[netmod] Re: [NMOP] Re: Re: Re: Semantic meaning across representations and boundaries

[Qin Wu <[e-mail]>]

Hi, Frank:
TM Forum has developed Intent common model (https://www.tmforum.org/resources/introductory-guide/intent-common-model-v3-8-0-tr290/) which described intent objectives, scope, constraints, context, and expected outcomes.
In addition, TM Forum has developed A2A-T, A2A-T can be used to carry task prompt template as A2A payload which serve as semantic layer to describe intent objectives, scope, constraints, context and expected outcome, see
https://www.tmforum.org/resources/introductory-guide/ig1453-agent-to-agent-protocol-for-telecoms-v2-1-0/ for more details.

I am wondering how semantic model you describe below is different from what TM Forum is doing.
Second, for mechanisms that bind semantic models to existing representations, is this related to MCP mapping mechanism used in the northbound interface of the network device? Or similar to mapping from service level representation to network level representation used in the northbound of Network Controller/Network Management Agent?
Is there any tools or prototype to demonstrate or validate this idea?

-Qin
发件人: chong feng <[e-mail]>
发送时间: 2026年8月5日 21:28
收件人: Christopher Janz <[e-mail]>
抄送: Qin Wu <bill.wu=[e-mail]>; nmop <[e-mail]>; [e-mail]; [e-mail]; [e-mail]; [e-mail]
主题: [NMOP] Re: [netmod] Re: Re: Semantic meaning across representations and boundaries

Hi Chris, Qin, all,

Thanks for the very insightful discussion. I think the discussion is converging toward an important distinction: the semantic model itself and the mechanisms that bind semantic models to existing representations are related but different problems.

I agree with Chris that the fundamental challenge is to make meaning explicit and independent from its representation form. In this sense, ontology or semantic models provide a conceptual layer that describes concepts, relationships, constraints, and contextual meaning. This layer is not specific to YANG, REST API, or any particular data representation.

However, in practical infrastructure environments, we also need to consider how such semantic meaning can interact with existing deterministic systems. Most existing infrastructures will continue to use their current models and interfaces, such as YANG-based management systems, REST APIs, CLI, and other domain-specific models.

This is where I see the role of NAIM and RAIM. They are not intended to replace ontology or define another universal knowledge model. Instead, they provide semantic bindings between AI-consumable representations and existing deterministic representations.

NAIM is an attempt to define such a semantic binding for YANG-based network management systems, while RAIM (REST AI Interface Modeling) applies the same idea to REST API based systems:

https://github.com/agentic-intent-network/raim-spec

A possible layered view could be:

              Ontology / Semantic Model
                       |
              Semantic Binding Layer
                  / \
              NAIM RAIM
                |           |
              YANG REST API

The upper layer defines shared meaning, while NAIM/RAIM address how that meaning can be mapped into existing operational systems without requiring changes to those systems.

It is also possible to operate without a fully defined ontology initially:

              LLM / Agent
                    |
              NAIM / RAIM
                    |
             Existing Systems

In this case, NAIM/RAIM provide an intermediate representation that is easier for LLMs to consume than raw operational models, while preserving compatibility with existing infrastructure.

Therefore, I think AI-consumable models have two complementary aspects:

defining machine/LLM-understandable meaning across domains;

providing practical mappings from that meaning to existing deterministic systems.

The first is a semantic modeling problem, while the second is an integration and interoperability problem. Keeping these two aspects separate may help us better identify where different efforts (ontology, knowledge graphs, annotations, NAIM, RAIM, etc.) fit.

Best,Frank

Christopher Janz <[e-mail]<mailto:[e-mail]>> 于 2026年8月5日周三 下午8:25写道：

Qin — thank you for the careful review and detailed suggestions.

On your side meeting's aim: understood, it's to frame the problem space for AI-consumable models across the OPS area, not only YANG, and to weigh the tools against the gaps. I am contributing to that discussion, but also pointing to related recent efforts to better grasp core problems.

On terminology: I do not think I am introducing new terms. Form / meaning / use is Morris's original semiotic triad, and syntax / semantics / pragmatics are the standard names.

I have taken two of your points: slide 5 now labels the bottom layer "syntax (form)", and since constraints are ontological I have added "axioms / constraints" to the ontology line on slides 5 and 6.

But I'd push back on folding pragmatics into ontology. An ontology, per Gruber, is an explicit specification of a conceptualization: concepts, relations, their axioms or constraints, and the lexicon that names them. That is semantics. Pragmatics is a different axis: how concepts are used in context, who asserted concept or context and how surely, etc. The field keeps these apart deliberately, which is why provenance, named graphs and SKOS scope notes sit around an ontology rather than inside it, and it is where much of the operational value lives. The anomaly work illustrates it: action/reason/trigger is the ontology, while the lifecycle, the scores and the human-in-the-loop are pragmatics.

So a framing that I think meets your instinct: keep "ontology" for the conceptual core, its concepts and relations and constraints and lexicon, as you prefer, and treat it as one component of the whole. The whole is the model's full meaning, its semantics and its pragmatics. I call that whole the "semantic model", using "semantic" in the broad semantic-web sense; I am not wedded to the name.

On reference lexicon versus ontology vocabulary: it's true that lexicon and vocabulary are much the same word. The operative word is "reference". A reference lexicon is a vocabulary designated as an independent, shared anchor that systems bind to, kept deliberately thin: fixed identifiers, names, short definitions and examples, without the relations and axioms of a full ontology. That thinness is the point, since binding N systems to one shared reference scales where N-squared bilateral alignment does not. It supports reconciliation; it does not replace an ontology.

On problem versus solution space: agreed, AI-consumable YANG is largely the problem framing and the others are approaches, and I am happy to state it that way. I take your smaller points on slides 4, 7, 8 and 11.

On the IETF/IRTF boundary: whether and how NMOP and NMRG should collaborate is already a live question, as you know. Thomas has made a good specific suggestion in this exchange that points to a clean split: NMRG carries the ontology and reconciliation research, NMOP contributes the pragmatics and validation. I expect we will progress his proposal (probably narrowing to NMRG and NMOP lists) after the August vacation period.

Updated slides are here: https://github.com/ChrisFJanz/meaning-across-representations/blob/main/semantic_model_deck.pdf

Thanks again for the review and suggestions.

Best

Chris

On Tue, Aug 4, 2026 at 11:43 PM Qin Wu <bill.wu=[e-mail]<mailto:[e-mail]>> wrote:
Hi, Chris:
Thank for kicking off discussion on AI consumable model, the intention of AI Consumable Data Model Side Meeting is to articulate problem space of AI consumable model work in IETF OPS area, not limited to YANG and describe the key challenges and gaps, discuss how these challenges and gaps can be tackled or resolved with tool development, RDF/OWL based knowledge graph/ontology, YANG, or semantics enhancement, etc.
https://mailarchive.ietf.org/arch/msg/nmop/pKOnwWc7eFl5kKk7gWtD2tjCeGc/
If we limit the scope to AI only, ONSEN and NETMOD are not relevant, although we see many proposals in the past to discuss Guidelines for Translation of UML Information Model to YANG Data Model in the NETMOD.
I tend to agree there are more space in NMRG to tackle this kind of topics, especially ontology. But I think Ontology defines a machine-readable vocabulary, taxonomies, and conceptual relationships and acts as the overarching framework that synthesises Semantics, Constraints, and Context into a unified, machine-readable knowledge base. Therefore I have 2 concerns to this semantic model topic discussion:

1.       Such topic discussion makes boundary between IETF and IRTF become blurred.

2.       Introduce more new terms to describe what we have already had. What I would like suggest to stick to use ontology as a basis or umbrella
A few comments on attached slides with my biased humble opinions:
#Slide 2:
I think AI-consumable YANG or AI-consumable Data Model focuses on outlining the problem space, Ontology reconciliation, NAIM, knowledge graphs and annotation, multi-vendor normalisation, the ONSEN YANG↔TMF bridge are potential solutions to address these problems.
In addition, not clear how reference lexicon is different from Ontology vocabulary.

#Slide 4
When you say to create their semantic representations, it can be interpreted in different way:

1.       We can add more semantics in the form Data Model ,which can be consumed by AI such as semantic model defined in draft-ietf-nmop-network-anomaly-semantics

2.       You can create a totally brand new semantic representation, Intent ontology model can be a good example

#Slide 5

1．  Suggest to replace “form” with “syntax”, which is better to explain the relation with semantics

2．  I think ontology should comprise both PRAGMATICS and SEMANTICS, also suggest to replace PRAGMATICS with Context, Constraints, don’t need to introduce new terms to add confusion

3.       Call combination of PRAGMATICS and SEMANTICS as semantic model is a little bit narrow sense.

#Slide 6
1. Ontology should include Context information, Constraints information or PRAGMATICS
2. You can list RDF, OWL, YANG as examples for data model

# Slide 7
I think the commonality of all these cases exepct LIFT is to explore how to do intent translation using Ontology, knowledge Graph, Knowledge base, it is challenging to define common practice for intent translation.
We have many lessons in the past.

#Slide 8
I think NAIM and draft-prabhu-nmrg-prompt-schema-llm share lots of commonability.

#Slide 11
Interesting observation, I like it, I am wondering what lessons can we learnt from the past?

#Slide 13
I think AI Consumable YANG focus on problem space while the other proposals such as ontology reconciliation focus on solution space.

-Qin
发件人: Christopher Janz [mailto:christopher.janz<mailto:christopher.janz>=[e-mail]<mailto:[e-mail]>]
发送时间: 2026年7月31日 21:00
收件人: nmop <[e-mail]<mailto:[e-mail]>>; [e-mail]<mailto:[e-mail]>; [e-mail]<mailto:[e-mail]>; [e-mail]<mailto:[e-mail]>; [e-mail]<mailto:[e-mail]>
主题: [netmod] Semantic meaning across representations and boundaries

All – as I think has been broadly recognized during meetings last week and through various activities and discussion threads both leading up to and following them, a common theme and general problem seems to have emerged, in various ways, within and across NMOP, NETMOD, ONSEN, NMRG, and now as one element of the AIOps discussion.

These multiple discussions and problem contexts point at the same need: for semantic models for network information — meaning made explicit and independent of form – capable of supporting bridging of model and system boundaries and information consumption by machines.

I have put some form, structure and justification around a problem statement and tied it all to the various takes/contexts mooted and discussions in play. In the end I think there is a relatively focused but true research problem to be considered, which I have outlined and would suggest properly belongs in NMRG. For further discussion there … of course such a program would not preclude related activities and outcomes in other WGs, perhaps working against narrower/more specific targets.

Please have a look, and respond with your thoughts – if you wish, on the github repository here: https://github.com/ChrisFJanz/meaning-across-representations ... The deck is there, and also attached here for convenience.

Best

Chris


--
Nmop mailing list -- [e-mail]<mailto:[e-mail]>
To unsubscribe send an email to [e-mail]<mailto:[e-mail]>
_______________________________________________
netmod mailing list -- [e-mail]<mailto:[e-mail]>
To unsubscribe send an email to [e-mail]<mailto:[e-mail]>