1 of 83

3 of 83

3

Content

Foundations

Why Social Media Data
LLMs and its challenges
NeuroSymbolic Approach
Types of Knowledge Infused Learning and its advantages

Neurosymbolic Approach for an health application

Use-Case COVID-19

Hands-on Session

4 of 83

4

Content

Foundations

Why Social Media Data
LLMs and its challenges
NeuroSymbolic Approach
Types of Knowledge Infused Learning and its advantages

Neurosymbolic Approach for an health application

Use-Case COVID-19

Hands-on Session

5 of 83

Why Social Media Data?

5

6 of 83

6

The youngest adults stand out in their social media consumption

88% of 18- to 29-year-olds indicate that they use any form of social media.

By Pew Research Center “Social Media Use Report 2018”

5.2 Billion

Users

global scale of social media usage, it’s pretty eye-opening. As of October 2024, there are 5.2 billion active social media users worldwide (Statista, 2024). That’s more than 63% of the global population actively engaging on platforms like X, Facebook, TikTok, and Instagram.

What does this mean? Well, billions of posts are created every single day, and this constant flow of user-generated data provides us with a significant opportunity. Social media users do not just share what’s happening, it also reflects collective human behavior in real time.

Now, when we look at the world here, you can see some interesting usage patterns for different regions. For example:

Northern Europe has one of the highest social media usage rates at 78.8%, showing how deeply integrated these platforms are in everyday life.
in regions like Middle Africa and Eastern Africa, social media use is much lower at around 10%, likely due to differences in internet access and infrastructure.
And then there’s a global average of 63.8%, which highlights just how interconnected we are as a society, through SOCIAL MEDIA.

This is important because social media data gives us a real-time pulse on what people are thinking, feeling, and experiencing. It allows us to track trends, understand human responses during crises like the COVID-19 pandemic, and analyze behaviors at both local and global scales. This makes social media not just a tool for communication but a powerful resource for research in areas like public health, mental health, and societal well-being.

Ultimately, the more we understand how people use social media and what they express, the better we can leverage this data to address real-world problems.

7 of 83

7

“Information that comes directly from consumers,

often via social media, is deemed more helpful than data

from reports or government research.”

Insights from Social Media, How useful?

Social media isn’t just about sharing updates and scrolling through your feed, it can become one of the sources of insight for decision-making. For example: based on a survey from World Economic Forum, CEOs rank social media as their number one source for understanding trends, even above traditional media, government data, or commissioned research.

Of course, these insights are extracted from raw social media data through big social data analytics, this shows the potential how social media can be useful to understand the public or customer opinion about certain things, your company, products, services, OR health and well-being of the people.

Why makes social media so valuable? It’s because the data comes directly from people or consumers, it’s real-time and dynamic, and reflects what people are thinking, feeling, and saying. This makes it incredibly useful for several critical areas:

Consumer Sentiment Analysis for Brand Health Monitoring: Companies use social media to track how consumers feel about their products or services. This helps brands respond quickly and maintain their reputation.�
Competitor Benchmarking and Market Trends Prediction: Social media data allows companies to see how they stack up against competitors. By analyzing trends and public discussions, businesses can predict shifts in the market and adapt their strategies ahead of time.�
Political and Social Movement Monitoring: Social media is also a powerful tool for tracking opinions and public sentiment around political or social issues.�
Mental Health and Well-Being Monitoring: Perhaps one of the most impactful applications of social media data is in understanding mental health trends. Platforms reflect emotional expressions, distress signals, and conversations about anxiety, depression, and overall well-being. For example, during the COVID-19 pandemic, social media data helped track the rise in mental health problems and substance use.

Social media can be a critical source of insight for understanding collective behavior, tracking trends, and responding to public sentiment in ways that traditional data sources just can’t do.

8 of 83

8

Contexts where Social Media Matters: the Good & the Bad

A spectrum to demonstrate the good and the bad on social media.

Kursuncu, U., Purohit, H., Agarwal, N., & Sheth, A. (2021). When the Bad is Good and the Good is Bad: Understanding Cyber Social Health through Online Behavioral Change. IEEE Internet Computing, 25(01), 6-11.

Monitoring Epidemic Covid-19, Zika

Help

Fighting Depression

Disaster Relief

Marketing

Understanding & Predicting Consumer Behavior

Monitoring Opioid Usage

Extremism

Illicit Drugs

Disinformation

Harassment

More Good

More Bad

More Good

While social media can be a powerful tool for good, it can also be a double-edged sword.

On the positive side, social media plays a crucial role in areas like:

Marketing: Businesses can better understand and reach their audiences, better understand trends to provide more relevant and meaningful experiences, products and services.

Social media provides real-time insights into what people want, think, and feel, which helps organizations make better decisions.�

Monitoring Opioid Usage: Platforms can reveal signals of substance abuse trends, which are critical for public health interventions.�
Monitoring Epidemics: For example, during the COVID-19 or Zika outbreaks, social media was used to track disease spread and public health responses.

Fighting Depression: Online communities have become lifelines for people struggling with mental health issues by offering support and reducing stigma.

On the negative side, social media has challenges we can’t ignore:

Extremism, Harassment and Illicit Drugs: Social media can amplify harmful content, including online extremism, cyberbullying and illegal drug markets.

More specifically, from cyberbullying to severe forms of online abuse, harassment on social media has serious consequences for mental health and well-being.

What this spectrum shows us is that social media reflects both the good and the bad of human behavior.

For social media analysis, understanding this dynamic is critical. By analyzing social media content, we can focus on identifying early signals of mental health issues, tracking public sentiment, and addressing social problems like opioid abuse or misinformation.

It may help us to leverage the positive potential of social media while recognizing and mitigating its challenges.

9 of 83

1,213,046 death in the U.S.*

103,436,829 positive cases

7,077,725 death globally**.

776,973,432 positive cases

Multiple lockdowns, guidance for staying at home, social distancing, accelerated the use of technology, including social media.

9

COVID-19; Sudden Emergence leads to Rapid Adaptation

* https://covid.cdc.gov/covid-data-tracker/

** https://covid19.who.int/

COVID-19 changed everything almost overnight. The sudden emergence of the virus led to global disruptions that forced us all to adapt quickly; socially, economically, and technologically.

Some of the significant impact was:

In the U.S. alone, there were over 1.2 million deaths and more than 103 million positive cases reported.

Globally, it was over 7 million deaths and close to 777 million positive cases.

To slow down the spread, governments around the world implemented multiple lockdowns.

However, these lockdowns brought new challenges: staying at home, practicing social distancing, and navigating an unprecedented reliance on technology. And this is where social media played a big role.

Platforms like Twitter (now X), Facebook, and Reddit became essential tools for communication, connection, and information sharing. People used social media to stay informed about the virus, express their fears, or connect with friends and family during isolation. It also accelerated trends like remote work, online schooling, and even mental health discussions.

The rapid adoption of social media during the pandemic gave an opportunity to analyze public sentiment and emerging mental health trends in real-time.

Whether it was people sharing struggles with anxiety, job loss, or loneliness, the online discourse reflected the collective human experience during a global crisis.

10 of 83

Early detection of pandemics and outbreaks

Social media data analysis during the COVID-19 pandemic led to early outbreak predictions [1].

Pandemic leading to mental health crisis?
Real-time detection of mental health crises (e.g., depression, anxiety).
Awareness and resource allocation to respond emerging mental health issues.

10

COVID-19, Public Health and Social Media

Shi, B., Huang, W., Dang, Y., & Zhou, W. (2024). Leveraging social media data for pandemic detection and prediction. Humanities and Social Sciences Communications, 11(1), 1-18.

The COVID-19 pandemic brought not just a health crisis but also a mental health crisis, and social media played a major role in, perhaps amplifying, but also understanding these challenges.

During COVID-19, real-time analysis of social media data could have helped predict patterns, for early warnings that traditional systems sometimes missed, just because social media was a major platform with discussions about symptoms, hospitalizations, and hotspots early on.

But beyond the virus itself, the pandemic seemingly triggered a mental health crisis. According to the World Health Organization, there was a 25% increase in the prevalence of anxiety and depression worldwide. People faced isolation, job loss, financial uncertainty, and fear, which significantly impacted mental well-being.

This is where social media becomes very valuable:

It allows for real-time detection of mental health crises. By analyzing public posts, we can identify signals of rising depression, anxiety, or stress.

Social media can also help raise awareness and guide the allocation of resources to the areas and communities that need it most.

11 of 83

Prevalence in Mental Health Issues & Online Toxicity

11

51%

↑15 YoY

Teens experienced some form of Online Harassment [1]

52%

↑12 YoY

Online Harassment

Ever Experienced among American Adults [1]

37%

↑10 YoY

Severe Online Harassment, Sexual, physical threats, swatting, doxing and sustained harassment [1]

Anti-Defamation League (ADL), Online Hate and Harassment Report: The American Experience 2023. https://www.adl.org/resources/report/online-hate-and-harassment-american-experience-2023
Pew Research Center, Parenting in America Today, 2023. https://www.pewresearch.org/social-trends/2023/01/24/parenting-in-america-today/
CDC Adolescent Behaviors and Experiences Survey (ABES), 2021. https://www.cdc.gov/abes/index.html
Forsberg, J. T., & Thorvaldsen, S. (2022). The severe impact of the COVID-19 pandemic on bullying victimization, mental health indicators and quality of life. Scientific reports.

40%

Children struggling with anxiety or depression, reported by parents [2]

37% of U.S. adolescents had regular mental health struggles during COVID-19 pandemic [3].

“... increased prevalence in bullying, more mental health problems and significantly reduced quality of life compared to before the pandemic” [4]

one of the biggest challenges we’ve seen during and after the COVID-19 pandemic: the impact of pandemic on mental health, especially for young people.

37% of U.S. adolescents reported regular struggles with mental health during the pandemic, such as anxiety, depression.

When we look closer, about 40% of parents said their kids were dealing with anxiety or depression. At the same time, these online spaces became more toxic: 51% of teens experienced some form of online harassment, which was up by 15% year-over-year. Clearly, the pandemic intensified a lot of the issues people face online.

And it’s not just teens, adults are feeling this too. 52% of American adults reported experiencing online harassment, up 12% year-over-year. severe harassment affected 37% of users, which is up by 10% YoY.

A study from Scientific Reports found: there’s been an 'increased prevalence in bullying, more mental health problems, and a significantly reduced quality of life compared to before the pandemic.'

This situation shows how connected mental health struggles can be to what’s happening in digital spaces. For this study, it’s a big reason why we need to analyze social media data in real time. So that we can detect trends and early signs of mental health crises, and understand how toxic behaviors emerge, helping us work toward better solutions.

12 of 83

A Social Media Data Concern: Content Quality

Social data often contain noise and irrelevant content:

Semantic filtering in preprocessing�

Distinguish genuine mental health indicators in data from unrelated or satirical posts.

Contextual understanding through domain specific knowledge graphs in model learning.

Challenge; content moderation on social media platforms for problematic content or gain awareness about potential mental health crisis.

12

One of the biggest challenges in analyzing social media data is making sure of content quality. Social platforms generate vast amounts of data, but much of this is noisy data, such as spam, irrelevant posts, or bot-generated content.

Another issue is distinguishing genuine mental health indicators from unrelated or satirical posts. Social media often includes ambiguous language, sarcasm, or even memes.

To address this, we use semantic filtering during preprocessing to pick more meaningful signals.

By incorporating domain-specific knowledge graphs, our models can achieve a deeper contextual understanding, allowing us to filter out irrelevant content and focus on true mental health signals in the data.

As these issues are prevalent for most of social media data, they are also relevant to other challenges, such as effective content moderation. While social media platforms have heavily relied on human moderators in the past, they are increasingly turning to AI-driven approaches. This shift highlights the same challenges and the potential for innovative, scalable solutions.

13 of 83

Increasing Reliance on AI -Content Moderation

13

The abundance of online big (social) data enabled recent breakthroughs in AI

Limitations: Bias and toxicity have seeped into models�

Detection and Countering challenging

Ambiguity, Subjectivity, Context�

Impact of greater reliance on AI

AI may reinforce existing social biases; �racial, gender, sexual orientation.
Prohibitive Adverse Implications

the growing reliance on AI for content moderation—especially after the pandemic. When COVID-19 hit, social media platforms had to send thousands of human moderators home, leaving content moderation largely in the hands of AI models.

Now, AI has been a main focus for managing the abundance of big social data. These systems can scan billions of posts in real time to detect harmful content like misinformation, hate speech, or harassment, which is critical for keeping platforms safe and functional.

But, of course there are challenges:

Bias and toxicity often seep into these models.

They still struggle with ambiguity, subjectivity, and context, which are essential for distinguishing harmful content from meaningful discussions.

So, during the COVID-19 pandemic, social media wasn’t just a place for casual conversation, as it became a critical medium for information. People used these platforms to share experiences, symptoms, and concerns, while others sought support for their mental health. However, this flood of data also included harmful content that worsened public anxiety.

For example, during COVID-19, unmoderated misinformation spread panic, while posts about real struggles with mental health highlighted an urgent need for intervention. So the balance between filtering out harmful content and identifying meaningful signals, is where better AI systems, like neuro-symbolic AI, can make a real difference.

14 of 83

14

5% of all Google searches are health-related.

Source: https://googleblog.blogspot.com/2015/02/health-info-knowledge-graph.html

Healthcare data will experience a compound annual growth rate (CAGR) of 36% through 2025.

Source: https://healthitanalytics.com/news/big-data-to-see-explosive-growth-challenging-healthcare-organizations

FDA Sets Goals for Big Data, Clinical Trials, Artificial Intelligence.

Source: https://healthitanalytics.com/news/fda-sets-goals-for-big-data-clinical-trials-artificial-intelligence

When we take a step back and think about the role of big data and AI in healthcare and public health, we can see that it has been increasingly more critical.

For instance, 5% of all Google searches are health-related, meaning millions of people using Google every day to look up symptoms, treatments, or general health advice. This shows just how important online platforms are for understanding people’s health concerns and behavior in real time.

At the same time, healthcare data is growing at a very high rate, a massive amount of data coming from electronic health records, wearable devices, telemedicine platforms, and, of course, social media. This growth reflects not only the increasing reliance on technology but also how critical data is for making informed decisions about public health and patient care.

And to manage this growth, the FDA is stepping in to set goals for using big data, clinical trials, and AI. They recognize the potential of AI to process and analyze large, complex datasets.

By combining social media analysis with AI-driven approaches, we can better understand what’s happening in public health, raise awareness about mental health challenges, and support decision-makers in allocating resources where they’re needed most.

15 of 83

15

Source: https://healthitanalytics.com/news/how-artificial-intelligence-is-changing-radiology-pathology

16 of 83

Information is cheap. Understanding is expensive.

Karl Fast,

Professor of UX Design,�Kent State University

16

AI is about converting data into knowledge, insights and actions.

“

We live in an era where there’s no shortage of data, thanks to internet and social media.

But the main point here is that the data by itself is just information. The real challenge is how we can transform this data into meaningful knowledge and actionable insights, and that’s where AI comes in.

Traditional AI models, like machine learning or deep learning systems, are great at processing large amounts of data, but they often lack contextual understanding, suffering from ambiguity.

neurosymbolic AI can address this gap, which combines neural networks (learning from data) with symbolic reasoning (which brings structured, knowledge-based understanding). By integrating domain-specific knowledge graphs, we give AI systems access to structured information that helps them reason and interpret context more effectively.

17 of 83

Challenges with Current LLMs

17

18 of 83

Explainability for People, not just Designers and Developers

18

LLAMA

NeuroSymbolic AI

Domain Knowledge: PHQ 9

LLAMA + Domain Knowledge Output

Dalal, S., Tilwani, D., Gaur, M., Jain, S., Shalin, V., & Seth, A. (2023). A Cross Attention Approach to Diagnostic Explainability using Clinical Practice Guidelines for Depression. arXiv preprint arXiv:2311.13852.

19 of 83

Knowledge-Verified Prediction via Linking to KGs

19

Really struggling with my bisexuality which is causing chaos in my relationship with a girl. I am equal to worthless for her. I’m now starting to get drunk because I can’t cope with the obsessive, intrusive thoughts, and need to get out of my head.

288291000119102: High risk bisexual behavior

365949003: Health-related behavior finding

307077003: Feeling hopeless

365107007: level of mood

225445003: Intrusive thoughts

55956009: Disturbance in content of thought

26628009: Disturbance in thinking

1376001: Obsessive compulsive personality disorder

Multi-hop traversal on medical knowledge graphs

Obsessive-compulsive disorder is a disorder in which people have obsessive, intrusive thoughts, ideas or sensations that make them feel driven to do something repetitively

Gaur, M., Desai, A., Faldu, K., & Sheth, A. (2020). Explainable ai using knowledge graphs. In ACM CoDS-COMAD Conference. Link, slide.

Rawte, V., Chakraborty, M., Roy, K., Gaur, M., Faldu, K., Kikani, P., ... & Sheth, A. P. TDLR: Top Semantic-Down Syntactic Language Representation. In NeurIPS'22 Workshop on All Things Attention: Bridging Different Perspectives on Attention., link

20 of 83

Knowledge-Verified Prediction via Process KGs Structures

20

Process Knowledge Structure in C-SSRS

C-SSRS: Columbia Suicide Severity Rating Scale

I wish I could give a shit about what would make it to the front page. I have been there and got nothing. Same as my life. I do have a gun.’, ’I thought I was talking about it. I am not on a ledge or something, but I do have my gun in my lap.’, ’No. I made sure she got an education and she knows how to get a job. I also have recently bought her clothes to make her more attractive. She has told me she only loves me because I buy her things.

1. Wish to be dead - Yes

2. Non-specific Active Suicidal Thoughts - Yes

3. Active Suicidal Ideation with Some Intent to Act - Yes

4. Label: Suicide Behavior or Attempt

Interpretable for System Users i.e., Clinicians and Patients

(1,2,3 verify adherence to the clinical guideline on diagnosis which a clinician understands)

47%

70%

LLMs

Process Knowledge (Ours)

Agreement with Experts

Sheth, A., Gaur, M., Roy, K., Venkataraman, R., & Khandelwal, V. (2022). Process Knowledge-Infused AI: Toward User-Level Explainability, Interpretability, and Safety. IEEE Internet Computing, 26(5), 76-84., link

21 of 83

Generative AI has Significant Potential for Harm!

21

Article Link

22 of 83

Recent Case of Character.ai

22

https://apnews.com/article/chatbot-ai-lawsuit-suicide-teen-artificial-intelligence-9d48adc572100822fdbc3c90d1456bd0

23 of 83

23

Article Link

24 of 83

How Current Language Models Work

24

What is Mark Zuckerberg’s net worth?

Did you mean: net worth

Did you mean: salary

Did you mean: rich for

net worth:

0.00567%

Image Source

net worth

salary

rich for

Language Models Predict based on Context-Specific Distributional Mappings

Prediction Context

25 of 83

25

26 of 83

26

Longer list of Failures …

LLM

Limited accuracy in complex decision-support requests

ChatGPT showed only 56% accuracy in medical queries (Wei et al., 2023), raising concerns about trustworthiness in clinical use.

Lack of domain-specific expertise

General-purpose LLMs struggle with specialized medical knowledge, leading to errors in diagnosis and treatment recommendations (Szymanski et al., 2024).

Inability to handle & Follow guidelines

LLMs often rely on outdated or incomplete information, failing to incorporate the latest medical research or evolving clinical guidelines (Sheth et al., 2024).

Potential for generating harmful or biased content

LLMs can provide inaccurate or harmful suggestions, particularly if the input data is biased or not representative of diverse patient populations (Gupta et al., 2023).

For Decision-Support Assistance

? Data - Why train on Voluminous open web data?

? Knowledge

Representing Domain-Specific Information
Representing Relevant Facts about the World
Representing Domain-relevant Decision Processes, incl. societal/professional value (laws, rules, guidelines, protocols)

? Human Expertise

- How to Ensure Knowledge and data are Leveraged correctly?

27 of 83

Role of Knowledge in understanding content and deeper analysis through Neurosymbolic AI

27

28 of 83

28

Symbolic AI Statistical AI Neuro-symbolic AI

Where are we in AI Evolution now?

29 of 83

29

Knowledge Graph (Labeled Nodes and Edges)

NeuroSymbolic Reasoning

System 2

Neural Network and Deep Learning

Decisions/Actions

System 1

Low-level Data

Sensors, Text, Image, and Collection

Symbolic Explicit Knowledge Representation

Neural Implicit/Parametric Knowledge Representations

Expert Human

Amit Sheth, Kaushik Roy, Manas Gaur, Neurosymbolic Artificial Intelligence (Why, What, and How), IEEE Intelligent Systems, 38 (3), May-June 2023

NeuroSymbolic AI

30 of 83

30

Knowledge and

Experience

System 1:

Perception:

DL/Neural AI

System 2:

Cognition:

Symbolic AI

Data to Concepts,

Abstractions, Understanding

NeuroSymbolic: System 1 (Neuro) + System 2 (Symbolic)

31 of 83

31

Knowledge and

Experience

System 1:

Perception:

DL/Neural AI

System 2:

Cognition:

Symbolic AI

Data to Concepts,

Abstractions, Understanding

Natural Language (NL)-Processing (P) to NL-U (Understanding)

32 of 83

32

Neural Network

Abstract / Contextualization

ACT

DECIDE

reasoning

Planning

Inference

Apply Process Knowledge: User has Specific concerns due to X, Y, Z Concepts

Action:

Further Interact with System User on their concerns

Explicit Knowledge

Data

Contextualization

is at the heart of

understanding

Natural Language (NL)-Processing (P) to NL-U (Understanding)

33 of 83

33

From NLP to NLU: Deeper understanding of content

34 of 83

34

Neurosymbolic Customized and Compact (NeSy-CC) Copilots

A Granular Look at The Features of a NeSy-CC Systems

#Grounding #Instructability

#Alignment �#Explainability #Intrepretability

#Analogy�#Reliability

#Consistency

#Planning

#Reasoning

35 of 83

35

Shallow Infusion

Sheth, Gaur, Kursuncu, & Wickramarachchi, (2019). Shades of knowledge-infused learning for enhancing deep learning. IEEE Internet Computing, 23(6), 54-63., link

Machine Learning Model

How well the model learned the task?

Shapley plots on Feature Importances or Dependencies

PT_t = t^th topic/phrase extracted from free form input text

KS_c = c^th concept in a knowledge source ( graph, base, ontology, and/or lexicon

Mapping

36 of 83

36

Semi-Deep Infusion

Sheth, Gaur, Kursuncu, & Wickramarachchi, (2019). Shades of knowledge-infused learning for enhancing deep learning. IEEE Internet Computing, 23(6), 54-63., link

Machine Learning Model

37 of 83

37

Deep Infusion

Sheth, Gaur, Kursuncu, & Wickramarachchi, (2019). Shades of knowledge-infused learning for enhancing deep learning. IEEE Internet Computing, 23(6), 54-63., link

Backpropagation

Connector acting like a toggle switch

38 of 83

38

Content

Foundations

Why Social Media Data
LLMs and its challenges
NeuroSymbolic Approach
Types of Knowledge Infused Learning and its advantages

Neurosymbolic Approach for an health application

Use-Case COVID-19

Hands-on Session

39 of 83

COVID-19 Use Case

39

40 of 83

As of December 14th, 2024

~7M deaths and >704M confirmed cases globally
~1.2M death with >111M confirmed cases in the US
Massive, once in a century societal impact on health, economy & well-being

40

Massive impact of pandemic on health and society

Photo: The European Society for Medical Oncology

41 of 83

Impact on mental Health

Mental Health: Depression, Anxiety
Addiction: Substance use/abuse

41

Massive impact of pandemic on health and society

Photo: unsplash.com

Source: https://www.statista.com/statistics/1241055/us-adults-mental-health-changes-covid-vs-last-ten-years-by-gender/

42 of 83

Social media reveals impact

Photo: American Psychological Association

"All the things are being shut down by #Covid19 but my anxiety & depression 🙁"

"A feeling of hopelessness. Seems I am in a dark age. #coronavirus #COVID19"

“I drive the streets of #LA looking 4 my #Homeless kids,drug & alcohol #addicted Often, I find them emaciated & delusional.”

“i blame my parents for manipulating me into thinkin i’m nothing without them and i blame myself for believing it >:| #abusiveparents”

Mental Health

Addiction

43 of 83

43

Twitter Data: 12 billion tweets analyzed, capturing public sentiment and mental health signals during the COVID-19 pandemic.
Subreddit Data: 2.5 million subreddit posts offering deep community-based insights on mental health topics like depression and anxiety.
News Articles: 700,000 COVID-19-related articles providing a broader societal and policy context.
Knowledge Graphs: A combination of domain-specific resources like DSM-5 and Drug Abuse Ontology (DAO) and general-purpose graphs like DBpedia and Wikidata.
Neologisms: Captured emerging terms such as “Zoom fatigue” and “coronasomnia” from social media, enabling real-time adaptation to evolving language trends.

The Massive Social Media Corpus

44 of 83

44

Technical Approach Overview

Vedant Khandelwal, Manas Gaur, Ugur Kursuncu, Valerie Shalin, and Amit Sheth. "A Domain-Agnostic Neurosymbolic Approach for Big Social Data Analysis: Evaluating Mental Health Sentiment on Social Media during COVID-19." In Proceedings of the IEEE International Conference on Big Data, 2024

45 of 83

Technical Approach Overview

45

46 of 83

Domain-Specific Topic and Language Modelling

46

Topics describing each subreddits are identified through:

Skip Gram model to generate n-grams
LDA over subreddits
LDA over bigrams of subreddits

Relevant topics were identified constraining through Topic Coherence measure.
We utilize UCI topic coherence model which is Pointwise Mutual Information.

Sub-reddit language model is trained through:

Skip Gram model to generate n-grams
Word2Vec over ngrams of subreddits

47 of 83

Some of the Topics Identified after LDA

47

Anxiety	Depression, Cognitive distortions, panic attacks, hopelessness, physical sensations.
Depression	Mood swings, weight gain, rapid cycling, depressive episode, Impulsivity, mood swings, antisocial conduct, personality disorder
Addiction	Buying oxycodone, pain management, chronic pain, alienation, crippling alcohol, dependent on crack

48 of 83

DSM-5: Background

48

2013, 5th Edition Diagnostic and Statistical Manual of Mental Disorders (DSM-5) is a psychiatric bible that can cure 46.4% of adult US population suffering from Mental Illness.

There are 21 Diagnostic categories of which 20 are specific to Mental Health

49 of 83

DSM-5 Catalog

49

Neurodevelopmental Disorders

Schizophrenia Spectrum

Psychotic Disorders

Bipolar and Related Disorders

Depressive Disorders

Anxiety Disorders

Obsessive-Compulsive and Related Disorders

Trauma- and Stressor-Related Disorders

Dissociative Disorders

Sleep-Wake Disorders

Feeding and Eating Disorders

Elimination Disorders

Suicidal Behavior/Ideation Disorders

Sexual Dysfunctions

Gender Dysphoria

Disruptive Impulse Control and Conduct Disorders

Substance Use and Addictive Disorders

Neurocognitive Disorders

Personality Disorders

Paraphilic Disorders

50 of 83

DAO: Drug Abuse Ontology

50

Conceptual framework interconnecting sets of Drug-focused and Health-related concepts.

The advantage of DAO is that it is not limited to medical terminology, but also includes commonly used lay and slang terms for mental health conditions and associated symptoms.

Concept	315
Relations	31
Instances	814

Drug Abuse Ontology

Lokala, Usha, Raminta Daniulaityte, Francois Lamy, Manas Gaur, Krishnaprasad Thirunarayan, Ugur Kursuncu, and Amit P. Sheth. "DAO: An Ontology for Substance Use Epidemiology on Social Media and Dark Web." JMIR Public Health and Surveillance (2020).

51 of 83

Content Enrichment

Mental Health - Drug Abuse (MHDA) Knowledge Base :

It is obtained by aggregating mental health and drug abuse related entities from PHQ-9, SNOMED-CT, DSM-5, DAO, MeSH Terms.�

Entities:

Entities are extracted from data sources.
Candidate entities are filtered using Knowledge bases
Further filtered set of entities are used to enrich lexicon categories.

51

Candidate Entities

Enriched Lexicon

Domain Knowledge

DAO

DSM-5

52 of 83

Neologisms

52

The system captures emerging terms like "coronapocalypse" and "Zoom fatigue," reflecting shifts in public discourse during key COVID-19 milestones.
These neologisms, derived from semantic filtering, enhance contextual understanding, ensuring the model remains relevant to evolving societal language trends.

53 of 83

Content Enrichment

Semantic Filtering

Removal of irrelevant noisy data
Finding mapping between Lexicons and Tweets Phrases

Location Extraction

Obtained data about US states, county, city and alias information from OpenStreetMap, data.gov.us and Geonames Ontology.
Filter tweets based on the location metadata

53

54 of 83

Technical Approach Overview

54

55 of 83

55

Semantic Proximity: alignment with MHDA-Kb.

Removal of ambiguity.
Example: “palpitations and social anxiety are killing me” → “anxiety is killing me”

Semantic Mapping:

Every comment or post is pre-categorized into one of various subreddit on Reddit.
We trained topic model for each Mental Health-related subreddit and generate compound topics.
We define semantic mapping as a procedure to Match compound topics from sub-reddit to those obtained from tweets

Hit Score Calculation

56 of 83

Hit Score Calculation

56

Medical Knowledge Bases

LDA

LDA over Bi-grams

Hit

Score

DSM-5

Lexicon

DAO

Drug Abuse Ontology

*Gaur, Manas, et al. "" Let Me Tell You About Your Mental Health!" Contextualized Classification of Reddit Posts to DSM-5 for Web-based Intervention." Proceedings of the 27th ACM International Conference on Information and Knowledge Management. 2018.

N-Gram Key Phrases

57 of 83

Hit Score Calculation

57

S: Index of a particular tweet

D: Concepts extracted from the lexicons of different category {Depression, Addiction, Anxiety, HealthCare, Financial, StayAtHome}

H^S: Collection of Hit Score calculated for S

ng^S: Ngrams extracted from S

LDA^S: Compound topics extracted

bLDA^S: Compound Ngram Topics extracted

H(a,b): Number of hits of a that maps with hits in b.

nhs^S_D: Index Score

*Gaur, Manas, et al. "" Let Me Tell You About Your Mental Health!" Contextualized Classification of Reddit Posts to DSM-5 for Web-based Intervention." Proceedings of the 27th ACM International Conference on Information and Knowledge Management. 2018.

Semantic Proximity

Semantic Mapping

57

58 of 83

Tweet Examples From Dataset

Anxiety: “All things are being shut down by #COVID19 but my anxiety and depression🙁”

Depression: “A feeling of hopelessness. Seems I am in a dark age. #SARSCOV-2”

Addiction: “I drive the streets of #LA looking 4 my #Homeless kids, drug and alcohol Often, I find them emaciated and delusional.”

HealthCare: “Meanwhile: NHS staff to be asked to treat coronavirus patients without gowns #novelcorona”

StayAtHome: “The stress, uncertainty and isolation of #COVID19 can be even more frightening for people in abusive relationships. #DomesticViolence #COVID19 #stayhome”

Financial: “The <username> has three new credits to help your business through these rough times, including immediate assistance to keep your employees in your payroll. #COVIDreliefUT #business #SmallBusiness #COVID19 #COVID”, “Our child care system is on the verge of collapsing beneath the economic burden of this pandemic. If we don't act, millions of parents will be unable to return to work and our economic recovery will suffer. <username> and I have a plan to fix it—before it's too late. #COVID #creditfreeze #relieffund”

58

59 of 83

Technical Approach Overview

59

60 of 83

Previous Work: Architecture

60

SEDO

Semantic Encoding and Decoding Optimization. It is a procedure to modulate word embedding (vectors) of a word.

Reddit with

DSM-5 labels

Word Embedding Model

Correlation Matrix (Q)over word vectors

Medical Knowledge Bases

Domain

Experts

Correlation Matrix (P)

over DSM-5 Lexicon or DAO

SEDO

Optimize P, Q & Z

DSM-5 Lexicon

DSM-5 Vocabulary Matrix

Word-modulated Word Embeddings

DSM-5 Classification

Cross Correlation Matrix (Z)

between word vectors and DSM-5 Lexicon or DAO

HLF+VLF+FGF

Feature set

DAO

*Gaur, Manas, et al. "" Let Me Tell You About Your Mental Health!" Contextualized Classification of Reddit Posts to DSM-5 for Web-based Intervention." Proceedings of the 27th ACM International Conference on Information and Knowledge Management. 2018.

61 of 83

Semantic Encoding and Decoding Optimization (SEDO)

61

We have incorporate background knowledge in DSM-5-DAO to classification process utilizing SEDO.

We introduce SEDO as an approach for obtaining a discriminative weight matrix between the DSM-5 lexicon and Reddit embedding space

SEDO modulates the embeddings of each word in the Reddit content of the user based on proximity of the word to DSM-5 category.

Correlation Matrix (Q)over word vectors

Correlation Matrix (P)

over DSM-5 Lexicon or DAO

Cross Correlation Matrix (Z)

between word vectors and DSM-5 Lexicon or DAO

SEDO

Optimize P, Q & Z

*Gaur, Manas, et al. "" Let Me Tell You About Your Mental Health!" Contextualized Classification of Reddit Posts to DSM-5 for Web-based Intervention." Proceedings of the 27th ACM International Conference on Information and Knowledge Management. 2018.

62 of 83

Semantic Encoding and Decoding Optimization

62

12808 Words

300 dimension embedding

20 DSM-5 Categories

R

Reddit Word Embedding Model

DSM-5 -DAO Lexicon

W

Solvable Sylvester Equation

Here, as we can see on the left hand side we have correlation matrix of words, on the right hand side we have correlation matrix of DSM-5 category, further we try to learn the weight matrix W, by utilizing the sylvester equation, which has been used in computer visitation within the context of Zero shot learning. Here as we can see we are utlizing the three correlation matrix in the equation as, Correlation between DSM-5 category multiplied by W plus, W multiplied to correlation of words which equals to (one plus delta, which is a regularisation parameter) multiplied to cross correlation between category and words. Finally, we received a weight matrix of the size 12808*20. Such that, for every word we have a weight vector of length 20 is associated, where each dimension of vector corresponds to one DSM-5 category.

63 of 83

Model Training for Covid-19

63

SEDO

Semantic Encoding and Decoding Optimization. It is a procedure to modulate word embedding (vectors) of a word.

Tweets Ngrams mapped to MHDA Lexicon

Word Embedding Model

Correlation Matrix (Q)over Tweet word vectors

Correlation Matrix (P) over MHDA Lexicon

SEDO

Optimize P, Q & Z

MHDA Vocabulary Matrix

Word-modulated Word Embeddings

Tweet Classification

Cross Correlation Matrix (Z)

between Tweet word vectors and MHDA Lexicon

Modulated Tweet Embedding

*Gaur, Manas, et al. "" Let Me Tell You About Your Mental Health!" Contextualized Classification of Reddit Posts to DSM-5 for Web-based Intervention." Proceedings of the 27th ACM International Conference on Information and Knowledge Management. 2018.

64 of 83

Experimental Setup

Purpose: Validate the neurosymbolic approach for dynamic sentiment analysis in mental health discourse during the COVID-19 pandemic.

Experiments Conducted:

Baseline Classification: Evaluate binary classification for Depression, Addiction, and Anxiety categories.
Triangulation Study: Validate generalizability using external datasets.

Analyze empirical significance of pre-trained vs. fine-tuned SEDO weight matrices.

Comparison with LLMs: Evaluate performance against state-of-the-art language models like LLama, Phi, and Mistral.
Focus: Precision, recall, F1 score, and computational efficiency.

64

65 of 83

Results - Baseline Classification

The table compares precision, recall, and F1 scores for four models across categories. Values in red parentheses indicate the percentage performance drop when the SEDO matrix is excluded.
The Neurosymbolic Balanced Sub-sample Random Forest (BSRF) consistently achieves the highest F1 scores across all categories, underscoring its effectiveness in handling imbalanced data.
The significant decrease in performance metrics (precision, recall, and F1-score) without the SEDO matrix (-21% to -29%) demonstrates the critical role of SEDO optimization in improving classification accuracy.

65

66 of 83

Results - Triangulation Study

This study assesses the generalizability and robustness of the proposed SEDO framework by applying it to previously unseen datasets annotated for depression, addiction, and anxiety.
The datasets consist of social media posts collected from published datasets specific to each category—depression, addiction, and anxiety.
One with pre-trained SEDO weight matrix and the other with a fine-tuned matrix for improved classification.
Fine-tuning the SEDO matrix on new datasets led to significant performance improvements, demonstrating.

66

67 of 83

Results - Comparison with LLMs

The neurosymbolic model was compared with three state-of-the-art LLMs: LLama (7B parameters), Phi (2.7B parameters), and Mistral (7B parameters). These LLMs were used in an open-source, instruct-tuned, zero-shot setting.
The evaluation dataset included 1,000 tweets per category (depression, addiction, and anxiety), collected across three timeframes—April-May 2020, August-September 2020, and December 2020-January 2021.
The neurosymbolic approach outperformed the LLMs in all metrics (precision, recall, F1-score) for all three categories, achieving F1-scores between 88.84% and 91.85%, compared to 68.95%-78.56% for LLMs. This highlights its adaptability and computational efficiency over traditional LLMs.

67

68 of 83

68

A calculated Social Quality Index (SQI) aggregates mental health components (Depression, Anxiety), Addiction and Substance Use Disorders.

Social Quality Index (SQI)

vecteezy.com

Change in SQI informs comparisons between states.
Raw transformed SQI into relative state rankings changing over time.

We defined an empirical Social Quality Index called SQI that aggregates over different MHDA categories. Our analysis focuses on relative SQI between states and how is it changing over time.

We developed this empirical index to provide a clear, measurable way to assess social and mental health conditions across different regions.

The SQI aggregates four critical components:

Depression

Anxiety

Addiction

Substance Use Disorders

By bringing these factors together, we can create a single, comprehensive measure that reflects the overall mental health and well-being of a population.

We focus on two key aspects:

Relative SQI between states: This allows us to compare how states rank in terms of their mental health and substance use challenges.

How SQI is changing over time: Tracking this helps us understand trends—are some states improving, while others are getting worse? And what factors might be driving these changes?

This is important because the SQI gives us a data-driven tool to analyze how mental health and substance use disorders evolve across regions.

SQI helps us move from raw data to actionable insights. It’s not just about knowing that mental health challenges exist, how they’re changing, and what we can do to address them.

69 of 83

69

e.g., IN, NH, OH, OR, WA, WY �are worsening.

Results: Relative State Rankings Reveal Patterns

SQI Ranking April 4 - 10

SQI Ranking March 14 - 20

SQI Ranking March 21 - 27

SQI Ranking March 23-April 3

Darker: Better Social Quality

We have grouped several states together, to obtain clusters of states with similar pattern of change in SQI, for a four week time period.Here as we can see for the cluster of states consisting of Indiana, New Hempshire, Ohio, Oregon, Washington, Wyoming are worsening over the four period timeline.

—---------

Here, we’re looking at the relative changes in SQI rankings across states over a four-week period. To make sense of the trends, we grouped states into clusters with similar patterns of change.

One clear example is the cluster of states including Indiana (IN), New Hampshire (NH), Ohio (OH), Oregon (OR), Washington (WA), and Wyoming (WY). As you can see across these maps, these states show a consistent worsening trend, lighter colors over time indicate a drop in their SQI rankings, reflecting a decline in social quality.

The darker the green, the better the social quality and in these states, that darker shade is fading week by week. This pattern suggests growing challenges in areas like mental health and substance use.

Visualizing these changes over time helps us pinpoint regions that need more attention and intervention.

70 of 83

70

�IL, NY, MD, AZ, NM, MA

WI, RI, NV, NJ, CT, LA, OK

WA, KS, IN, WY, OH, OR, NH

Relative SQI Ranking

Results: Three of the Observed Temporal Patterns

March 14-20 March 21-27 March 28-April 3 April 4-10

Based on pattern of change in SQI, we have put different group of states in clusters. The states such as Wisconsin, Rhode Island, Nevada, New Jersey, Connecticut, Los Angeles, Oklahoma shows a non-linear pattern. Where as, states Kansas, Indiana, Washington, Wyoming, Ohio, Oregon and New Hempshire showing a linear decline in the SQI, and Intresetingly states such as illinois, New York, Maryland, Arizona, Massachusetts show a linear improvement in SQI. In Further slides, we will be discussing about the different factors affecting these change in SQI.

—----------

Here, we’re looking at some really interesting temporal patterns in the SQI rankings over a four-week period during the early days of the pandemic. Based on these patterns, we’ve clustered states into distinct groups based on how their SQI is changing over time.

First, let’s look at the non-linear group—states like Wisconsin, Rhode Island, Nevada, New Jersey, Connecticut, Louisiana, and Oklahoma. You can see this blue line on the graph, rising sharply, stabilizes for a bit, and then starts to decline. These states show fluctuating trends in their SQI, indicating a complex and dynamic pattern.

Next, we have states showing a linear decline in their SQI, represented in green. These states include Kansas, Indiana, Washington, Wyoming, Ohio, Oregon, and New Hampshire. Over time, these states consistently move downward, suggesting worsening mental health and substance use trends. This could point to ongoing challenges, like economic stress or lingering effects from the pandemic.

On the other hand, we see states showing a linear improvement in their SQI, highlighted by the orange line. States like Illinois, New York, Maryland, Arizona, and Massachusetts are in this cluster. These states steadily move upward, indicating an improvement in their relative mental health and substance use outcomes.

What’s really interesting here is that these patterns, whether non-linear, improving, or declining, reveal that the SQI isn’t static. Different regions are experiencing shifts at varying rates, and understanding these trends is key.

In the next few slides, we’ll look deeper into the factors potentially driving these changes, whether it’s socio-economic conditions, public health interventions, or specific events affecting mental health and substance use in these states. By analyzing these patterns, we can get a clearer picture of where interventions are working and where more support is needed.

71 of 83

Results: Cluster --Improving SQI Ranking

71

SQI bad SQI better SQI better SQI better

Frequency

Depression: 125037

Addiction: 92897

Anxiety: 81891

Total: 299825

Frequency

Depression: 113830

Addiction: 81810

Anxiety: 74080

Total: 269720

Frequency

Depression: 81463

Addiction: 60166

Anxiety: 45998

Total: 187627

Frequency

Depression: 59088

Addiction: 49086

Anxiety: 46887

Total: 155061

IL, NY, MD, AZ, NM, MA.

March 14-20 March 21-27 March 28-April 3 April 4-10

For a cluster with Improving SQI ranking, we see that the volume of tweets with MHDA concepts, have been decreasing, ie, from about 300k in first week to 155k in the last week.

—--------

Here, we’re looking at a cluster of states, Illinois, New York, Maryland, Arizona, New Mexico, and Massachusetts, showing an improving SQI ranking over time.

What’s particularly interesting is the volume of tweets mentioning mental health and substance use concepts (MHDA). You can see that this volume is decreasing week by week:

In the first week (March 14–20), we had nearly 300,000 mentions.

By the final week (April 4–10), this number drops to 155,000 mentions.

This decline in MHDA-related discussions coincides with the improvement in the SQI ranking for this cluster. While there are still significant mentions of concepts like 'feel down,' 'social anxiety,' and 'fear of social situations,' the overall frequency is decreasing.

This pattern suggests a potential improvement in mental health indicators or a reduction in the intensity of public discussions around these issues in these states.

72 of 83

72

Results: Cluster --Declining SQI Ranking

March 14-20 March 21-27 March 28-April 3 April 4-10

SQI good SQI worse SQI worse SQI worse

WA, KS, IN, WY, OH, OR, NH

Frequency

Depression: 88491

Addiction: 24373

Anxiety: 37725

Total: 146589

Frequency

Depression: 68491

Addiction: 37846

Anxiety: 53189

Total: 159526

Frequency

Depression: 81746

Addiction: 59756

Anxiety: 78885

Total: 220387

Frequency

Depression: 123244

Addiction: 84879

Anxiety: 94999

Total: 303122

For a cluster with declining SQI, we see that the volume of tweets with MHDA concepts, have been increasing, ie, it goes from 146k in the first week to 303k for the last week.

—---------

Here, we’re looking at a cluster of states, Washington, Kansas, Indiana, Wyoming, Ohio, Oregon, and New Hampshire, showing a declining SQI ranking over time.

What stands out here, is the volume of tweets discussing mental health and substance use concepts (MHDA). Unlike the previous improving cluster, this volume is increasing significantly:

In the first week (March 14–20), we see about 146,000 mentions.

By the final week (April 4–10), this jumps to over 303,000 mentions—more than double.

The most frequently observed concepts include 'social anxiety,' 'feel down,' 'fear of social situations,' and 'addicted to meth.' This steady rise in mentions aligns with the worsening SQI for this cluster, suggesting an increasing burden of mental health and substance use issues.

These findings highlight the importance of monitoring public discourse in real time. The growing discussion signals that communities in these states may need additional interventions or support to address these challenges.

73 of 83

Results: Cluster --A Non-Linear SQI Ranking

73

WI, RI, NV, NJ, CT, LA, OK.

SQI worse SQI better SQI better SQI worse

Frequency

Depression: 91,480

Addiction: 103549

Anxiety: 88293

Total: 283322

Frequency

Depression: 62825

Addiction: 81400

Anxiety: 54184

Total: 198409

Frequency

Depression: 58223

Addiction: 76232

Anxiety: 41484

Total: 175949

Frequency

Depression: 78061

Addiction: 87463

Anxiety: 63865

Total: 229389

March 14-20 March 21-27 March 28-April 3 April 4-10

74 of 83

Explanation: Two threads of influence

74

External events

(business and school closing)

Short term Human Coping Processes (content changes in focus of attention)

SQI

75 of 83

Results: Influence of External Events

75

SQI worse

Cluster 4:

CT, LA, NJ, NV, OK, RI, WI.

School Closures: CT, LA, NJ, NV, RI, WV, WI

Business Closures: CT, LA, NJ, RI, WV, WI

Social Distancing Reg: LA, NJ, RI, WV, WI

Business Relief: WI

Unemployment increase:

CT 2.5K %, LA 2.5K %, NJ 1.2K %,

NV 1.2K %, OK 1.2K %, RI 2.5K %, WI 1.2K %.

Stay at home: CT, LA, NJ, OK, RI, WI, WV

Extension School: CT, WV

Major Disaster: NJ

Business Relief: NJ

Unemployment increase:

CT 180%, LA 0 %, NJ 64 %,

NV 0 %, OK 99 %, RI -23%, WI 99 %.

Major Disaster: CT, WV

Strict Social Dist: CT, RI

Extensions deadlines: CT

Medical shortage: NJ

Extension Stay home: OK

Extension School: RI

Extension Business Closure: RI

Business Relief: NJ, RI

Individual Relief: RI

Unemployment increase:

CT 0%, LA 5 %, NJ 3 %,

NV 11 %, OK 7 %, RI 0%, WI -5 %.

Extension School: CT

Extension Stay home: LA

Strict Social Dist: NJ

Business Relief: WI

Cluster 5:

FL, GA, MI, NE, TN, VA, WV.

School Closures: FL, GA, MI, TN, VA, WV,

Business Closures: WV, MI

Social Distancing Reg: FL, MI, NE, TN, VA, WV,

Business Relief: FL, GA, MI, NE, TN, VA

Individual Relief: TN, VA

Unemployment increase:

FL 600%, GA 650%, MI 180%,

NE 70%, TN 180%, VA 180%,

WV 600%

Stay at home: MI, WV

Shelter in Place: GA

Business Closure: GA, TN

Extension School: GA, WV

Major Disaster: FL

Business Relief: TN

Individual Relief: TN

Unemployment increase:

FL 3.1K%, GA 3K%, MI 1.8K%,

NE 200%, TN 700%, VA 1.6K%,

WV 1.7K%

Stay at home: FL, VA

Shelter in Place: TN

Major Disaster: GA, MI, TN, VA, WV

Strict Social Dist: GA

Extension School: GA, MI

Unemployment increase:

FL -25%, GA 190%, MI 27%,

NE 8%, TN 26%, VA 33%,

WV 0%

Extension School: GA

Extension Stay home: MI

SQI worse

SQI better

March 14-20 March 21-27 March 28-April 3 April 4-10

Among many external factors, the events that are finance related appear to impact SQI. Specifically, Business and individual relief announcements, business closures, stay at home, and increase in unemployment, have apparent effects, which are illustrated here with two state clusters that have U-shaped relative SQI patterns. The states of Florida, GerogiaA, Michigan, Nevada, Tennessee, Virginia, West Virginia, delayed the closure of businesses and the announcements of stay at home and shelter in place orders . Accordingly, weekly increases in their unemployment rates were also delayed. We observe that whenever the individuals and businesses are given financial reliefs, the SQI is better; whereas, whenever the unemployment increase is much more significant than the previous week, the SQI is worse. Further, a week announcement of stay at home or shelter in place order is usually followed by a week with better SQI. These findings implies that financial factors (e.g., unemployment and relief packages) have mainly the most effect in the social quality of people, and the specific government interventions have significant impact.

—---------------

Here, we’re looking at the influence of external events on SQI, focusing on two state clusters: Cluster 4 (CT, LA, NJ, NV, OK, RI, WI) and Cluster 5 (FL, GA, MI, NE, TN, VA, WV).

What stands out here is that financial events, like business relief announcements, business closures, unemployment increases, and stay-at-home orders—have a clear impact on SQI.

For example:

In states like Florida, Georgia, Michigan, Tennessee, and Virginia, delayed closures and late announcements of stay-at-home orders also delayed the weekly rise in unemployment rates.

We observe a U-shaped pattern in SQI:

When financial relief is provided to individuals or businesses, the SQI improves (in green).

On the other hand, when there’s a significant spike in unemployment, SQI tends to worsen.

Interestingly, a stay-at-home or shelter-in-place order is often followed by a week with better SQI, suggesting a stabilizing effect on social quality.

The overall key takeaway here is that financial factors, such as relief packages and unemployment trends, have the most noticeable influence on social quality.

This highlights how specific government interventions can play a critical role in shaping public well-being during crises.

76 of 83

Hashtag Content Mirrors SQI

(steadily improving states)

76

SQI:

SQI bad SQI better SQI better SQI better

Hashtag:

#

Cluster 7:

IL, NY, MD, AZ, NM, MA.

March 14-20 March 21-27 March 28-April 3 April 4-10

In this cluster of states, including Illinois, New York, Maryland, Arizona, New Mexico, and Massachusetts; we see how hashtag content on social media reflects changes in SQI over time.

In Week 1, when the SQI was relatively bad, hashtags like #trumppandemic cast blame on leadership, showing public anger, while another hashtag #kag2020 (Keep America Great) emerges as a defensive reply. We also see hashtags like #coronaapocalypse, reflecting public panic and recognition of the pandemic's potentially severe implications.

However, as the weeks progress and SQI improves, we observe a shift in focus toward adaptive and coping responses to the crisis. For example:

#hydroxychloroquine appears as a hopeful treatment, highlighting early attempts to combat the virus.

Hashtags like #quarantinelife and #lockdown normalize new behaviors, such as staying at home, which helps rationalize public anxiety or agoraphobia.

Finally, hashtags like #lightitblue emerge to motivate and support health care workers, emphasizing solidarity and hope as communities adapt to the new reality.

This shift in tone, moving from panic and blame to adaptation and support, mirroring the improvement in SQI over time, showing how public sentiment on social media aligns with broader trends in social quality.

—----------

Relatively bad week 1 is associated with terminology consistent with anger. #trumppandemic casts blame on the US president while #kag2020[KeepAmericaGreat2020] is a defensive reply. #coronaapocalypse is a recognition of the threat for possible grave implications of the pandemic and reflecting panic in the public.

And over the weeks, we see a improvement in SQI, because of which we begin to see an emphasis on an adaptive/coping response to the crisis. #hydroxychloroquine appears to be a promising treatment while #quarantinelife and #lockdown becomes the new normal, which rationalizes agoraphobia. #lightitblue supports and motivates the essential health care workers who will help to overcome the threat.

Week1 : #flattenthecurve describes the rate of increase in anxiety among the population due to rise in pandemic cases.

#trumppandemic: because of the inefficient response from the govt. which induced the panic among the population

#coronapocalypse: The COVID-19 was compared as a apocalypse because of its severe outbreak in the country which is causing the fear of social situation among the population

#kag2020(keepamericagreat2020): extreme fear as america topped in the ranking of countries severely affected by coronavirus pandemic.

Week4: #hydroxychloroquine because of its similarity with the drug benzodiazepine. [https://www.medintensiva.org/en-hydroxychloroquine-potentially-lethal-drug-articulo-S2173572717300577]

#quarantinelife can be related to the introverts meaning, the fear of being judged or rejected, rise in social phobia with COVID.

[https://www.technologyreview.com/2020/04/02/998440/lockdown-was-supposed-to-be-an-introverts-paradise-its-not]

#lightitblue was the initiative to light up the buildings in blue color in support of the essential frontline workers in the COVID-19 pandemic. Motivating them to overcome the rising irrational fear of pandemic and keep supporting (and helping) people survive the pandemic

#lockdown to keep the each other safe from the virus and prevent the spread of infection people instead it is giving rise to fear from social gathering or meet up (an initial sign to agoraphobia)

77 of 83

Hashtag Content Mirrors SQI

(steadily declining states)

77

SQI:

SQI better SQI worse SQI worse SQI worse

Hashtag:

#

Cluster 1:

WA, KS, IN, WY, OH, OR, NH

March 14-20 March 21-27 March 28-April 3 April 4-10

In this cluster of states, including Washington, Kansas, Indiana, Wyoming, Ohio, and New Hampshire, we see an interesting progression in hashtag content that mirrors the decline in SQI over time.

In Week 1, when the SQI is relatively good, hashtags like #wewillprevail and #familiesfirst reflect a sense of motivation, community support, and resilience during the early days of stay-at-home orders. We also see hashtags like #lightitblue and #thankatrucker, which highlight public support for essential workers, from healthcare professionals to truck drivers, helping communities through the lockdown.

However, as weeks progress and the SQI declines, we see a shift in the tone of hashtags. Terms like #hydroxychloroquine appear, suggesting growing dependency or desperation for solutions. Hashtags such as #stayathome and #breaking signal rising frustration and panic. With extended lockdowns, increasing unemployment, and dwindling financial resources, the tension builds, even leading to calls for reopening workplaces and potential unrest.

This shift from optimism and support to panic and distress aligns with the decline in social quality, showing how public sentiment evolves under prolonged stress and uncertainty.

—---------------------

The relatively good first week is associated with hashtags such as #wewillprevail and #families first to motivate and keep each other supported during the stay at home orders and lockdown. #ligthitblue and #thankatrucker shows the supports to essential works who are helping them survive in this state of lockdown.

And over the weeks here we see the decline in quality of SQI, which shows emphasis on #hydroxycholorquine which is similar to drug (benziodiazepine) that can be interpreted as people being dependent on sedatives. #stayathome and #breaking shows how after being at home for longer time and with increase in unemployment and no remaining financial resources, there is a state of panic building up leading to riots to open up work spaces.

Week1: #wewillprevail people supporting each other during the pandemic outbreak to release the pressure due to isolation, lockdown, school closure, and business closure.

#lightitblue Motivating them to overcome the rising irrational fear of pandemic and keep supporting (and helping) people survive the pandemic, was the initiative to light up the buildings in blue color in support of the essential frontline workers in the COVID-19 pandemic.

#familiesfirst to support people living alone and improving lack of motivation

#thankatrucker in support of the supply truck drivers reducing their extreme fear of pandemic

Week4: #hydroxychloroquine similar to drug (benzodiazepine) which can be interpreted as people being dependent on sedatives during the pandemic lockdown.

#stayhome staying in home for a longer time and increase in unemployment in the states is causing depression among the population

#ppe Because of unavailability of the PPE has raise issues on social phobia and fear from social situation.

#breaking The population coming out of their home for riots to open up the states, because of no financial resources remaining to survive on. This creates a condition of panic attacks which is leading in people breaking the lockdown.

78 of 83

78

Content

Foundations

Why Social Media Data
LLMs and its challenges
NeuroSymbolic Approach
Types of Knowledge Infused Learning and its advantages

Neurosymbolic Approach for an health application

Use-Case COVID-19

Hands-on Session

79 of 83

HANDS-ON Session

Modulating word embedding with Zero-shot learning
Neologism

79

Link to the notebook

Complete Github Repo:

80 of 83

Conclusion

Neurosymbolic AI integrates symbolic reasoning with neural networks to enhance adaptability.

Tackling LLM challenges improves efficiency in dynamic, noisy data environments like social media.

Applications in health domains showcase impactful use-cases, e.g., COVID-19 analysis.

The hands-on session equipped participants with practical skills to implement word embedding modulation and neologism.

80

NeuroSymbolic AI

Open Source Gen AI

Instructability

Alignment

Grounding

81 of 83

Conclusion

81

Instructability

Grounding

Alignment

The capability of AI systems to be taught and guided by humans to cause intentional behaviours.
Features:

Skill Acquisition
Knowledge-Gap management
Human-AI Interaction

Explainability
Interpretability
Observability

The process of establishing meaningful connections between AI representations and the real world, ensuring AI systems understand and interact with their environment effectively.
Features:

Symbol Grounding
Pragmatic Grounding
Compositional Grounding

Ensuring AI systems' goals, actions, and behaviors are consistent with end user expectations, for example human values and norms.
Features:

Value-based / Ethical Orientation
Task Orientation
Collaborative Functioning

Image by pngtree.com

Image by kjpargeter on Freepik

Image by gun21awan740843 on vecteezy

Interaction

Feedback

Human

AI

Human-Values

AI Actions

Real World Concepts

Last Layer

(AI Representation)

Commands, Queries and Responses

Corrections and Learnings

Image by pngtree.com

Image by medium

Image by emiltimplaru, santima.studio on vecteezy

Image by oval on clker.com

Image by pngtree.com

82 of 83

82

Primary funding support by NSF Awards #: 2133842, 2335967, 2119654, 2350302, WIPRO, BOSCH, others.

Learn more:

Neurosymbolic AI at AIISC and other projects: see http://wiki.aiisc.ai
Website - demos, open data/tools, tutorials, workshops, papers
LinkedIn - http://linkedin.com/company/aiisc
YouTube - http://youtube.com/aiisc (demos, tutorials, dissertations, keynotes, invited talks)

83 of 83

Thank You!

83