Turn Unstructured Data into Business Actions

Trisotech.com
Kw-copilot.ai
Turn Unstructured Data
into Business Actions
The Role of Generative AI (GenAI)

Trisotech.com
Kw-copilot.ai
Intelligent Digital Assistant (IDA)
We have been working on creating Intelligent Digital Assistants (IDA) for a few years:
 We are creating IDA architecture meant to support Knowledge Workers.
 A key capability for our IDA architecture is the ability to consume natural language
communications.
 Currently developing the third generation of IDAs.

Trisotech.com
Kw-copilot.ai
Empowering Knowledge Workers with
Intelligent Digital Assistants

Trisotech.com
Kw-copilot.ai
Intelligent Digital Assistant (IDA)
 Some of the goals of our IDA architecture

Trisotech.com
Kw-copilot.ai
We favor an Assistant/Co-pilot approach

Trisotech.com
Kw-copilot.ai
Leveraging a Nero-Symbolic
Artificial Intelligence Architecture

Trisotech.com
Kw-copilot.ai
Leveraging Speech Acts in IDAs
Identifying and interpreting the intent of a user’s natural language communication to enable
meaningful actions and responses.
o Intent Detection: Understanding the user’s goal behind their words.
• Example: "Book a flight to New York.“
• Identified Intent: Directive (user wants the system to perform a task).
o Context Awareness: Determining the function of a statement based on context.
• Example: "Can you complete this?"
• Context: In a conversation regarding a report.
• Interpreted Intent: Directive (polite request for action).
o Action Mapping: Associating speech acts with specific actions or responses.
• Example: "Send an email to John about the meeting.“
• Mapped Action: Draft and send an email with a predefined template or subject line.

Trisotech.com
Kw-copilot.ai
The advent of Generative AI
The new wave of Large Language Models (LLMs) research has offered new
and enhanced language capabilities.

Trisotech.com
Kw-copilot.ai
The advent of Generative AI

Trisotech.com
Kw-copilot.ai
Rebalancing the Nero-Symbolic Mixture
 Rebalancing: Our reliance on symbolic decisions for explicit knowledge encoding is
being reduced, as new neural capabilities can now effectively deliver semantic lifting that
complements or even partially substitutes what previously required symbolic decisions.

Trisotech.com
Kw-copilot.ai
Net Effect of
Rebalancing
Our neuro-symbolic architecture
becomes more adaptive, scalable,
and versatile.

Trisotech.com
Kw-copilot.ai
Named Entity Recognition (NER):
What is it?
Identifies specific entities such as names, dates, locations, or
organizations within text.
 Example:
o From: "Apple launched the iPhone 15 in Cupertino, California, on September 12,
2023.“
o Entities: “Apple” (Organization), “iPhone 15” (Product), “Cupertino, California”
(Location), “September 12, 2023” (Date)

Trisotech.com
Kw-copilot.ai
Named Entity Recognition (NER):
Transition
Previous way
 Worked with predefined, static categories
(e.g., Person, Location, Organization).
 Required manual updates to include new
entity types or adapt to domain-specific
needs.
 Limited understanding of complex or
ambiguous contexts (e.g., "Apple" as fruit
vs. company).
 Requires significant effort for
customization (e.g., adding new rules,
retraining).
New way leveraging LLMs
 Easily adapts to emerging entities or
domain-specific vocabularies without
retraining.
 Highly context-aware, distinguishing
between entities based on usage and
sentence context.
 Highly customizable through prompt
engineering or fine-tuning

Trisotech.com
Kw-copilot.ai
Data Structuring:
What is it?
 Transforms unstructured text into structured formats (e.g., tables, JSON, or graphs).
 Example:
o From: "The supplier is Bill Shep. His phone number is 540-817-4567. He lives in 3807 Rustic
Av, Liverpool, WV,. He is a Sales Engineer with Pennington Sipper, Inc. His email is
billshep@pensip.com. .“
o JSON:
{
"Name": "Bill Shep",
"Type": "Supplier",
"Phone": "540-817-4304",
"Address: "3807 Rustic Av, Liverpool, WV,25252
"Position": "Sales Engineer",
"email": "billshep@pensip.com"
}

Trisotech.com
Kw-copilot.ai
Data Structuring:
Transition
Previous way
 Relied on rule-based approaches, regex,
templates, or predefined parsing.
 Limited flexibility; required custom rules
or scripts for new data formats.
 Required manual adjustments for new
fields or entities.
 Highly flexible, adapting to a wide range
of text inputs without explicit
programming.
 Extracts and structures data using
contextual understanding, even with
ambiguous input.
 Easily customizable via prompt
engineering or fine-tuning.

Trisotech.com
Kw-copilot.ai
Topics Identification:
What is it?
Detects overarching themes or subjects in text (e.g., business,
sports, politics, technology)
 Example:
o From: "The supplier is Bill Shep. His phone number is 540-817-4567. He lives in
3807 Rustic Av, Liverpool, WV,. He is a Sales Engineer with Pennington Sipper,
Inc. His email is billshep@pensip.com. Schedule a sales meeting on August 23 at
07:00 with him and prepare a draft supplier introduction email.“
o Identified Topics: "Contact manager," "Calendaring," "Email messaging"

Trisotech.com
Kw-copilot.ai
Topics Identification:
Transition
Previous way
 Relied on statistical models and rule-
based systems
 Required structured or semi-structured
text.
 Required preprocessing like stop-word
removal.
 Limited to predefined or explicitly derived
topics from the dataset.
 Struggled to identify emerging trends or
to adapt to rapidly evolving topics.
 Infers topics contextually
 Works on raw, unstructured text with
minimal preprocessing required.
 Highly flexible and can handle dynamic or
emerging topics.
 Excels at detecting emerging trends and
adapting to dynamic content.

Trisotech.com
Kw-copilot.ai
Intent Detection:
What is it?
 Understanding the user’s goal behind their words.
 Example:
o From: "Can you help me schedule a meeting and send a summary to my team?“
o Traditional: Confusion or single-label assignment.
o LLM: Recognizes both "scheduling" and "summarizing" as intents.

Trisotech.com
Kw-copilot.ai
Intent Detection:
Transition
Previous way
 Limited to explicit keywords or patterns;
struggles with ambiguous contexts.
 Works well for predefined intents but
struggles with overlapping or emergent
intents.
 Requires significant manual effort to scale
to new domains or applications.
 Deep contextual understanding, even for
complex text.
 High accuracy across diverse datasets due
to broad pretraining, even for rare or
unique inputs.
 Scales efficiently to diverse applications
with minimal adjustments.

Trisotech.com
Kw-copilot.ai
Text Generation:
What is it?
 Produces coherent and contextually relevant text (e.g., chat
responses, summaries, creative writing).
 Example:
o From: "Write a short message for a chatbot responding to a customer asking
about the return policy."
o Generated Text: "Thank you for reaching out! Our return policy allows you to
return items within 30 days of purchase. Please ensure the items are in their
original condition. Let us know if you need further assistance!"

Trisotech.com
Kw-copilot.ai
Text Generation:
Transition
Previous way
 Uses template-based approaches, rule-
based systems.
 Required pre-designed templates or
significant engineering for new tasks.
 Limited scalability; more suited for
structured or repetitive tasks.
 Dynamically adapts to new domains,
tones, and tasks with minimal
adjustments.
 Excels in producing nuanced, domain-
specific, and complex text.
 Easily customized via prompt engineering
or fine-tuning.
 Scales efficiently for diverse and complex
applications.

Trisotech.com
Kw-copilot.ai
Streamline Knowledge Worker Processes
 Knowledge worker processes are notorious for being impervious to RPA
and BPA due to their highly collaborative nature.
 With our IDA architecture, we can:
oIdentify Human Interaction Gaps.
oUnderstand Roles in a Collaboration.
oAddress Communication Needs.
oDetermine Intent and Act.

Trisotech.com
Kw-copilot.ai
Demo

Trisotech.com
Kw-copilot.ai
Using BPMN for Effective Prompt
Management
 Centralize all calls to prompts in one process.
 Store prompts in datastores
 The interactions between the prompt and the LLM is complex
 Consider provisions for “stress” testing prompts
Notional BPMN Diagram to ease understanding

Trisotech.com
Kw-copilot.ai
Guidelines to Minimize Hallucinations
 Use simple and explicit prompts.
o Ambiguity in prompts leads to misinterpretation, so keeping them simple and explicit ensures the model
understands the task clearly.
 Describe what the prompt does in a preamble.
o A preamble provides context and clarifies the purpose of the task, aligning the model’s response with your
expectations.
 Provide examples.
o Examples demonstrate the desired output format, guiding the model to better generalize and produce accurate
responses.
 Ask the LLM to evaluate the prompt for potential risks.
o Asking the LLM to self-evaluate identifies sources of misinterpretation and encourages safer, more responsible
outputs.
 Avoid using prompts that infer or categorize the contents of the free text into a result.
o "The apples have shipped"->Fruit shipment
o Prompts that infer without explicit context risk fabricating information, so focusing on grounded details ensures
accuracy.

Trisotech.com
Kw-copilot.ai
Key Take Aways
 Generative AI is Transformative
o LLMs revolutionize processing and acting on unstructured data.
o Enable context-aware, adaptive, and efficient natural language understanding.
 Rebalanced our Neuro-Symbolic architecture based on enhanced NLP Capabilities from LLMs
• Topic Identification – Detect emerging themes dynamically.
• NER – Context-aware identification of entities.
• Data Structuring – From unstructured text to actionable formats.
• Sentiment Analysis – Nuanced emotional understanding.
• Text Generation – Domain-specific, contextual content creation.
 Practical Considerations for Success
o Speech acts enhance functional intent understanding enabling meaningful actions and responses.
o Effective prompt management ensures reliability and ethical outputs.

Trisotech.com
Kw-copilot.ai
Any questions?
THANKS!
27

Turn Unstructured Data into Business Actions

More Related Content

Similar to Turn Unstructured Data into Business Actions (20)

More from Denis Gagné (20)

Recently uploaded (20)

Turn Unstructured Data into Business Actions