Leveraging the Power of AI and Schematron for Content Verification and Correction

Schematron And
Artificial Intelligence
© 2023 Syncro Soft SRL. All rights reserved
Octavian Nadolu, Syncro Soft
octavian_nadolu@oxygenxml.com
@OctavianNadolu

AI in Schematron and SQF
Agenda
● Artificial Intelligence (AI)
● OpenAI/Generative Pre-trained Transformer
● Schematron and Schematron Quick Fixes (SQF)
● Implementing AI in Schematron and SQF
● Examples of AI-driven Schematron and SQF
Solutions
● Benefits of Using AI in Schematron and SQF

What is Artificial Intelligence?
Artificial Intelligence (AI) is a branch of computer
science dealing with the simulation of intelligent
behavior in computers.

Artificial Intelligence History
● AI history dates back to ancient times
● In 1950s AI began to take shape - researchers
began to use computers to try and simulate
human intelligence
● AI program by mathematician Alan Turing
● Expert system by Edward Feigenbaum in the
1970s and the emergence of neural networks
in the 1980s
● Today, AI is used to automate processes,
improve efficiency, and solve complex
problems

AI in Natural Language Processing
● Natural language processing is a subfield of AI that focuses
on enabling machines to understand and generate human
language
● Machine learning involves training algorithms to learn
patterns in data
● Deep learning is a type of machine learning that uses neural
networks

Generative Pre-trained Transformer(GPT)
● Network models that uses the transformer architecture
● A type of LLM (Large Language Model)
● Pre-trained refers to the model being trained on a large corpus of data
● An application is ChatGPT developed by OpenAI

Transformers
● A transformer is a deep learning model architecture used for
processing data
● The transformer architecture is based on the idea of self-attention
● Introduced in a research paper titled "Attention Is All You Need" in
2017

Embeddings
● A mathematical representations of words, sentences, or documents in
a continuous vector space
● Encode similar words with similar embeddings
● Embeddings have become a fundamental component of many NLP
tasks

OpenAI
OpenAI is an open-source research organization that works to advance
artificial intelligence (AI)

OpenAI Application
OpenAI has trained language models that are very good at understanding
and generating text
● Text Summarization
● Natural Language Processing
● Text Generation
● Machine Translation
● Text Classification

What is Schematron?
A language for making assertions in documents

Schematron and AI
Automatically verify documents using an AI model
Examples of verification with AI:
– Is active/passive voice used in the description?
– Is this written according to the style guide?
– Does the topic answers to this <question>?
– Is spelling and grammar correct?

Schematron Quick Fix (SQF)
● SQF is an extension of the ISO Schematron
● User-defined fixes for Schematron assert/report

SQF and AI
● Correct problems in documents using AI
● Examples:
– Rephrase to use active voice
– Rephrase to have 20 words
– Rephrase paragraph to answer to the following question
– Correct spelling and grammar

Implementation of AI in Schematron
● Using XSLT functions
● Using extension functions
– xs:string ai:transform-content(xs:string instruction, xs:string content)
– xs:boolean ai:verify-content(xs:string instruction, xs:string content)
<xsl:function name="ai:chatGPT">
<xsl:param name="userInput"/>
<xsl:variable name="url" select="'https://api.chatgpt.com/v1/chatbot/question'"/>
<xsl:variable name="requestBody" select="concat('{', '"text":"',
$userInput,'"', '}')"/>
<xsl:variable name="response" select="document(concat($url, '?apiKey=',
'<your_api_key>'))//response"/>
<xsl:sequence select="$response"/>
…..
</xsl:function>

Functions Built-in Prompt
● A prompt represents instructions and context given to a language
model to perform a task
● Functions can provide a specific built-in prompt
– ai:verify-content()
“You are a technical writer and you need to verify the following and respond with
true or false:“ + Is active voice used in the description? + content
– ai:transform-content()
“You are a developer and you need perform the following task:“ + Rephrase
to use active voice + content
https://github.com/f/awesome-chatgpt-prompts

Examples of Schematron Rules

Check if the text is consistent
● Example of a rule that checks if the text is easy to read and understand
The text in not easy to read and understand

Check if the text is consistent
● Rule that verifies if the text is easy to read and understand
<sch:rule context="p">
<sch:assert test="ai:verify-content('Is the text easy to read and understand?', .)">
The text in not easy to read and understand</sch:assert>
</sch:rule>

Correct text consistency
● Example fix that makes the text easy to read and understand
Correct the consistency of the text

Correct the text consistency
● SQF fix that corrects the text to be easy to read and understand
<sqf:fix id="rephrase">
<sqf:description>
<sqf:title>Correct the consistency of the text</sqf:title>
</sqf:description>
<sqf:replace match="text()" select="ai:transform-content(
'Correct the text to be easy to read and understand', .)"/>
</sqf:fix>

Check the text voice
● Example of a rule that checks if the text uses active voice
In the description we should use active voice

Check the text voice
● Rule that verifies if the text voice is active
<sch:rule context="shortdesc">
<sch:assert test="ai:verify-content('Is active voice used?', .)">
In the description we should use active voice.</sch:assert>
</sch:rule>

Correct the text voice
● Example fix that reformulates the text to use active voice
Reformulate the text to use active voice

Correct the text voice
● SQF fix that that reformulates the text to use active voice
<sqf:description>
<sqf:title>Reformulate the text to use active voice</sqf:title>
</sqf:description>
<sqf:replace match="text()" select="ai:transform-content('
Reformulate to use active voice', .)"/>
</sqf:fix>

Answer to question
● Example of a rule that checks if the text answers a specified question
The test does not answer to the question "What is OpenAI?"

Check the question
● Rule that verifies if the text does answer a specific question
<sch:rule context="p[@id='openai']">
<sch:assert test="ai:verify-content('Does it answers to the question: What is OpenAI?', .)">
The test does not answer to the question "What is OpenAPI?" </sch:assert>
</sch:rule>

Reformulate text
● Example fix that reformulates the text to answer a specific question
Reformulate the text to answer to the question:
What is OpenAI?

Reformulate text
● SQF fix that that reformulates the text to answer the
question “What is OpenAI?”
<sqf:description>
<sqf:title>Reformulate the text to answer to the question:
What is OpenAI?</sqf:title>
</sqf:description>
'Reformulate the text to answer to the question: What is OpenAI?', .)"/>
</sqf:fix>

AI and Schematron
● Advantages of using AI with Schematron
– Verify your documents using AI power
– Define the instructions to be sent to the AI engine
– Control the content to be verified
– Control the content that is sent
– Automation of the process
● Challenges
– High cost for validation as you type or multiple validations
– The response from the AI server in not instant
– Responses can sometimes be inaccurate

AI and Schematron Quick Fixes
● Advantages of using AI with SQF
– Use the power of AI to correct the content
– Control the content that is sent
– Control the content to be modified
– Automation of the process
● Challenges
– The response from the AI server is not instant
– Responses can sometimes be inaccurate

AI and Schematron Quick Fixes
● Use AI just to correct the problems
– Reduce the cost
– You can use the validation as you type
– The user decides when to call the AI

Check the number of words
● Example of a rule that checks the number of words from the
description
The description must contain less than 50 words.

Check the number of words
● Rule that verifies the number of words from the shortdesc element
<sch:report test="count(tokenize(.,'s+')) > 50">
The description must contain less than 50 words.</sch:report>
</sch:rule>

Correct text to contain less words
● Example fix that reformulates the phrase to contains less than 50 words
Reformulate phrase to contain less than 50 words

Correct text to contain less words
● SQF fix that reformulates the phrase to have less than 50 words
<sqf:description>
<sqf:title>Reformulate phrase to contain less than 50 words</sqf:title>
</sqf:description>
'Reformulate phrase to contain less than 50 words', .)"/>
</sqf:fix>

Check if text should be a list
● Example of a rule that checks it the text should be converted to a list
The text should be converted to a list.
<body>
<p> - Is active/passive voice used in the description?
- Is this written according to the styleguide?
- Does the to pic answers to this "question"?
- Is spelling and grammar correct?</p>
</body>

Check if text should be a list
● Rule that verifies the text from a paragraph should be converted to a
list
<sch:rule context="p">
<sch:report test="contains(., '- ')">
The text should be converted to a list</sch:report>
</sch:rule>

Create a list from phrases
● Example of fix that generates an unordered list from a set of phrases
Create a list from the phrases from the paragraph
<p> - Is active/passive voice used in the description?
- Is this written according to the styleguide?
- Does the to pic answers to this "question"?
- Is spelling and grammar correct?</p>
<ul>
<li>Active/passive voice used?</li>
<li>Written according to styleguide?</li>
<li>Topic answers question?</li>
<li>Spelling and grammar correct?</li>
</ul>

Create a list from phrases
● SQF fix that creates a list from a set of phrases
<sqf:fix id="replace">
<sqf:description>
<sqf:title>Create a list from the phrases from the paragraph</sqf:title>
</sqf:description>
<sqf:replace match="text()">
<xsl:value-of select="ai:transform-content(
'Create a Dita unorderd list with an item from each phrase', .)"
disable-output-escaping="yes"/>
</sqf:replace>
</sqf:fix>

User Entry
The instruction to correct the problem
is specified by the user

Check technical terms
● Example of a rule that checks if the technical terms are not explained
adequately
The text uses technical terms that are not explained adequately

Check technical terms
● Rule that verifies if the technical terms are explained adequately
<sch:report test="ai:verify-content('Are the technical terms explained ambiguous?', .)"
The text uses technical terms that are not explained adequately.</sch:report>

Correct terms
● Example fix that allows the user to specify how to reformulate the
phrase
Specify how to reformulate the phrase

Correct terms
● SQF fix that allows the user to specify the prompt that will be sent to
the AI
<sqf:fix id="reformulateUser">
<sqf:description>
<sqf:title>Specify how to reformulate the phrase</sqf:title>
</sqf:description>
<sqf:user-entry name="userInput" default="'
Reformulate phrase and replace the ambiguous terms with a more accurate one'">
<sqf:description><sqf:title>How to correct:</sqf:title>sqf:description>
</sqf:user-entry>
<sqf:replace match="text()" select="ai:transform-content($userInput, .)"/>
</sqf:fix>

Generate AI Fix
● Generate the fix from the Schematron message
● Uses the Schematron rule context as the content to correct
– System prompt: Act as a developer. Perform the following task:
The description must contain less than 50 words
– User prompt: The content from the “shortdesc” element
– Operation: Replace the text from “shortdesc” with the one from AI
<sch:report test="count(tokenize(.,'s+')) > 50">
The description must contain less than 50 words.</sch:report>
</sch:rule>
Generate AI Fix

Automatic Fix
● Advantages:
– No need to create fixes in Schematron
– A fix is generated automatically from each Schematron message
– Fixes can also be generated for other error messages
● Challenges
– The message need to be as concise as possible so that the AI can use
– Sometimes you need to change the context for the fix or perform different
operations

Create Schematron using AI
● An assert that verifies the number of words to be 10
● An assert that verifies if there is an email in text
<sch:assert test="count(tokenize(., 's+')) = 10">There should be exactly 10.</sch:assert>
<sch:assert test="matches(., 'b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+.[A-Z|a-z]{2,}b')">
There is no email in the text</sch:assert>
AI
AI

Conclusion
● Schematron can use the power of AI to verify content
● Correct content using SQF and AI
● Generate fixes automatically using AI
● Develop Schematron with AI
● AI constantly improving and growing

Resources
● https://openai.com/
● https://platform.openai.com/docs/guides/chat
● http://schematron.com/
● http://schematron-quickfix.github.io/sqf

Questions?
Octavian Nadolu
Project Manager at Syncro Soft
octavian.nadolu@oxygenxml.com
Twitter: @OctavianNadolu
LinkedIn: octaviannadolu

Leveraging the Power of AI and Schematron for Content Verification and Correction

More Related Content

Similar to Leveraging the Power of AI and Schematron for Content Verification and Correction (20)

More from Octavian Nadolu (20)

Recently uploaded (20)

Leveraging the Power of AI and Schematron for Content Verification and Correction