Declarative markup in the time of “AI”

Controlling the semantics of tokenized strings

Elisa Beshero-Bondar

GitHub: @ebeshero | Mastodon: epyllia@indieweb.social

Balisage 2023

Link to these slides: https://bit.ly/declare-string

(a paper responding to a prediction that large language models will render descriptive markup unnecessary.)

Natural Language Processing
- process text as sequence of tokens
- measure clusters, co-occurrences
Text Encoding
- OHCO (express structure)
- overt declaration
- usually structural nested containers
- can mark tokenized grams
  - linguistics markup (word-by-word with attributes)

How do we study texts in Digital Humanities?

John Tenniel's illustration of Humpty Dumpty talking to Alice about words

Text generation by approximation
- moving context windows,
- training data
- stochastic process (random sampling from the training data)
- Stats-based "word math": word-embeddings, vector distance
Some intrepid digital humanists learn how to train their own models.
All of us tinker with an enormous one from Google or Meta or OpenAI or in your code editor / word processor:
- good for code correction
- content completion
- writing stuff for you

NLP and Large Language Models

John Tenniel's illustration of Humpty Dumpty talking to Alice about word semantics

“Medium” scale curated projects
Combine XQuery/XSLT with Python
- to reach NLP libraries like NLTK or SpaCY for linguistics data
- to make nifty data viz for fun and profit
XQuery to pull our metadata
- Example: All the artists and performers connected by years in the teiHeaders
A fun XML and Python project:
- Pull all the stage directions in a collection of plays. Determine if tragedies in the collection apply a distinct range of action verbs for character movement in the stage directions.

Text Encoding + NLP methods

(a little unusual, but not new)

John Tenniel's illustration of Humpty Dumpty talking to Alice about word semantics

Work with collation software to compare five different versions of Frankenstein, each one marked up in XML
- (Read about this project in past Balisage papers) For now, a quick summary. . .
- We convert XML nodes into strings, send it through the collation software, and then convert the output into new XML nodes
- Conversion of XML to "stringified" tags + text
- CollateX software makes pair-wise comparisons of five big tokenized strings
- Comparison output is structured data
  - as alignment table
  - as a graph
  - as nodes in a TEI-similar critical apparatus

Special case: Collating XML documents for the Frankenstein Variorum

Including some markup in the comparison

Manuscript (from Shelley-Godwin Archive):

<lb n="c56-0045__main__2"/>It was on a dreary night of November 
<lb n="c56-0045__main__3"/>that I beheld <del rend="strikethrough" 
xml:id="c56-0045__main__d5e9572">
       <add hand="#pbs" place="superlinear" xml:id="c56-0045__main__d5e9574">the frame on
         whic</add></del> my man comple<del>at</del>
<add place="intralinear" xml:id="c56-0045__main__d5e9582">te</add>
<add xml:id="c56-0045__main__d5e9585">ed</add>

1818 (from PA Electronic edition)

<p xml:id="novel1_letter4_chapter4_div4_div4_p1">I<hi>T</hi> was on a dreary 
night of November, that I beheld the accomplishment of my toils.</p>

What matters for meaningful comparison?
- Text nodes
- <del> and <p> markup
What doesn't matter?
- <lb/> elements, attribute nodes
- *In real life (but not for this example), we include the <hi> elements as meaningful markup because sometimes they are meaningful for emphasis.

Normalized strings to compare

MS (from Shelley-Godwin Archive):

It was on a dreary night of November that I beheld 
&lt;del&gt;the frame on whic&lt;/del&gt; my man 
comple&lt;del&gt;at&lt;/del&gt;teed

1818 (from PA Electronic edition)

&lt;p&gt;IT was on a dreary 
night of November, that I beheld
the accomplishment of my toils.&lt;/p&gt;

Tokenize them!

MS (from Shelley-Godwin Archive):

["It", "was", "on", "a", "dreary", 
"night", "of". "November", "that", 
"I", "beheld" 
"&lt;del&gt;the frame on whic&lt;/del&gt;",
"my", "man", 
"comple", "&lt;del&gt;at&lt;/del&gt;", "teed"]

1818 (from PA Electronic edition)

["&lt;p&gt;", "IT", "was", "on", "a", "dreary", 
"night", "of", "November,", "that", "I", "beheld",
"the", "accomplishment", "of", "my", "toils.", "&lt;/p&gt;"]

Project decision: Treat a deletion as a complete and indivisible event:

a ”long token”. This helps to align other witnesses around it.

Nodes on the other side of collation

Real output from the project

(Embedded markup is a little more complicated than our previous example)

<app>
	<rdgGrp n="['that', 'i', 'beheld']">
		<rdg wit="f1818">that I beheld</rdg>
		<rdg wit="f1823">that I beheld</rdg>
		<rdg wit="fThomas">that I beheld</rdg>
		<rdg wit="f1831">that I beheld</rdg>
		<rdg wit="fMS">&lt;lb n="c56-0045__main__3"/&gt;that I beheld</rdg>
	</rdgGrp>
</app>
<app>
	<rdgGrp n="['&lt;del&gt; the frame on whic&lt;/del&gt;',
               'my', 'man', 'comple', 
               '', '&lt;mdel&gt;at&lt;/mdel&gt;', 'te', 'ed', 
               ',', '.', '&lt;del&gt;and&lt;/del&gt;']">
		<rdg wit="fMS">&lt;del rend="strikethrough" 
          xml:id="c56-0045__main__d5e9572"&gt;
			&lt;sga-add hand="#pbs" place="superlinear" 
          sID="c56-0045__main__d5e9574"/&gt;the
	      frame on whic &lt;sga-add eID="c56-0045__main__d5e9574"/&gt; &lt;/del&gt; my man
		  comple &lt;mod sID="c56-0045__main__d5e9578"/&gt; 
          &lt;mdel&gt;at&lt;/mdel&gt;
		  &lt;sga-add place="intralinear" sID="c56-0045__main__d5e9582"/&gt;te
          &lt;sga-add eID="c56-0045__main__d5e9582"/&gt;
          &lt;sga-add sID="c56-0045__main__d5e9585"/&gt;ed
		  &lt;sga-add eID="c56-0045__main__d5e9585"/&gt;
          &lt;mod eID="c56-0045__main__d5e9578"/&gt;
          &lt;sga-add hand="#pbs" place="intralinear"sID="c56-0045__main__d5e9588"/&gt;, 
          &lt;sga-add eID="c56-0045__main__d5e9588"/&gt;.
		  &lt;del rend="strikethrough"
		  xml:id="c56-0045__main__d5e9591"&gt;And&lt;/del&gt;</rdg>
	</rdgGrp>
	<rdgGrp n="['the', 'accomplishment', 'of', 'my', 'toils.']">
		<rdg wit="f1818">the accomplishment of my toils.</rdg>
		<rdg wit="f1823">the accomplishment of my toils.</rdg>
		<rdg wit="fThomas">the accomplishment of my toils.</rdg>
		<rdg wit="f1831">the accomplishment of my toils.</rdg>
	</rdgGrp>
</app>

That ugly but orderly and information-rich combination of "flattened" tags and markup is the basis for all the visualized interfaces we build in our digital edition, the Frankenstein Variorum.

New nodes constructed from output collation data
- Display variance data as "hotspots" in each edition (XHTML + CSS, JS)
- Each version displays its own distinct "hotspots" and displays the other versions at this point

An SVG visualization showing comparative lengths of each collation unit, constructed from collation XML output

Achieving the perfect collation is so. much. work.

How would the language model AI’s handle string collation?

Isn't this what AI is supposed to help us with?

Can the new “AI” collate texts?

Maybe it's optimal for this...
- An LLM operates on word tokens / word embeddings data
- Can it make pair-wise comparisons, based on its tokenizing algorithm?
  - Can it reliably identify differences in pairs of strings? (A to B, A to C, A to D. B to C, B to D, etc.)
  - Can it organize same/similar versions of a text in a group?

Let's test this...

Short test

Can you prepare a data structure to show where these texts are the same and where they differ?

Manuscript: It was on a dreary night of November that I beheld the frame on which my man compleated. And with an anxiety that almost amounted to agony I collected instruments of life around me that I might infuse a spark of being into the lifeless thing that lay at my feet.

1818 edition and 1831 edition: It was on a dreary night of November, that I beheld the accomplishment of my toils. With an anxiety that almost amounted to agony, I collected the instruments of life around me, that I might infuse a spark of being into the lifeless thing that lay at my feet.

Successful result should identify four differences

1818 edition and 1831 edition: It was on a dreary night of November, that I beheld the accomplishment of my toils. With an anxiety that almost amounted to agony, I collected the instruments of life around me, that I might infuse a spark of being into the lifeless thing that lay at my feet.

Manuscript: It was on a dreary night of November that I beheld the frame on which my man compleated. And with an anxiety that almost amounted to agony I collected instruments of life around me that I might infuse a spark of being into the lifeless thing that lay at my feet.

First result: a nice data table

+---------------------+---------------------------+-------------------------+
|       Manuscript     |      1818 edition         |     1831 edition        |
+---------------------+---------------------------+-------------------------+
| It was on a dreary  | It was on a dreary        | It was on a dreary      |
| night of November   | night of November         | night of November       |
| that I beheld the   | that I beheld the         | that I beheld the       |
| frame on which my   | accomplishment of my      | accomplishment of my    |
| man compleated.     | toils.                    | toils.                  |
| And with an anxiety | With an anxiety that      | With an anxiety that    |
| that almost amounted| almost amounted to agony, | almost amounted to agony|
| to agony I collected| I collected the           | I collected the         |
| instruments of life | instruments of life       | instruments of life     |
| around me that I    | around me,                | around me,              |
| might infuse a spark| that I might infuse a     | that I might infuse a   |
| of being into the   | spark of being into the   | spark of being into the |
| lifeless thing that | lifeless thing that lay   | lifeless thing that lay |
| lay at my feet.     | at my feet.               | at my feet.             |
+---------------------+---------------------------+-------------------------+

Nothing is wrong with this, but it leaves the comparison to the reader. It doesn't mark the moments of variance.

Elisa: Let's try to refine this. Where the texts are identical, just show one block of text on the screen. Then when the texts are different, use the table to put the differences side by side.

ChatGPT: Sure! Let's refine the representation to show the identical text blocks once and use a table to display the differing portions side by side. Here's an updated representation:

In this refined representation, the identical text block "It was on a dreary night of November" is displayed once at the top, indicating that this portion is the same in all versions. The table below then shows the differing text portions side by side, highlighting the variations between the Manuscript, 1818 edition, and 1831 edition.

Elisa: Your approach is what I had in mind, good. But you've forgotten where the source texts are different. Can you redo this, taking another look at the source texts?

ChatGPT: Apologies for the oversight. Let's revise the representation to include the differing portions as well. Here's the updated version:

In this refined representation, we have the text differences table with the Manuscript, 1818 edition, and 1831 edition side by side. The portions that are the same in all versions are omitted for brevity, and only the differing sections are displayed.

(Wha-at? you didn't omit anything!)

Summary

Good / promising
- Understands "grouping" the versions
- When prompted it can find some similarities
Problematic
- Doesn't identify all the similarities
- Forgets the actual similarities and differences when re-prompted
- Always prefers the longest possible block alignment
- Seems blind to particular distinctions
Similar results when prompting ChatGPT to output a critical apparatus in TEI XML
- (See lots more examples in my full paper)

Most promising results: Claude.ai

Elisa: You are preparing a critical edition in TEI using critical apparatus markup. Can you prepare a sample of code for comparing two editions based on the following three versions of the same passage?

Remember...
Successful result should identify four differences

Claude.ai: Here is a sample TEI encoding of the critical apparatus comparing those three versions of the passage:

<p>It was on a dreary night of November, that I beheld the 
<app>
  <rdg wit="#MS">frame on which my man compleated</rdg>
  <rdg wit="#1818 #1831">accomplishment of my toils</rdg>. 
</app>
With an anxiety that almost amounted to agony, I collected the instruments of 
life around me, that I might infuse a spark of being into the lifeless thing that 
lay at my feet.</p>

The <app> elements contain the variants, with each <rdg> containing the text as it appears in each witness. The witnesses are identified in the wit attributes using shorthand labels (#MS for the manuscript, #1818 and #1831 for the two editions). This allows you to encode the differences between the versions in a structured way. Additional notes or details about the variants could also be added in the encoding.

Elisa: There's at least one error in your critical apparatus, if we're being strict about punctuation differences. Can you catch and correct it?

Claude.ai: You're right, my initial sample critical apparatus overlooked a punctuation difference between the manuscript and editions. Let me take another look and update the encoding:

<p>It was on a dreary night of November, that I beheld the
<app>
  <rdg wit="#MS">frame on which my man compleated</rdg>
  <rdg wit="#1818 #1831">accomplishment of my toils</rdg>.
</app>  
With an anxiety that almost amounted to agony, I collected the instruments of life 
around me, that I might infuse a spark of being into the lifeless thing that lay
<app>
  <rdg wit="#MS">at my feet</rdg>
  <rdg wit="#1818 #1831">at my feet.</rdg>  
</app>
</p>

Claude.ai:

The editions end the sentence with a period, while the manuscript does not. To accurately encode the differences, I've wrapped the final punctuation in an <app> element as well. Thanks for catching my oversight! Proper encoding requires accounting for all variations, including minor punctuation.

( ...no, that wasn't a real difference...)

Elisa: That wasn't the difference. Both sets of witnesses do end with a period. It's a different moment in the text...

Claude.ai: You're absolutely right, my apologies. Let me take another careful look at the punctuation in the passages:

The key difference is that the manuscript has a comma after "November", while the 1818 and 1831 editions do not. Here is the corrected encoding.

<p>It was on a dreary night of November<app>
  <rdg wit="#MS">,</rdg>
  <rdg wit="#1818 #1831" />  
</app> that I beheld the
<app>
  <rdg wit="#MS">frame on which my man compleated</rdg>
  <rdg wit="#1818 #1831">accomplishment of my toils</rdg>.
</app>
With an anxiety that almost amounted to agony, 
  I collected the instruments of life around me,
  that I might infuse a spark of being into the 
  lifeless thing that lay at my feet.</p>

Claude.ai's TEI apparatus (take 2)

(...Claude has this backwards: 1818 and 1831 have the comma, and the MS lacks it.

Why does it make that mistake?)

Claude's collation performance

Better "out of the box" than ChatGPT
Still unreliable

Troubling things about LLMs attempting collation

Persistent blind spots for one or two tokens that are different
Punctuation differences get lost quickly
Re-prompting doesn't always help
See more examples in the paper
- AI can't tell that transposed text is different
- even drops tokens

Q: So, can the new “AI” collate texts?

...But that wasn't really the point of this paper!

Talking to an LLM chat bot seems declarative...

...but the declarative semantics of a prompt are slippery, inconsistently processed.

A: not really, not validly, not reasonably, not accurately, like a drunk person...

Declarative semantics make it possible to reason about representations; imperative semantics impede.

Michael Sperberg-McQueen, "What does descriptive markup contribute to digital humanities?" (2015)

The technology of descriptive markup

is not value-free or value-neutral

Declarative reasoning in the input is displaced by other imperatives
- like needing to respond to me with explanatory text
Code output is formally correct
- but its semantics are dubious
LLMs can "talk" in formally correct code and natural language syntax...
- ...but they frequently lose context and can't reliably calculate

Conjectures about text-generative AI

A stochastic parrot?

Source: Midjourney prompted by THE DECODER, from

https://the-decoder.com/language-models-defy-stochastic-parrot-narrative-display-semantic-learning/

More conjectures. . .

Does chat get in the way of declaration?
- ChatGPT and Claude can't reliably identify meaningful units of comparison
- We need to train them, but prompt engineering won't be enough.
- Declarative methods should control the imperative programming "under the hood"

A stochastic parrot?

Source: Midjourney prompted by THE DECODER, from

https://the-decoder.com/language-models-defy-stochastic-parrot-narrative-display-semantic-learning/

Let's take a moment to trace how some declarative markup moves through imperative processing...

How declarative methods take control in our project’s XML and Python (1)

Our “long token” method: Controls the token size with XML markup in the source documents:
- Prevents the collation software from aligning small words/punctuation in completely different passages
- Effectively brackets a passage away from micro-comparisons with other versions
- Exerts control over the collation process

inlineVariationEvent = ['head', 'del', 'mdel', 'add', 'note', 'longToken']

Python variable storing a list of "longToken" element names for special treatment in tokenizing

Our Python processing supplies each of these flattened nodes as a single complete indivisible "token" for collation

How declarative methods take control in our project’s XML and Python (2)

”Masking” some element nodes from the collation
- Instruct the collation software to ignore some text during the collation, based on markup

inlineVariationEvent = ['head', 'del', 'mdel', 'add', 'note', 'longToken']

ignore = ['sourceDoc', 'xml', 'comment', 'include', 
          'addSpan', 'handShift', 'damage', 
          'unclear', 'restore', 'surface', 'zone', 'retrace']

Python variable storing a list of element names to ignore

Our Python processing skips over these flattened nodes, not feeding them to the collation software.

Invoke XML Pulldom library in Python:
- ”Pull” the element nodes from the source XML
- Construct a "partial DOM tree"
  - position newline characters as needed to control tokenization
  - Is this something in between a node and string?
- Pre-process the source tree so you control the tokens delivered to the collation software

How declarative methods take control in our project’s XML and Python (3)

 def extract(input_xml):
    """Process entire input XML document, firing on events"""
    doc = pulldom.parse(input_xml)
    output = ''
    for event, node in doc:
        if event == pulldom.START_ELEMENT and node.localName in ignore:
            continue
        # ebb: The following handles our longToken and longToken-style elements:
        # complete element nodes surrounded by newline characters 
        # to make a long complete token:
        if event == pulldom.START_ELEMENT and node.localName in inlineVariationEvent:
            doc.expandNode(node)
            output += '\n' + node.toxml() + '\n'
        # stops the problem of forming tokens that fuse element tags to words.
        elif event == pulldom.START_ELEMENT and node.localName in blockEmpty:
            output += '\n' + node.toxml() + '\n'
        # ebb: empty inline elements that do not take surrounding white spaces:
        elif event == pulldom.START_ELEMENT and node.localName in inlineEmpty:
            output += node.toxml()
        # non-empty inline elements: mdel, shi, metamark
        elif event == pulldom.START_ELEMENT and node.localName in inlineContent:
            output += '\n' + regexEmptyTag.sub('>', node.toxml())
            # output += '\n' + node.toxml()
        elif event == pulldom.END_ELEMENT and node.localName in inlineContent:
            output += '</' + node.localName + '>' + '\n'
        # elif event == pulldom.START_ELEMENT and node.localName in blockElement:
        #    output += '\n<' + node.localName + '>\n'
        # elif event == pulldom.END_ELEMENT and node.localName in blockElement:
        #    output += '\n</' + node.localName + '>'
        elif event == pulldom.CHARACTERS:
            # output += fixToken(normalizeSpace(node.data))
            output += normalizeSpace(node.data)
        else:
            continue
    return output

The editors of the Frankenstein Variorum edition declare

We want:
- An entire deleted passage to be represented as one single, complete event in the text
- That passage must be aligned to a fully comparable unit in the other versions of this document
- When the author crossed out a passage and indicated another to add,
- generate a single long moment of comparison

<app>
	<rdgGrp n="['&lt;del&gt;to his statement, which was delivered&lt;/del&gt;', 
			'to him with interest for he spoke']">
       <rdg wit="fThomas"><del rend="strikethrough">to his statement, 
			which was delivered</del> <add>to him with interest
			for he spoke</add></rdg>
	</rdgGrp>
	<rdgGrp n="['to his statement, which was delivered']">
		<rdg wit="f1818">&lt;longToken&gt;to his statement, which was
				delivered&lt;/longToken&gt;</rdg>
		<rdg wit="f1823">&lt;longToken&gt;to his statement, which was
				delivered&lt;/longToken&gt;</rdg>
		<rdg wit="f1831">&lt;longToken&gt;to his statement, which was
				delivered&lt;/longToken&gt;</rdg>
	</rdgGrp>
</app>

Output declaration

Notice:

trace of the source edition's <longToken> elements in the <rdg> text nodes
@n attribute on the <rdgGrp> elements shows the basis for comparison: the normalized version delivered to the collation software
normalized string removes the <longToken> and <add> elements because we decided these tags were not relevant to the comparison.

What I learned...

Don't expect stochastic processes to speak declaratively
- LLM loses purposeful meaning in calculating text to generate
- subject to shifting context window
- Latent, lurking biases
- Secret training data
Declarative markup is a powerful precision instrument
- Stable, purposeful meaning applied to calculation
- Reliable guide to imperative processes
- Openly announces its perspective and biases
  - Visible, tractable, and usually documented
  - You may not agree with my collation method but you can see what drives it and why it works as it does.

Kordjamshidi P, Roth D, Kersting K., "Declarative Learning-Based Programming as an Interface to AI Systems," Frontiers in Artifical Intelligence 5:755361, 2022 Mar 14. doi:10.3389/frai.2022.755361. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8967162/

The earlier AI problem solvers were expert systems that attempted to model the way experts reason and make decisions using a set of logical rules. Programming languages like Lisp and Prolog were designed to make programming such systems easy even for non-expert users. The idea was to represent the domain knowledge using a set of logical rules, and use the rules in a logical reasoning process hidden from the programmers.

. . .

The expert programs could go beyond an independent set of rules and turn to logical programs with a Turing-complete expressivity, supporting logical inference, for example, by unification and resolution.

A call for declarative methods in today's AI

Kordjamshidi P, Roth D, Kersting K., "Declarative Learning-Based Programming as an Interface to AI Systems," Frontiers in Artifical Intelligence 5:755361, 2022 Mar 14. doi:10.3389/frai.2022.755361. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8967162/

...We emphasize the need to use some fundamental declarative ideas such as first-order query languages, knowledge representation and reasoning techniques, programming languages for multi agent systems, database management systems (DBMS), and deductive databases (DDB). We need to place these ideas within and around ML formalisms including classical ML tools, deep learning libraries and automatic differentiation tools, and integrate them with innovative programming languages and software development techniques, as a way to address complex real-world problems that require both learning and reasoning models.

Is it possible?

Could a text-generative AI system operate in a fully declarative way, or even just a more declarative way?

For my little collation experiment:

What if AI could recognize alternative ways of knowing how to read and compare input strings?
What if AI could evaluate when it's best to switch off its word-embedding calculations and apply another method for "measuring" text?
What if we could ask an AI to access
- not only its LLM word-embeddings data,
- but also a structured database of marked-up resources,
- and also some software libraries of our choosing?

Thanks for listening!

What do you think?

Declarative markup in the time of "AI"

By Elisa Beshero-Bondar

Declarative markup in the time of "AI"

Slides on a talk about the significance of declarative markup during a time of AI driven by large language models. For presentation at Balisage: The Markup Conference, 2023.

1,130

Elisa Beshero-Bondar PRO

Professor of Digital Humanities and Chair of the Digital Media, Arts, and Technology Program at Penn State Erie, The Behrend College.

Declarative markup in the time of “AI”

Controlling the semantics of tokenized strings

How do we study texts in Digital Humanities?

NLP and Large Language Models

Text Encoding + NLP methods

(a little unusual, but not new)

Special case: Collating XML documents for the Frankenstein Variorum

Including some markup in the comparison

Normalized strings to compare

Tokenize them!

Nodes on the other side of collation

Real output from the project

Achieving the perfect collation is so. much. work.

How would the language model AI’s handle string collation?

Isn't this what AI is supposed to help us with?

Can the new “AI” collate texts?

Let's test this...

Short test

Successful result should identify four differences

First result: a nice data table

Summary

Most promising results: Claude.ai

Remember... Successful result should identify four differences

Claude's collation performance

Troubling things about LLMs attempting collation

Q: So, can the new “AI” collate texts?

...But that wasn't really the point of this paper!

A: not really, not validly, not reasonably, not accurately, like a drunk person...

Conjectures about text-generative AI

More conjectures. . .

Let's take a moment to trace how some declarative markup moves through imperative processing...

How declarative methods take control in our project’s XML and Python (1)

How declarative methods take control in our project’s XML and Python (2)

How declarative methods take control in our project’s XML and Python (3)

The editors of the Frankenstein Variorum edition declare

Output declaration

What I learned...

A call for declarative methods in today's AI

Is it possible?

Declarative markup in the time of "AI"

More from Elisa Beshero-Bondar

Remember...
Successful result should identify four differences