I have been thinking about grounding a lot in my conversations with clients, and I really like the distinction you are making between grounding as retrieval and grounding as maintaining coherent state across identity, time, space, and meaning.
In the enterprise, though, coherence alone is not enough.
An AI system also needs to know which interpretation is authoritative, which version applies, who approved it, under what conditions it is valid, and what actions are permitted.
A system can retrieve the right document and maintain the right state, yet still make the wrong decision because it applied the wrong definition, policy, jurisdiction, or organizational context.
Within the enterprise, I would add two more dimensions to your “at least four” grounding spectrum:
1. Normative grounding
What is allowed?
What is required?
What is preferred?
What is prohibited?
Under which conditions?
2. Authoritative grounding
Which definition governs this use case?
Which authority approved it?
Which version applies?
What should happen when sources conflict?
I tend to group these two dimensions under the idea of authoritative context: the governed context that tells an AI system not only what information means, but how it should be interpreted and used.
The first falls under boundaries (what are the shacl restrictions that determine whether actions can be undertaken or not). The second is provenance, which is related to temporal event management. Agreed on both - they are part of the grounding, and definitely need to be there.
agree with the first but the second i'm a bit on the fence because provenance is a bit narrow for me. I would see authority as depending on provenance and temporal context, but also on applicability, precedence, mandate, and conflict-resolution rules.
If you can capture it, yes. precedence is not THAT hard to capture, but you need to have a reasonably comprehensive dataset to establish it - (cf. Conflict-Resolution, of which precedence is the superclass). Mandate ultimately comes down to ascertaining restrictions. Conlict-resolution rules fall into the same category as precedence, in that they both determine operational priority, and usually resolve to ranking of named graph query patterns. Applicability is the odd man (odd, pattern, anyway) out - and comes down to determining meta-patterns (metaphors).
It’s in French, but here is the translation of part of the introduction:
"In our work, we consider the task of automatic processing aimed at building, from
a corpus of road accident reports, interpretations compatible with these
last ones, and to offer illustrations in the form of sequences of still images...
In the first part of the article, we present the overall architecture of the computer system. In the second part, we propose elements of analysis for an operative solution to the
the articulation of lexical and grammatical meanings, in the form of text segmentation in different spatio-temporal phases. In the third part, we present a modeling of traffic areas and vehicle movement, ensuring the transition to the image."
I have been thinking about grounding a lot in my conversations with clients, and I really like the distinction you are making between grounding as retrieval and grounding as maintaining coherent state across identity, time, space, and meaning.
In the enterprise, though, coherence alone is not enough.
An AI system also needs to know which interpretation is authoritative, which version applies, who approved it, under what conditions it is valid, and what actions are permitted.
A system can retrieve the right document and maintain the right state, yet still make the wrong decision because it applied the wrong definition, policy, jurisdiction, or organizational context.
Within the enterprise, I would add two more dimensions to your “at least four” grounding spectrum:
1. Normative grounding
What is allowed?
What is required?
What is preferred?
What is prohibited?
Under which conditions?
2. Authoritative grounding
Which definition governs this use case?
Which authority approved it?
Which version applies?
What should happen when sources conflict?
I tend to group these two dimensions under the idea of authoritative context: the governed context that tells an AI system not only what information means, but how it should be interpreted and used.
The first falls under boundaries (what are the shacl restrictions that determine whether actions can be undertaken or not). The second is provenance, which is related to temporal event management. Agreed on both - they are part of the grounding, and definitely need to be there.
agree with the first but the second i'm a bit on the fence because provenance is a bit narrow for me. I would see authority as depending on provenance and temporal context, but also on applicability, precedence, mandate, and conflict-resolution rules.
If you can capture it, yes. precedence is not THAT hard to capture, but you need to have a reasonably comprehensive dataset to establish it - (cf. Conflict-Resolution, of which precedence is the superclass). Mandate ultimately comes down to ascertaining restrictions. Conlict-resolution rules fall into the same category as precedence, in that they both determine operational priority, and usually resolve to ranking of named graph query patterns. Applicability is the odd man (odd, pattern, anyway) out - and comes down to determining meta-patterns (metaphors).
Interesting to think about.
These reflections evoke for me both the « world models » imagined by Yann Le Cun in particular and the epistemological reflection of Edgar Morin (cf. https://www.linkedin.com/posts/marc-henri-hurt-5b010346_miniontolandenmhhurt-activity-7467226708740370432-YOQz ), but I do not know the latest propositions of Yann Le Cun, and I have not reread Morin.
But the paragraph beginning with "Spatial grounding..." made me think of a very old study to which here is the link : https://talnarchives.atala.org/TALN/TALN-1997/taln-1997-court-001.pdf .
It’s in French, but here is the translation of part of the introduction:
"In our work, we consider the task of automatic processing aimed at building, from
a corpus of road accident reports, interpretations compatible with these
last ones, and to offer illustrations in the form of sequences of still images...
In the first part of the article, we present the overall architecture of the computer system. In the second part, we propose elements of analysis for an operative solution to the
the articulation of lexical and grammatical meanings, in the form of text segmentation in different spatio-temporal phases. In the third part, we present a modeling of traffic areas and vehicle movement, ensuring the transition to the image."