No difference between a data ontology and a world ontology
John F Sowa
John,
John> you don't have to believe me. You can just read those references.
I write my first program in Algol in 1975 and work with AI tasks (pattern recognition in chemistry) since 1977 and symbolic AI since 1980. I also remember all this AI-related fevers with LISP, Prolog, was participant of expert systems projects (something like frame system in 1980, first ontology editor for expert system was in 1985), last ontology editor (good for ISO 15926) is available in open source, we stopped development 8 years ago: https://github.com/TechInvestLab/dot15926. Therefore I was participant of all this trends in AI and ontology field, you need not tell me read something to be aware of.
John> It also analyzes 40 years of English documentation: specifications, user manuals, emails, comments in the COBOL code, etc,. And it relates the English to the programs to check for errors. Please reread slides 47 to 55 of cogmem.pdf.
I am know that your project have analyzed simple patterns from English text, it is definitely was in your slides. To me this is NLP (natural language processing), today in AI they use NLU (natural language understanding).
John> For our new Permion system, we use a new version of Prolog that supports tensors (and LLMs) as a built-in primitive.
For our question answering system that we use to help students with our courses of ontology engineering, systems thinking, systems engineering etc., we use call of LLM backend. This is 6 lines of python code. And configuration of a call is done with prompt, and this prompt is in Russian (we used multilanguage LLM). And this is not simple LLM, this is finetuned “Instruct” type that more aligned with our question answering tasks.
Most of development in LLM is goes to reasoning inside LLM because you usually not knowing what precisely you need to reason/infer with extreme variety of user questions even in narrow domain. You may need to infer something even before you will go to tool choice step (one of the tools can be e.g. prover, another is computer algebra system, or ).
There are many work in neural emulation of algorithm execution, graph processing, formal reasoning with knowledge graph (https://towardsdatascience.com/graph-ml-in-2023-the-state-of-affairs-1ba920cb9232 -- is a review of it). And there is trade-off in neural implementation of such algorithm and external “ordinary” virtual machine for algorithm, graph algorithm, prover/reasoner (along Toolformer line, https://arxiv.org/abs/2302.04761, «Toolformer: Language Models Can Teach Themselves to Use Tools»).
I completely agree with you that it is very helpful to take any program language (including Assembler and Prolog, but today in AI they use Python for everything) and call LLM from this. Moreover, you can write LLM module execution in Prolog, why not! Language is not important here. More important overall architecture. And in today’s architectures they often use retrieve with vectors instead of triple stores/knowledge graph.
John> . For the IBM project that defeated the Jeopardy champion, the basic reasoning engine was Prolog.
Your words about IBM Watson (http://ontolog.cim3.net/forum/ontolog-forum/2011-02/msg00178.html): «But in any case, my main point was that the IBM Watson system was more successful than Cyc in question answering because it got down to addressing real questions from the very beginning. They did not waste time defining millions of axioms». The less ontology work, the better, sure )))
They can use Fortran, Prolog, LISP or even Assembler. But I attended ontolog forum meeting that was devoted to this project. And they tell about working with full texts only: Watson is supercomputer for full texts processing, no encoding of it to any ontology! They use upper ontology of a couple of hundred terms to enhance rate of valid answers about 12-14%. They couldn't encode all the text in their database (screenplays, fiction, wikipedia, and so on), because something subtly got lost in compressing the information into a small set of relationships -- and the questions were about anything, including what wouldn't survive this compression of information when recoded into a knowledge graph. Similar approach was in “ISO 15926-2 industrial usage” when from all huge knowledge graph (mainly taxonomy of terms) they used only 201 type from upper ontology.
Modern vector database systems work with just this: they select relevant information for processing, and then pass it to the LLM -- and there this processing goes for many questions without loss of information when modeling some entities with rigid ontology lattice. And they have not even upper ontology here.
If we are talking about answering some known questions in advance (like finding some understandable type of error in program code and documentation), then there is no problem to make an algorithm for this -- even in Prolog, even in Fortran. To me your slides of Cobol error seeking project is about it.
Another topic is inference. You have Boolean algebra inference in Prolog. But I mentioned Bayesian inference and excess Bayesian (quantum-like). Role of these two in knowledge work in conditions of uncertainty is important. And this is my point: ontology should work with spectrum of formality, not only common logic type.
Best regards,
Anatoly
From: ontolog...@googlegroups.com <ontolog...@googlegroups.com> On Behalf Of John F Sowa
Anatoly,
From the very beginning of AI (Turing 1950, Machine translation 1950s, Theorem proving 1950s, and Neural networks 1940s), all the major methods were introduced. And all those methods have been under continuous development since then. If you don't believe me, please look at the summary of developments from the 1980s with a huge number of references: https://jfsowa.com/ikl/
I put together the citations of projects that implemented those developments. But you don't have to believe me. You can just read those references. If you think that there are more I should add, I would be happy to add them. I haven't updated that web page in the past few years because I have been busy with other projects. I'm thinking of adding a new section on LLMs, but right now I'm overwhelmed with other very important projects.
Anatoly> I consider type of the project that you have in your slides is very particular. Your project is something like prover for huge codebase.
No! It also analyzes 40 years of English documentation: specifications, user manuals, emails, comments in the COBOL code, etc,. And it relates the English to the programs to check for errors. Please reread slides 47 to 55 of cogmem.pdf.
The basic theorem prover is Prolog, which you can download for free. That language was introduced in the late 1970s, and it has been a fantastic success for every projec that used it. As cogmem.pdf shows, it can be used with large knowledge bases of every kind. For every application in the cogmem.pdf slides, you could download a free version of Prolog to do all that reasoning. The most important new feature (in 1999) was the Cognitive Memory system, which is described at the beginning of those slides. That is the "secret sauce", which is combined with Prolog to implement every application in cogmem.pdf. Please study those slides. They show a technology that can make GPT far more general and ***reliable***.
But there are many other reasoning engines that were developed in the late 20th c, which could be used with or instead of Prolog. The largest one is Cyc, which also has a huge ontology and precise and reliable theorem provers. For the IBM project that defeated the Jeopardy champion, the basic reasoning engine was Prolog.
For our new Permion system, we use a new version of Prolog that supports tensors (and LLMs) as a built-in primitive. But just the free version of Prolog is sufficient for doing all the applications in cogmem.pdf.
Anatoly> Software Industry today have another hot topics and another instruments, most of all writhing programs according to specifications in natural language, i.e. formalization task. Second is explanation task, “why my program not works as intended”. And you have here GitHub Copilot with huge success. It appeared that programming tasks benefit from common sense and therefore LLM for code have advantage over many other methods.
In summary, I totally agree with your requirements because that is what our new Permion.ai system does. LLMs, by themselves, cannot do reasoning of any kind. But LLMs plus Prolog are fantastic. But our Permion system is not open source. And the reason is simple: development require $$$. And without $$$, you can't do anything. I just told you the secret: combine LLMs with Prolog and make it open source, if you wish.
But without a reasoning engine -- LLMs by themselves -- you have a very good machine translation system. Everything else is flaky or halucinogenic or dangerous.
John
From: "Anatoly Levenchuk" <ai...@asmp.msk.su>
John,
I consider type of the project that you have in your slides is very particular. Your project is something like prover for huge codebase. There are many code analysis tools that deal with this (especially for security bugs), and your project is one of them (but for «business bugs»), but proprietary service, not a product. Software Industry today have another hot topics and another instruments, most of all writhing programs according to specifications in natural language, i.e. formalization task. Second is explanation task, “why my program not works as intended”. And you have here GitHub Copilot with huge success. It appeared that programming tasks benefit from common sense and therefore LLM for code have advantage over many other methods.
For Summit or Series I am not advising topics like “prover for huge codebase in Cobol” or “LLM in Software engineering», this is not about ontology. Ontology is about modeling (description with formalization) of the world. Creativity usually have iterative evolution algorithm that need formalization and deformalization steps (I modeled it after 3 practices, namely STA-methodology, TRIZ, ToC):
1. Formalize domain with common sense (give strong folk ontology typing to all entities, find relationships with strong types).
2. Find contradictions, it is errors in knowledge.
3. Deformalize (cancel most of the strong typing. E.g. go to another concept theory, from theory theory to prototype theory and use metaphors and conceptual blending).
4. Have a guess (from noise/chaos: more or less random!) what can be not so folk concepts in the domain.
5. Formalize with new conceptual schema.
6. Go to 2 and repeat till no contradictions in formal representation
LLM good for formalization and deformalization steps here and hallucinating guesses from noise to suggest alternative conceptualizations. This process can be done in 3 architectures:
- Inside huge LLM with some tooling and memory. Memory addition is very hot topic today, cycles based on memory that added to LLM.
- Adversarial with several LLM that have different tasks for different steps. This is some kind of cognitive architecture. LLM usually consists from multiple modules each of them is neural network but together they are complex structure (like brain that have many parts, neocortex only one of this parts, and all parts is wet neural networks that interconnected). Give them time and memory to iterate.
- LLM+people (interactively). This is usual today (all this LLM thing with sparks of reasoning appear only in March 2023, therefore you have not so many industrial solutions to 1 and 2 architectures).
Software engineering here is only one of domains. In this domain DDD (domain driven design) as ontology engineering is one of the most interesting practices, but not finding errors in legacy program code that is purely “reasoning with specialized model checkers” project. Model checkers, reasoners, provers is very helpful if you already have some code and some description that can be formalized to the same types that they already using in code. But LLM now works not only with code.
I support claim that “ontology” and “ontology engineering” is slightly outdated terms, while we can use them. Many people and now LLM-based systems (not pure LLM standalone!) do ontology engineering work (i.e. conceptualization and formalization in different forms) but not call it “ontology engineering”. We should be finding such practitioners and gather all of them. Computational ontology is not only about concept maps but about multiple representations of concepts, including connectionist ones. E.g. for knowledge works and search many people (including my company) use vector database + LLM, not triple stores + LLM for knowledge representations. We in ontolog forum should be less outdated in this regard, be more up to date. World modeling in AI (ontology engineering) includes representation learning with such ontology engineering techniques as contrastive representation learning, deep learning, self-supervising learning etc. for me is all about alternative conceptualization of ontology engineering. Mapping between this local and distributed knowledge representation is very interesting thing. My point is that creativity needs iterations between both representations:
-- search of contradictions that convenient in theory theory representations (concept maps, triple stores, common logic etc.)
-- cross-domain work (ontology merge, intuitive guesses, work with creativity that is depend on chaos and noise as novelty source) in prototype theory representations (vector databases, LLMs, prompt libraries, etc.)
-- creativity practice for alternative conceptualization creation (not documenting of folk ontology but inventing new ontology. Not FCA as formal concept analysis, but formal-informal concept synthesis).
E.g. see «Vector Database Use Cases» from one of vendors, https://qdrant.tech/use-cases/ -- 10 years ago it all can be from ontology-based sulutions, but not today. There is specter of formality: nonformal – probability based (quantum-like as excess Bayesian and classical Bayesian) – formal (Boolean, i.e. radical form of Bayesian). Ontology should be considered as concteptualisation on all these levels with all possible concept theories (https://plato.stanford.edu/entries/concepts/).
My suggestion is not to be centered on software verification with formal methods, but find contemporary ontology engineering that is masked by different names and today is everywere. E.g. DDD as ontology engineering -- https://en.wikipedia.org/wiki/Domain-driven_design. See there references about “Relationship to other ideas”, there not mention of «Ontology» there. “Ontology (information science)” mentioned only in “See also” section! Event storming as ontology engineerint practice for DDD is not regarded as ontology engineering -- https://en.wikipedia.org/wiki/Event_storming. But it definitely ontology engineering practice! All contemporary LLM work tend to deal with this formalization and deformalization (including deformalization for explanation purposes) techniques.
Best regards,
Anatoly
Anatoly Levenchuk
John,
Then may be will be helpful to cite my first two letters with multiple links, to have more context for this thread:
1)
I send several times notions about neural networks and its properties. Here is some works about it:
1. Neural networks are universal (i.e. any function! Including functions of reasoners, Turing machine etc.) approximator. There are multiple works about it: https://arxiv.org/abs/2301.00942 (and there see «2.3.1 Universal approximation results»), https://arxiv.org/abs/2204.07415 and many more. Neural network theoretically can emulate any device, including humans, classical computers with prover and reasoner algorithms etc.).
2. There are emergence properties (including reasoning skills) that emerge in neural networks (brains or artificial neural networks) with size. Here is key illustration of it: https://ai.googleblog.com/2022/04/pathways-language-model-palm-scaling-to.html. Multiple works speaks about it: https://arxiv.org/abs/2302.02083
Contemporary huge neural network have architecture “mixture of experts”. This is system composed of experts and this system have (as any system) emergent properties. E.g. GPT-4 consists of 16 experts and have 1.8 trillion parameters in all of them. It is huge and demonstrate multiple emergent properties.
Emergence properties can be greater with Toolformer architecture, https://arxiv.org/abs/2302.04761 (you should not thinking about “neural network alone” without usage of a tools – it is necessarily due to “no free lunch theorem”). For me it is similar to contemporary “accelerators” architecture of CYC (see "Representation and Reasoning Lessons Learned in Building Cyc", https://ontologforum.org/index.php/ConferenceCall_2022_02_02, video https://bit.ly/3L9YlDF).
3. Quantum-like (not quantum! But quantum-like: math of quantumness, not quantum physics) computation is common in biology and have relation to neural networks: https://www.sciencedirect.com/science/article/pii/S0303264720301994 is about quantum-like modeling, and multiple works that suggests that human decision-making is quantum-like, e.g. https://www.nature.com/articles/s41562-019-0804-2
Emergent Quantumness in Neural Networks -- https://arxiv.org/abs/2012.05082
4. Claude 2 is already in public and comparable with GPT-4 in its properties -- https://www.anthropic.com/index/claude-2 («For example, our latest model scored 76.5% on the multiple choice section of the Bar exam, up from 73.0% with Claude 1.3. When compared to college students applying to graduate school, Claude 2 scores above the 90th percentile on the GRE reading and writing exams, and similarly to the median applicant on quantitative reasoning. Users can input up to 100K tokens in each prompt, which means that Claude can work over hundreds of pages of technical documentation or even a book. Claude can now also write longer documents - from memos to letters to stories up to a few thousand tokens - all in one go. In addition, our latest model has greatly improved coding skills. Claude 2 scored a 71.2% up from 56.0% on the Codex HumanEval, a Python coding test. On GSM8k, a large set of grade-school math problems, Claude 2 scored 88.0% up from 85.2%»). This is not superhuman performance but comparable with humans performans.
5. To have better performance you simply should have multiple time calls to the network, e.g. in some kind of genetic algorithm, providing multiple calls to LLM, e.g. – most hot topic today is “LLM with memory” and Claude 2 have 100K tokens memory “right out of the box”. Other LLM can add memory with additional frameworks like in https://medium.com/@kamaljp/unleashing-the-llm-potential-with-memory-7-types-of-memory-to-use-in-langchain-376930ffa9f7
I am not mentioned many other hot topics, e.g. one of the most discussed now https://arxiv.org/abs/2306.11644 (Textbooks Are All You Need).
In my new book I also modeled several practices (systems think-activity methodology, TRIZ, theory of constraints) to show how use steps of formalization/ontology engineering for seeking errors in knowledge and deformalization for ontology merge steps in knowledge evolution. In human thinking this is about constantly moving between S1 and S2 mode of thinking according to Daniel Kaneman -- https://en.wikipedia.org/wiki/Thinking,_Fast_and_Slow
By the way: I am good English reader, but not listener, speaker and writer ))) And Google translate or even https://www.deepl.com/translator (it is better than Google) not very helpful in domain-specific texts. I think all this will change in a couple of years!
2)
there multiple works that perform test generation and discrepancies finding in codebases, automatic documentation creation etc. Е.g. «Is Your Code Generated by ChatGPT Really Correct? Rigorous Evaluation of Large Language Models for Code Generation», https://arxiv.org/abs/2305.01210, «Competitive programming with AlphaCode», https://www.deepmind.com/blog/competitive-programming-with-alphacode (AlphaCode achieved an estimated rank within the top 54% of participants in programming competitions by solving new problems that require a combination of critical thinking, logic, algorithms, coding, and natural language understanding).
Multiple such systems already in production, most popular of them is GitHub Copilot, https://github.com/features/copilot, and extensions like Microsoft Security Copilot, https://www.microsoft.com/en-us/security/business/ai-machine-learning/microsoft-security-copilot.
And there are a lot of researches that have very interesting results, https://github.com/saltudelft/ml4se
Sorry, but I do not know what you expect from contemporary software engineering AI tools that is similar to particular your case from the slides. Sure if you will have 8 weeks and availability of contemporary code models, you can do many more work than find several errors in huge codebase. May be you will find hundreds of errors!
And I do not suggest evaluate you program on common computer engineering benchmark (but developers of LLM for code usually have their leaderboard positions, may be you will be interested: https://paperswithcode.com/sota/code-generation-on-humaneval )
Best regard,
Anatoly
John F Sowa
John,
Then may be will be helpful to cite my first two letters with multiple links, to have more context for this thread:
1)
I send several times notions about neural networks and its properties. Here is some works about it:
1. Neural networks are universal (i.e. any function! Including functions of reasoners, Turing machine etc.) approximator. There are multiple works about it: https://arxiv.org/abs/2301.00942 (and there see «2.3.1 Universal approximation results»), https://arxiv.org/abs/2204.07415 and many more. Neural network theoretically can emulate any device, including humans, classical computers with prover and reasoner algorithms etc.).
2. There are emergence properties (including reasoning skills) that emerge in neural networks (brains or artificial neural networks) with size. Here is key illustration of it: https://ai.googleblog.com/2022/04/pathways-language-model-palm-scaling-to.html. Multiple works speaks about it: https://arxiv.org/abs/2302.02083
Alex Shkotin
--
All contributions to this forum are covered by an open-source license.
For information about the wiki, the license, and how to subscribe or
unsubscribe to the forum, see http://ontologforum.org/info/
---
You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/e08b80b7309d4fbfbb373e20ec76abbd%40bestweb.net.
Alex Shkotin
Anatoly,
--
Anatoly Levenchuk
John,
you already write this claims here in ontolog forum and I already familiar with them and disagree.
1. Argument from authority is not very good. E.g. Hinton already change his opinion about current developments in LLM-based methods, https://www.wired.com/story/geoffrey-hinton-ai-chatgpt-dangers/ (and this text is of May 2023, now is July – LLM is more elaborated). What about old experts in AI (including myself), there is Clarke’s three Laws about it (https://en.wikipedia.org/wiki/Clarke%27s_three_laws):
1. When a distinguished but elderly scientist states that something is possible, he is almost certainly right. When he states that something is impossible, he is very probably wrong.
2. The only way of discovering the limits of the possible is to venture a little way past them into the impossible.
3. Any sufficiently advanced technology is indistinguishable from magic.
My argument about approximating of any function by neural networks is from mathematics, there are multiple theorems about it, find links in 1. in first my letter that is about theorems. Main problem is computational budget, but this is solvable problem if you go from classical computers to e.g. optical (here is evaluation of optical transformers -- https://arxiv.org/abs/2302.10360, «we estimated that optical computers' advantage against current 300-fJ/MAC digital processors could grow to >100,000×»).
2. Nobody using vanilla neural network, thus all claims about it is wrong. This is to me like “motors cant fly”. Yes, motors themselves can’t fly, but motors as part of airplane can fly! The same thing with neural networks. Emergence (new properties) appear inside neural networks and with neural networks that assembled with other neural networks and other types of computing algorithms like reasoners, usual computer code for input-output, etc. Now we have Cambrian explosion in cognitive architectures. Vector databases if important part of this novel cognitive architectures, not only LLMs and InstructLLMs.
LLM is simply huge neural network that learned something. InstructLLM is neural network that learn slightly more than vanilla LLM. Chain-of-Thoughts is one of the methods of improving results of question answering (this is like genetic algorithm on neural network, it can generate novelty). Yes, all this architectures can do abduction.
3. My point is not about LLM but about ontology work in spectrum of formality. E.g. pseudocode level for computations. It appears that fully formalize we need only very important things, but to find what important in some situation we need perform not very formal computations. And this is not about fuzzy sets. This is about Bayesian inference and quantum-like inference.
Best regards,
Anatoly
James Davenport
The question of whether any function can be approximated/defined by a neural network depends on whether we are talking about extensional or intensional definitions.
In the extensional sense, as Anantoly points out, there are mathematical proofs that, given the extensional definition of a function, we can approximate/define it by a neural net.
But, of course, we cannot generally give the extensional definition, such as the ‘right’ move in all Go situations, since (a) we don’t know it
(b) it would be computationally intractable.
As John points out, some intensional definitions seem very hard if not impossible, e.g. “the best possible Go player”.
James Davenport
Hebron & Medlock Professor of Information Technology, University of Bath
National Teaching Fellow 2014; DSc (honoris causa) UVT
Former Fulbright CyberSecurity Scholar (at New York University)
Former Vice-President and Academy Chair, British Computer Society
From: ontolog...@googlegroups.com <ontolog...@googlegroups.com>
On Behalf Of Anatoly Levenchuk
Sent: 15 July 2023 09:50
To: so...@bestweb.net; 'ontolog-forum' <ontolo...@googlegroups.com>; ontolog...@googlegroups.com
Subject: [Ontology Summit] RE: No difference between a data ontology and a world ontology
|
CAUTION: This email came from outside of the University. To keep your account safe, only click on links and open attachments if you know the person who sent the email, or you expected to receive this communication. |
--
All contributions to this forum are covered by an open-source license.
For information about the wiki, the license, and how to subscribe or
unsubscribe to the email, see http://ontologforum.org/info/
---
You received this message because you are subscribed to the Google Groups "ontology-summit" group.
To unsubscribe from this group and stop receiving emails from it, send an email to
ontology-summ...@googlegroups.com.
To view this discussion on the web visit
https://groups.google.com/d/msgid/ontology-summit/005201d9b6f9%2457800560%2406801020%24%40asmp.msk.su.
Anatoly Levenchuk
James,
I fully agree with you. Then “learn” in deep learning is about learning of intensional definition of a function, and “inference” is about executing learned function. Sure, we can speak only about approximations, including optimisations (e.g. “best player” is about maximum of player function functional, calculus of variations. But it still possible to approximate! E.g. SoTA approximation of best possible Go player was Lee Sedol, then AlphaGo, then AlphaZero. Best model/approximation of a cat is another cat, preferably the same cat).
Neural network can guessing (hallucinate with temperature property, that is level of random noise as a base of novelty in results).
And Prolog programs definitely can do all of this, even Assembler program can do all of this! Machine learning algorithms with distributed representations can be written in any Turing-complete language.
My point is about ontology in distributed representations and ontology-based reasoning. Common logic, Bayesian networks, quantum-like inference needs different representations.
Best regards,
Anatoly
unsubscribe to the forum, see http://ontologforum.org/info/
---
You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/LO6P302MB00165BF5AEDEF8C0DCA8DC80E035A%40LO6P302MB0016.GBRP302.PROD.OUTLOOK.COM.
Kingsley Idehen
LLM is more elaborated). What about old experts in AI (including myself), there is Clarke’s three Laws about it (https://en.wikipedia.org/wiki/Clarke%27s_three_laws):
1. When a distinguished but elderly scientist states that something is possible, he is almost certainly right. When he states that something is impossible, he is very probably wrong.2. The only way of discovering the limits of the possible is to venture a little way past them into the impossible.
3. Any sufficiently advanced technology is indistinguishable from magic.
Hi Anatoly,
I agree with points 1-3. That said, in regards to LLM-based Natural Language Processors like GPT, I agree with fundamental point that John Sowa is making -- based on experimentation aimed at pushing the boundaries of these tools. For instance, I've been testing and experimenting with these tools in a variety of ways since last November, and nothing I've discovered refutes the following fundamental point:
A good natural language processor != a fluid knowledge base.
Thus, the way to make best (and safe) use of them is to loosely couple them with domain-specific knowledge bases.
The great news (from my experimental experience) is that LLMs like GPT have been designed with loose coupling in mind, courtesy of the following integration options:
1. Plugins
2. Fine-tuning via declarative query languages (e.g., SQL and SPARQL)
3. External function integration via callbacks i.e., you can hook their APIs into an programming language (or environment) that support the Open API standard and then use callback to facilitate very powerful integration.
BTW -- I've just started experimenting with Claude2 from
Anthropic, and it's a very close match to ChatGPT (in GPT4 mode).
In addition, behavior described in ChatGPT (e.g., for a
line-of-business smart agent) is portable to Claude, due to the
common use of standards for structured data representation [2].
Related:
[1] Collection of HowTo oriented posts from my ChatGPT related experiments
[2] Tweet
about OpenLink Agent portability across ChatGPT and Claude2
-- Regards, Kingsley Idehen Founder & CEO OpenLink Software Home Page: http://www.openlinksw.com Community Support: https://community.openlinksw.com Weblogs (Blogs): Company Blog: https://medium.com/openlink-software-blog Virtuoso Blog: https://medium.com/virtuoso-blog Data Access Drivers Blog: https://medium.com/openlink-odbc-jdbc-ado-net-data-access-drivers Personal Weblogs (Blogs): Medium Blog: https://medium.com/@kidehen Legacy Blogs: http://www.openlinksw.com/blog/~kidehen/ http://kidehen.blogspot.com Profile Pages: Pinterest: https://www.pinterest.com/kidehen/ Quora: https://www.quora.com/profile/Kingsley-Uyi-Idehen Twitter: https://twitter.com/kidehen Google+: https://plus.google.com/+KingsleyIdehen/about LinkedIn: http://www.linkedin.com/in/kidehen Web Identities (WebID): Personal: http://kingsley.idehen.net/public_home/kidehen/profile.ttl#i : http://id.myopenlink.net/DAV/home/KingsleyUyiIdehen/Public/kingsley.ttl#this
Anatoly Levenchuk
Kingsley,
I think that we speak about the same thing, different names. E.g. vector database is used exactly for implementing domain-specific knowledge base with LLM. Most implementations of LLM in corporate services base not in vanilla LLM but in some cognitive architecture (and plug-ins to commercial ChatGPT is only one variant of it). I am not experimenting with ChatGPT too much but already have domain-specific chat-bot in my company. It answers student’s domain-specific questions about our courses (is alternative to full-text search). It based on most popular framework for LLM, LangChain (https://github.com/hwchase17/langchain). Moreover it answers in Russian because textbooks in our courses is in Russian. E.g. screenshot is about question «What is ontology» consists from paragraph with answer that composed by AI (LLM+vector base) and then link to parts of the courses that was loaded to database and most relevant to answer the question. We work with this domain-specific LLM-based technology since 28 March 2023 and use GPT-3.5-turbo (it is cheaper and sufficient to answer simple questions). 
My point is about multiple level of formality in ontology representation: common logic, pseudocode, natural language, vector database (as domain-specific knowledge base), LLM etc. (with local and connectionist/distributed representation). And logic inference, Bayesian inference and quantum-like inference to use these ontologies for decision making.
Best regards,
Anatoly
Kingsley Idehen
Kingsley,
I think that we speak about the same thing, different names. E.g. vector database is used exactly for implementing domain-specific knowledge base with LLM. Most implementations of LLM in corporate services base not in vanilla LLM but in some cognitive architecture (and plug-ins to commercial ChatGPT is only one variant of it). I am not experimenting with ChatGPT too much but already have domain-specific chat-bot in my company. It answers student’s domain-specific questions about our courses (is alternative to full-text search). It based on most popular framework for LLM, LangChain (https://github.com/hwchase17/langchain).
Yes, I am aware of LangChain and what it offers with regards to
LLM integration that does enable loose-coupling with knowledge
bases. It is indeed a popular framework for achieving such goals.
Moreover it answers in Russian because textbooks in our courses is in Russian. E.g. screenshot is about question «What is ontology» consists from paragraph with answer that composed by AI (LLM+vector base) and then link to parts of the courses that was loaded to database and most relevant to answer the question. We work with this domain-specific LLM-based technology since 28 March 2023 and use GPT-3.5-turbo (it is cheaper and sufficient to answer simple questions).
My point is about multiple level of formality in ontology representation: common logic, pseudocode, natural language, vector database (as domain-specific knowledge base), LLM etc. (with local and connectionist/distributed representation). And logic inference, Bayesian inference and quantum-like inference to use these ontologies for decision making.
Best regards,
Anatoly
The use of ontologies for decision making can take many forms. For instance, via informing reasoning and inference that's incorporated in the query solution production pipelines. We even use such techniques in our bot, very similar to what you illustrate above, but courtesy of SPARQL and function integration with ChatGPT i.e., rather than using Langchain, we leverage our own Virtuoso RDBMS which has native support for HTTP, websockets, SPARQL etc..
In conclusion, I think we all agree that these LLMs need loose coupling with domain-specific knowledge bases + plus a bit of reasoning and inference middleware to be useful beyond hype-generating demos :)
Kingsley
Anatoly Levenchuk
Kingsley,
yes, we are going similar ways. My point is that there are multiple levels of expected formality in knowledge representation / world modeling / ontology engineering and you need tolerate/relax formality in one cases and additionally formalize in another cases. In general you need knowledge bases that is memories in different levels of formality, e.g. full text repositories, vector databases, triple stores etc. and LLM and reasoners for inference on these knowledge bases. And “ontology” is not term that applied to mathematical logic (Boolean algebra) level of formality, but to world modeling overall. Thus reasoners should be of different nature, tolerable to uncertainty, may be consisting of loosely coupled multiple sub-reasoners worked on common knowledge base. Answer to question “what is in the world there” is in all parts of this knowledge base, not only in formal logical part with triple- or relational store. Distributed ontology representations (like in LLMs) is a valid subject of ontology as a discipline, thus also in the scope of ontology forum and ontology summit. Now it is valid also to “computational ontology” because we can compute these connectivist representations.
Best regards,
Anatoly
--
All contributions to this forum are covered by an open-source license.
For information about the wiki, the license, and how to subscribe or
unsubscribe to the forum, see http://ontologforum.org/info/
---
You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/d01cb9bb-fc6e-c4e4-32b2-06ebd3e21ce0%40openlinksw.com.
John F Sowa
Anatoly Levenchuk
By the way there are multiple projects that already working for formal ontology works with usage of LLM, e.g. https://github.com/KRR-Oxford/DeepOnto (DeepOnto -- A package for ontology engineering with deep learning, https://arxiv.org/abs/2307.03067).
Applying deep learning techniques, particularly language models (LMs), in ontology engineering has raised widespread attention. However, deep learning frameworks like PyTorch and Tensorflow are predominantly developed for Python programming, while widely-used ontology APIs, such as the OWL API and Jena, are primarily Java-based. To facilitate seamless integration of these frameworks and APIs, we present Deeponto, a Python package designed for ontology engineering. The package encompasses a core ontology processing module founded on the widely-recognised and reliable OWL API, encapsulating its fundamental features in a more "Pythonic" manner and extending its capabilities to include other essential components including reasoning, verbalisation, normalisation, projection, and more. Building on this module, Deeponto offers a suite of tools, resources, and algorithms that support various ontology engineering tasks, such as ontology alignment and completion, by harnessing deep learning methodologies, primarily pre-trained LMs. In this paper, we also demonstrate the practical utility of Deeponto through two use-cases: the Digital Health Coaching in Samsung Research UK and the Bio-ML track of the Ontology Alignment Evaluation Initiative (OAEI).
Relevant Publications
- [1] Yuan He‚ Jiaoyan Chen‚ Denvar Antonyrajah and Ian Horrocks. BERTMap: A BERT−Based Ontology Alignment System. In Proceedings of 36th AAAI Conference on Artificial Intelligence (AAAI-2022). /arxiv/ /aaai/
- [2] Yuan He‚ Jiaoyan Chen‚ Hang Dong, Ernesto Jiménez-Ruiz, Ali Hadian and Ian Horrocks. Machine Learning-Friendly Biomedical Datasets for Equivalence and Subsumption Ontology Matching. The 21st International Semantic Web Conference (ISWC-2022, Best Resource Paper Candidate). /arxiv/ /iswc/
- [3] Jiaoyan Chen, Yuan He, Yuxia Geng, Ernesto Jiménez-Ruiz, Hang Dong and Ian Horrocks. Contextual Semantic Embeddings for Ontology Subsumption Prediction. World Wide Web Journal （WWWJ-2023). /arxiv/ /wwwj/
- [4] Yuan He‚ Jiaoyan Chen, Ernesto Jiménez-Ruiz, Hang Dong and Ian Horrocks. Language Model Analysis for Ontology Subsumption Inference. Findings of the Association for Computational Linguistics (ACL-2023). /arxiv/ /acl/
I repeat here my point about spectrum of formality that is needed for world modeling. Work with LLMs and inference with neural network to me is about world modeling thus is ontology engineering. We need shift ontolog forum from pure logic (e.g. common logic) ontology representations to another forms of computational ontology representations, especially natural language an multimodal representations (e.g. texts with images in textbooks or engineering standards like ISO standards) and connectionist models like LLMs and chat-LLMs (additionaly trained LLMs). Add here ontology-related LoRA models that is domain-specific and knows something domain-specific about objects in pictures, https://stable-diffusion-art.com/lora/ and here is repository of 1501 LoRA models (I can think about it as 1501 domain-specific ontologies resembled «visual microtheories» that used for visual representations but connected to texts and world ontology that name entities in anime world), https://civitai.com/tag/lora (there are mostly anime-related models/microtheories, but I remember how we here in ontolog summit 2014 have Hackathon project «Reference data for Anime and Manga: Semantic Linking and Publishing of Diverse Data-Sets», http://ontolog.cim3.net/file/work/OntologySummit2014/2014-02-27_OntologySummit2014_Hackathon-Launch/wip/Agroskin_OntologySummit2014_Hackathon-Launch_Anime-Manga-Dataset-Semantics--VictorAgroskin_20140227b.pdf ). If we have such Hackathon today and keep it state-of-the-art (SoTA) we should use LLMs and LoRA models to model anime world! In 2014 year we used LLM in wet neuron network of our brains to perform formalization work and ontology editor .15926 to write results, but essence of ontology work (defining objects that is interesting in a world) was performed by neural networks in our brain. The same will be with artificial neuron networks (ANN, including pretrained ones, i.e. LLMs).
Best regards,
Anatoly
--
All contributions to this forum are covered by an open-source license.
For information about the wiki, the license, and how to subscribe or
unsubscribe to the forum, see http://ontologforum.org/info/
---
You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/bf3c21b25d3548e4a099134aa8cee8a4%40bestweb.net.