From NLP to NLU: Why we need varied, comprehensive, and stratified knowledge, and how to use it for Neuro-symbolic AI
Download as PPTX, PDF1 like100 views
The document presents a keynote by Amit Sheth on the transition from Natural Language Processing (NLP) to Natural Language Understanding (NLU) and the importance of integrating comprehensive knowledge in neuro-symbolic AI. It addresses the limitations of current language models, including issues related to distributional semantics and common sense reasoning, while proposing knowledge-infused approaches to improve AI systems. The discussion covers various application domains such as healthcare, education, and disaster management, emphasizing the need for explainability and safety in AI implementations.
![Contents
3
1. Challenges with Current LMs
2. Possible Fixes and Limitations
3. Knowledge Infused Neuro-Symbolic AI
4. Before Transformers, Historical Context [Neuro Symbolic AI]
5. After Transformers, Current Context [Neuro Symbolic AI]
6. Future Context [Neuro-Symbolic AI] with Process Knowledge
Infusion.](https://image.slidesharecdn.com/knowledgenlpataaai2023keynote-230217013445-43486683/85/From-NLP-to-NLU-Why-we-need-varied-comprehensive-and-stratified-knowledge-and-how-to-use-it-for-Neuro-symbolic-AI-3-320.jpg)
![Challenges with Current LMs
4
Tycoon
Did you mean: tycoon
Did you mean: typhoon
Did you mean: typography
tycoon:
0.00567%
Distributional Semantics:
Language Models are trained to compute the
distributional plausibility of language tokens
from enormous amounts of training tokens.
World Model Semantics:
Relationships and effects among the objects
that the language tokens describe, grounded
in the physical or conceptual reality of the
world humans experience (ontological
commitment).
Do World Model Semantics Arise as an
Emergent Capability of Distributional
Semantics at Scale? [No- not directly, not
specifically, not adequately]
Sheth, A., Ramakrishnan, C., & Thomas, C. (2005). Semantics for the semantic web: The implicit, the formal and the powerful. International Journal on Semantic Web and
Information Systems (IJSWIS), 1(1), 1-18., 2005, link
World Model Semantics from Distributed Semantics at Scale?
1.](https://image.slidesharecdn.com/knowledgenlpataaai2023keynote-230217013445-43486683/85/From-NLP-to-NLU-Why-we-need-varied-comprehensive-and-stratified-knowledge-and-how-to-use-it-for-Neuro-symbolic-AI-4-320.jpg)
![Challenges with Current LMs
5
Formally
∀i ⊆ [N], N = Vocabulary Size,
Model the probability:
P(tokeni, …. , tokenN | token1, … tokeni)
Example Usage
Once Trained on Enormous amounts of Data,
Predict:
1. y ~ P(S = “Obama just won the 2032
Election”)?
2. P(S) does not pass a threshold test
=> y = 0, or False.
Tycoon
Did you mean: tycoon
Did you mean: typhoon
Did you mean: typography
tycoon:
0.00567%
Recency : World Model Semantics from Distributed Semantics at Scale
1.](https://image.slidesharecdn.com/knowledgenlpataaai2023keynote-230217013445-43486683/85/From-NLP-to-NLU-Why-we-need-varied-comprehensive-and-stratified-knowledge-and-how-to-use-it-for-Neuro-symbolic-AI-5-320.jpg)
From NLP to NLU: Why we need varied, comprehensive, and stratified knowledge, and how to use it for Neuro-symbolic AI
- 1. From NLP to NLU: Why we need varied, comprehensive, and stratified knowledge, and how to use it for Neuro-symbolic AI Keynote at KnowledgeNLP-AAAI’23 Amit Sheth Director, AI Institute of South Carolina University of South Carolina amit@sc.edu #AIISC, http://aiisc.ai
- 2. AIISC portfolio in Core AI & Translational AI 2 Knowledge-infused Learning (Neuro- symbolic /Hybrid AI) Knowledge Graph Development Deep Learning Reinforcement Learning Natural Language Processing/ Understanding/Generation Multimodal AI (IoT/sensor, data streams, images, emoji) Collaborative & Personal Assistants Multiagent Systems Interpretability/ Explainability/Safety/ Trust/Ethics in AI Medicine/Healthcare/Nursing (Nutrition, Neurodevelopmental Disease, Asthma, Diabetes, Hypertension, Autism, Aphasia, Cognitive Disorders, Oncology,...) Neuroscience Brain Science Epidemiology Education Social Good/Harm (Disinformation, Harassment,Toxic Content, Deception, Extremism, Radicalization) Public Health (Mental Health, Addiction, COVID-19, Epidemics) Smart Manufacturing (Digital Twins, Factory of Future) Disaster Management (Response, Resilience Pharma: drug discovery, vigilance Autonomous Systems (Vehicles) Automated Planning Computer Vision Gaming Cognitive Science Science & Engg: Radiation, Astrophysics, Civil Infra & Transportation Law
- 3. Contents 3 1. Challenges with Current LMs 2. Possible Fixes and Limitations 3. Knowledge Infused Neuro-Symbolic AI 4. Before Transformers, Historical Context [Neuro Symbolic AI] 5. After Transformers, Current Context [Neuro Symbolic AI] 6. Future Context [Neuro-Symbolic AI] with Process Knowledge Infusion.
- 4. Challenges with Current LMs 4 Tycoon Did you mean: tycoon Did you mean: typhoon Did you mean: typography tycoon: 0.00567% Distributional Semantics: Language Models are trained to compute the distributional plausibility of language tokens from enormous amounts of training tokens. World Model Semantics: Relationships and effects among the objects that the language tokens describe, grounded in the physical or conceptual reality of the world humans experience (ontological commitment). Do World Model Semantics Arise as an Emergent Capability of Distributional Semantics at Scale? [No- not directly, not specifically, not adequately] Sheth, A., Ramakrishnan, C., & Thomas, C. (2005). Semantics for the semantic web: The implicit, the formal and the powerful. International Journal on Semantic Web and Information Systems (IJSWIS), 1(1), 1-18., 2005, link World Model Semantics from Distributed Semantics at Scale? 1.
- 5. Challenges with Current LMs 5 Formally ∀i ⊆ [N], N = Vocabulary Size, Model the probability: P(tokeni, …. , tokenN | token1, … tokeni) Example Usage Once Trained on Enormous amounts of Data, Predict: 1. y ~ P(S = “Obama just won the 2032 Election”)? 2. P(S) does not pass a threshold test => y = 0, or False. Tycoon Did you mean: tycoon Did you mean: typhoon Did you mean: typography tycoon: 0.00567% Recency : World Model Semantics from Distributed Semantics at Scale 1.
- 6. Challenges with Current LMs 6 Distributional Semantics: Language Models are trained to compute the distributional plausibility of language tokens from enormous amounts of training tokens. E.g., What was the color of the white horse of Napolean? It is not very distributionally plausible that a person asks the answer to a question they already know The answer is abundantly clear in the question But, (a) distribution semantics has problem with the low probability, and (b) it does not understand! More challenges arise due to lack of context. Common Sense : World Model Semantics from Distributed Semantics at Scale 1.
- 7. Challenges with Current LMs 7 Distributional Semantics: Language Models are trained to compute the distributional plausibility of language tokens from enormous amounts of training tokens. E.g., Mike’s mum had 4 kids; 3 of them are X,Y,Z. What is the name of 4th kid? - Missing information is not clear Answer is obviously Mike 1. Mike is not a pet 2. More than one child cannot have the same name 3. Kid names cannot be any word 4. Impossible answers even after the hint: Mike, Luis, Drake, and Matilda, and All other tokens in the input! Distribution semantics does not have commonsense and probability does not help. Challenge - Missing Information: World Model Semantics from Distributed Semantics at Scale 1.
- 8. Challenges with Current LMs 8 Distributional Semantics: Language Models are trained to compute the distributional plausibility of language tokens from enormous amounts of training tokens. E.g., Mike’s mum had 4 kids;3 of them are X,Y,Z. What is the name of 4th kid? Human’s first answers (We asked 10 fellow humans) 1. Mike 2. I think this is a trick question Distribution semantics does not know to make human-like assumptions about missing information. Challenge - Missing Information: World Model Semantics from Distributed Semantics at Scale 1.
- 9. Challenges with Current LMs 9 Distributional Semantics are a Cloud of Probabilities The director, AIISC, of, Amit Sheth All other words Prof I kind of understand (not really though) Dense Representational Spaces Hallucinations - It makes up things ! 1.
- 10. Recency Insufficiency of data alone Common-sense Hallucinations Challenges to be addressed for NLU 10 User-Explainability Application-level Safety Characteristics of NLU Capable Systems 1.
- 11. 1. 11 Challenges to be addressed for NLU
- 12. Possible Fix: Instruct GPT 12 The Instruct GPT Framework Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., ... & Lowe, R. (2022). Training language models to follow instructions with human feedback. arXiv preprint arXiv:2203.02155. link No Problem I will make things larger and use Instruction based Training Hallucinations - It makes up things ! 2.
- 13. 13 1. 40 humans to capture the breadth of knowledge corresponding to the data in LLMs seems small 2. The richness of human knowledge is compressed into a mere label 2.
- 14. 14 Prof: Amit Sheth Synonyms: Amit Sheth Degree:Ph.D. Synonyms: Doctorate Company: AI Institute, Abbr: AIISC Synonyms: AI Institute, South Carolina employee_of Let’s see what this model of understanding can yield Knowledge (Graphs) to the rescue Fix: Addressing Hallucinations: Recency, Common-Sense, and Implicit Entity Mentions, etc., 2.
- 15. 15 World as Concepts vs. World as Probabilities. Explicit model of recency and common-sense. Supply missing Knowledge (entities, relationships). Semantics supported by Knowledge (Graphs) User-level Explanations and Safety Constraints 2.
- 17. 3. Knowledge Infused Neuro-symbolic AI Using Graphs in Neural Network Pipelines
- 18. 18 3. Shallow Infusion Semi-Deep Infusion Shades of KiL - Shallow and 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
- 19. 3. Shades of KiL - Deep Infusion 19 Deep Infusion Sheth, Gaur, Kursuncu, & Wickramarachchi, (2019). Shades of knowledge-infused learning for enhancing deep learning. IEEE Internet Computing, 23(6), 54-63., link
- 20. Sheth, Gaur, Kursuncu, & Wickramarachchi, (2019). Shades of knowledge-infused learning for enhancing deep learning. IEEE Internet Computing, 23(6), 54-63., link Characteristics/Method Distributed Semantics Shallow Infusion Semi-Deep Infusion Deep Infusion Recency U-M- M M+ H Filling in Missing Information U U M H Hallucinations Unsatisfactory (U) Get by but not really solving the problem (M) Better but not fully solve the problem (M+) Broadly solve the problem (H) Characteristics of Knowledge Infusion 3. 20
- 21. 3. KiL - Generic Architecture 21
- 23. 4. KiL - SEDO (Shallow Infusion) 23 Gaur, Kursuncu, Alambo,, Sheth, Daniulaityte, Thirunarayan, & Pathak. (2018, October). " Let Me Tell You About Your Mental Health!" Contextualized Classification of Reddit Posts to DSM-5 for Web-based Intervention. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management (pp. 753-762)., link
- 25. 5. KiL - K-Adapter (Shallow Infusion) 25 Wang, R., Tang, D., Duan, N., Wei, Z., Huang, X. J., Ji, J., ... & Zhou, M. (2021, August). K-Adapter: Infusing Knowledge into Pre-Trained Models with Adapters. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021 (pp. 1405-1418), link
- 26. 5. KiL - KALA (Semi-Deep Infusion) 26 Kang, M., Baek, J., & Hwang, S. J. (2022, July). KALA: Knowledge-Augmented Language Model Adaptation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (pp. 5144-5167)., link
- 27. 5. KiL - TDLR (Semi-Deep Infusion) 27 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
- 28. Knowledge Contexts Leads to Performance Gains With Smaller Models Knowledge Contexts Leads to Performance Gains With Smaller Models and Smaller Datasets 5. TDLR - Results 28 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
- 29. Knowledge Infused Neuro-symbolic AI Integrating Lifted Neural Representations with Knowledge Graphs 6.
- 30. 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 6. 30
- 31. 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 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 <is symptom> 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 6. Knowledge Verified Interpretable Prediction through linking to KG and definitions 31 Gaur, M., Desai, A., Faldu, K., & Sheth, A. (2020). Explainable ai using knowledge graphs. In ACM CoDS-COMAD Conference. Link, slide.
- 32. 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 6. Knowledge Verified Interpretable Prediction through Process Knowledge Structures 32 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
- 33. Do you feel nervous? More than half the days Do you feel irritated or self destructive? Do you feel something extreme might happen to you? Are you able to relax? Do you feel nervous? More than half the days Do you feel Irritated? Are you bothered by becoming easily annoyed or irritable? Are you bothered by any relaxation troubles? Knowledge Infusion using Medical Questionnaire (MedQ) These questions are medically valid and safe, in right sequence.. Safety Check s 6. Knowledge Verified Interpretable and Safe Text Generation through Process Knowledge Structures 33 Roy, K., Gaur, M., Soltani, M., Rawte, V., Kalyan, A., & Sheth, A. (2023). ProKnow: Process knowledge for safety constrained and explainable question generation for mental health diagnostic assistance. Frontiers in Big Data, 5., link
- 34. ● If the system were to give user-level explanation, it will need to incorporate/use conceptual model (vocabularies, knowledge graph) used by the user. Purely data driven system can at best give explanations that ML engineers (developers) can use. ● Knowledge is multifaceted. I presented diverse knowledge to support different levels of abstractions for NLU. There will be different knowledge for abstractions involved in image understanding activity. ● Should you bring knowledge to the data level (embedding) or bring data to the knowledge level (learn from data and align with knowledge)? For less demanding intellectual activities (classification, prediction, recommendation) the former will do; for more demanding intellectual activities (decision making and taking actions with explanations), latter is most likely needed. Takeaway 34
- 35. 35 Acknowledgement Funding supported in part by NSF Award#: 2133842 EAGER: Advancing Neuro-symbolic AI with Deep Knowledge-infused Learning and other projects (see http://wiki.aiisc.ai) Learn more: ● Website - http://aiisc.ai (projects, people, opensource, demos, open data/tools, tutorials, workshops, papers) ● Wiki Page - http://wiki.aiisc.ai ● LinkedIn - http://linkedin.com/company/aiisc ● YouTube - http://youtube.com/aiisc (demos, tutorials, dissertations, keynotes, invited talks)
- 36. Sheth core group AIISC Artificial Intelligence Institute of South Carolina (#AIISC) Kaushik Roy Vedant Khandelwal Also, Megha Chakraborty, Vipula Rawte, Yuxin Zi 36 Contribution/special thanks (this talk):