataphora' as Martin's 'esphora' — a term adapted from Ellis (1971). It is very misleading indeed to credit Martin with Halliday & Hasan's original ideas.

David Rose Misunderstanding Textual Prominence

 After BEATRIZ QUIROZ asked on SYSFLING on 18 Mar 2023, at 02:19:

How would you analyse the following clauses in terms of the ideational (transitivity) and textual metafunctions:

Those who have guns have them legally, …

the majority of people who have guns have them to protect themselves

(no any other punctuation in the original clauses found on the internet)

If “those who have guns” and “the majority of people who have guns” are [nominal groups with] embedded clauses realising a Participant within their respective single clauses, what is the function of “them”? a[s] an Attribute in a attributive possessive clause picking out “guns” from the embedded clause realising the Carrier (IFG4, p. 289)? Or is this some kind of structure giving special textual prominence to “those who have guns” and “the majority of people who have guns”? Or both?

 

David Rose replied on SYSFLING on 18 Mar 2023, 10:07:

Here’s my auty answer. First question -Yes. Second question -No. The Qualifiers function to specify the Carriers’ identity, not to mark them textually. So much of this has been worked out or flagged for further work in English Text, which continually acknowledges the work of others who went before it. ...

 

Blogger Comments:

[1] To be clear, the 'special textual prominence' of those who have guns and the majority of people who have guns, that Quiroz seeks, is simply that each of these unmarked Themes is also coterminous with an information unit:

(The information analysis is based on the tonic falling on the first have and legally in the first clause, and on majority and protect in the clause complex.)

The important difference between the two is that the first has unmarked information structure (Given^New), whereas the second has marked information structure (New^Given).

[2] To be clear, on the one hand, the question is about the nominal groups serving as Carrier, not about the Qualifiers of such nominal groups, and on the other hand, a Qualifier relates to the Thing of the nominal group, so it does not "specify the Carrier's identity".

[3] To be clear, "so much of this" was first "worked out" by Halliday. English Text (1992) is merely Martin's later misunderstanding of Halliday's original theorising, as demonstrated here.

[4] This is misleading, because it is untrue. See David Rose Positively Judging Martin (1992).