Re: Issues about the semantics of the ontology-lexicon interface [was: Re: Why not to shortcut the "sense" object]

Thx Guido, this discussion is very useful (provided that we do not get into the infamous "sumo-threads" where each discussion used to get eventually to discussing the nature of matter and life :)).

On Oct 15, 2012, at 2:37 PM, Guido Vetere <gvetere@it.ibm.com> wrote:

> Aldo, Armando, 
> 
> A couple of things about what you said (on the rest, I generally agree). 
> 
> As for the name of the arrow (property?) linking senses and concepts, Aldo is right, maybe 'characterize' is not appropriate in this context (indeed, the notion comes from mathematics) and is not likely to be accepted by the community. But 'representedBy', if read from left to right (a sense is represented by a concept), could be even worse, since, in the mainstream of western semiotics, signs represent things and stand for them (aliquid pro aliquo), and not the other way around. Maybe we could adopt the classic (e.g. Odgen-Richard) 'refers to', even if the binding with the 'referential function' may be inappropriate. It looks like a trivial naming detail, but it may have an impact on the way people grasp the intended meaning of the model. 

The reason why I like "representedBy", despite its generic ambiguity, is that I see ontology entities firstly as constants from a vocabulary. As constants, they can perfectly "represent" senses. Indeed, this is quite inline even with formal ontology and logic (cf. Nicola Guarino's 2003 paper on conceptualizations).
Of course, constants of a vocabulary get a *formal* meaning/interpretation that is based on model theory, but this is another story, which gives us room to claim that lexical entries can have a (formal) semantics with ontology entities.
In other words, the way the ontology-lexicon interface works seems to be the following:

- a lexical entry has some sense (either local or general/conventional), which we can call "lexical meaning"
- a sense can get ("be represented by", or "be expressed by") a constant (ontology entity) in a formal vocabulary
- that constant has a formal interpretation provided by logical and domain axioms: this is a "formal meaning"

Unfortunately, logicians have substantially identified intension (which is the closest relative to lexical meaning) with the constants of a vocabulary. Therefore, the only original, operational, and useful semantic stuff that we have from logical models is extensional meaning. But we are not going to talk about that as well, right? ;). Since we are not doing that, ontology entities from OWL/RDF will be inevitably ambiguous, and depending on context, sometimes they can be considered as constants, and sometimes as meanings.

> 
> This leads to the more basic question about the logic nature of this relation, i.e. of what kind of logical things fill the pattern: Lexical unit --meaning--> Sense --refers to--> Ontological concept. If we give this graph a DL interpretation, as I tried to do, nodes could be first order unary predicates and arrows (restricted) first order binary predicates. In this reading, instances of Sense (e.g. cat#1) would be related to instances of Concepts (e.g. my cat). Aldo suggests that this model would be in conflict with the intuition that cat#1 may in many cases refer to cats in general, i.e. the whole class of cats. However, 'class vs instance' ('intensional' vs 'extensional', if you whish) is part of the systematic polysemy for many senses, if not for senses in general. Dictionary developers might want to use the same sense of 'cat' both for 'the cat is on the mat' and 'the cat is a feline'. Now, it is true that an axiom of the form cat#1 TYPE (Sense AND refersTo ONLY Cat) would not capture the intensional reading of the sense, but, conversely, setting 'refers to' to range on class names, as Aldo suggests, would not capture the extensional one.

Maybe there is a misunderstanding here. When I read your "cat#1" I'm interpreting it as a sense of the word "cat", not as a particular cat. 
Now, if I interpret you right, cat#1 would be a Sense that is represented by some OntologyEntity. 
On the contrary, if you mean a particular cat, I'm not following you anymore: why a cat should be a Sense?

> 
> In general, using class names as values for the property in question, e.g. by using OWL 2 punning, raises the question of providing the property with some extra formal semantics, since punning, as you know, is just a syntactic trick. As Aldo says, problems like this have been tackled by other specifications already, such as SKOS.  However, we here face the problem of dealing with any legacy ontology, which rely on standard set-theoretic semantics, instead of 'ad hoc' conceptual frameworks. Thus, we should come up with a model that preserves both the intended formal meaning of standard ontologies and the complexity of linguistic signification, which is not an easy task, and cannot be pursued just by naming conventions. 

You're right in general, but I think that this is too much for this Community Group: after all, we do not want to solve the harsh problems of higher-order logics applied to natural language semantics.
Anyway, punning is not much a trick (despite its name), but a regular logical way of interpreting constants in a theory by partitioning their interpretations. The fact that those interpretations do not interact as in a rocketing HOL is simply due to the limitations we accept for having a Web Ontology Language, which is in addition considered way too expressive …

> 
> In my opinion, much depends on what 'Sense' represents in our basic pattern. I understand well, this concept is currently associated to either definitions in dictionaries or synsets in wordnets, thus being a mostly lexicographic notion. A different ontology could model Sense as a class of socially constructed abstractions evoked in linguistic acts, independent from dictionaries and wordnets. In the former case, Sense could be a leaf class, and what we link through arrows are instances. In latter case, I think that 'Sense' should rather be the root of a class hierarchy, and what we link to lexemes should be Sense's subclasses, whose instances, in turn, represent meanings in their textual occurrences.  By the way, Senso Comune embraces an ontology like this.  So a good question to start with would be: what do we mean when we say 'Sense'? 

My impression is that we cannot (and shouldn't in my opinion) attempt to solve that kind of issues; on the contrary, it's very useful to abstract out of them. 
A sense can be profitably (and yes, ambiguously) figured out as any conceptualization associated with a lexical entry, be it an entry, a definition/gloss, an ID, a paraphrase, a reference to some other disambiguating source, or even (please do not shoot me) formal meanings and cognitive objects studied by neurolinguistics. 
In the particular community of linked data and the semantic web, we can refrain from discussing too much what a sense is, and begin to see how interesting are the links emerging out of those apparently different creatures.

Ciao
Aldo

> 
> Cheers, 
> 
> Guido Vetere
> Manager, Center for Advanced Studies IBM Italia
> _________________________________________________
> Rome                                     Trento
> Via Sciangai 53                       Via Sommarive 18
> 00144 Roma, Italy                   38123 Povo in Trento, Italy
> +39 (0)6 59662137                 +39 (0)461 312312
> 
> Mobile: +39 3357454658
> _________________________________________________ 
> 
> 
> Aldo Gangemi <aldo.gangemi@cnr.it>
> 13/10/2012 14:40
> 
> To
> public-ontolex <public-ontolex@w3.org>
> cc
> Aldo Gangemi <aldo.gangemi@cnr.it>, John McCrae <jmccrae@cit-ec.uni-bielefeld.de>, Armando Stellato <stellato@info.uniroma2.it>, Guido Vetere/Italy/IBM@IBMIT, Philipp Cimiano <cimiano@cit-ec.uni-bielefeld.de>
> Subject
> Issues about the semantics of the ontology-lexicon interface [was: Re: Why not to shortcut the "sense" object]
> 
> 
> 
> 
> 
> Hi all, I lagged behind in the last month, because of my recent installation in Paris. Yesterday I was traveling back from Galway (EKAW) and couldn't attend, apologies for that. 
> I have followed the recent discussion, and that's my contribution. I have renamed the thread, because it is now spanning over different topics related to the semantics ig the O-L interface. 
> 
> ---Senses--- 
> Concerning Philipp's summary, firstly I agree with the decision (?not yet approved, it seems?) of creating the intermediate Sense class: it's obviously needed, either for making room for lexical senses (definitely to be distinguished from ontology entities), or to be able to talk about senses (reifications of the meaning function). 
> Concerning the name, I vote for "sense", because sememes, acceptations, and others, are either very technical for the layman, or even wrong, as Philipp reminds us about the original notion of sememe. The only real alternative would be "meaning", but I'd rather keep that term for the top-level class of a meaning taxonomy, as I suggest in the following. 
> 
> In a previous mail, I proposed to consider also an additional solution, i.e. to create a taxonomy of meanings, which has ontology entities (as formal semantic objects) and lexical senses as special subclasses. The two solutions are compatible, and if we realize that a meaning taxonomy might be useful, it can be introduced anyway. 
> Think of the sense-synset issue raised by Philipp: I agree that synsets are not lexical senses, if we assume that a lexical sense should be expressed by only one lexical unit (cardinality exactly 1), but still they are senses, and it's completely reasonable to put synsets (as well as many other creatures of lexical semantics, including sememes, acceptations, frames, semantic verb classes, etc.) in a meaning taxonomy. 
> 
> Concerning the property names, I'm ok with both LexicalEntry – meaning –> Sense, and with Sense – representedBy –> OntologyEntity. 
> Maybe we could get rid of multiple related uses of the "mean" notion, which can be somehow disturbing: Meaning as a class, meaning as a property between lexical entries and senses, means as a property between lexical entries and ontology entities … it may look like we are playing with words … what about following the conventional naming patterns that employs the name of the property range? E.g. LexicalEntry – sense –> Sense ; LexicalEntry – meansOntologyEntity –> OntologyEntity. The advantage of using this apparently redundant naming is that at the instance level, the triple become very clear, e.g. Saxophone – sense –> wordsense-saxophone-1 ; Saxophone – hasOntologyEntity –> music:Saxophone. 
> I also prefer "representedBy" to "characterizes", because the second is very generic and not attested in any related literature. 
> 
> ---Property chaining over senses--- 
> Secondly, I agree with the decision to add a property chain in the model, which helps resolving the indirection produced by the Sense class: this is a good practice (a logical design pattern), used in many contexts. I do not see room for John's criticism about it: it does not increase the cognitive complexity (on the contrary, it facilitates the use of the model for those reluctant to catch on the sense-ontology-entity distinction), and the added computational complexity only holds when a DL reasoner materializes the ABox. 
> One mild problem here might be that we are making slightly different assumptions when we name "representedBy" the property between senses and ontology entities, but "means" the property between lexical entries and ontology entities. Since we do not have a rich axiomatization behind these names, we might be pragmatic and ignore the problem, however I deem important to justify it a little bit in the documentation. In practice, this approach seems to suggest that senses are actually "represented" by ontology entities, and this is clear and intuitive. It also suggests that lexical entries actually "mean" ontology entities, but this is far less clear and intuitive, since in no obvious way words mean stuff in ontologies … it's much better to say that words have conceptualizations that are represented in ontologies. Indeed this is the way we talk of lexical senses :). That's why my above suggestion was "hasOntologyEntity", which however I admit ti be too generic. In principle, the compositional name that best fits the property chain would be "hasSenseRepresentedByOntologyEntity", but it's way too long, specially for those willing to use that property as a shortcut. Other suggestions? 
> 
> ---GCIs on ontology hierarchies--- 
> Finally, a comment about Guido's observation that "cat#1 INSTANCEOF (Sense AND characterizes ONLY Animal)" is the right formalization for an example of the representedBy object property values. If I understand well, here we have two important issues. The first one can be solved by using OWL2, the second poses a more difficult challenge. 
> For the first issue, I think that Guido talks about OWL1, but anyway that axiom would give us a misinterpretation, because it would tell us that cat#1 is a sense that can only be represented by *individuals* from the class Animal, which is not what Guido wants I guess. This problem was described in detail by W3C SWBPD committee in 2004, and eventually some OWL1 solutions were recommended in the "Classes as values" design pattern. However, in OWL2 (lucky us) punning makes our lives easier, and a simple (partial!) solution is (in Manchester syntax) "cat#1 TYPES (Sense AND representedBy VALUE Animal)". 
> For the second issue, Guido points out that there are cases in which we need to refer to generic subclasses of an ontology entity (if it's a class): this cannot be expressed in OWL at all, since we cannot use the OWL vocabulary in the position for the domain vocabulary, In other words, the following is a wrong axiom even in OWL2: "cat#1 TYPES (Sense AND representedBy (subClassOf VALUE Animal)". 
> A viable design pattern is to create a property for meaning hierarchies, in the vein of skos:broader or wordnet:hypernym, so that we could declare e.g.: "cat#1 TYPES (Sense AND representedBy ([skos:broader] VALUE Animal)". 
> However, a property like skos:broader typically applies to concepts, and senses would probably be compatible. Much less are ontology entities compatible, even though SKOS seems to suggest a loose correspondence between concepts and rdfs/owl classes. In particular, we should materialize ontology class hierarchies as skos:broader hierarchies in order to reason over these constructs. 
> Another design pattern might resort to a specialized property, such as "broadlyRepresentedBy", e.g.: "cat#1 TYPES (Sense AND broadlyRepresentedBy VALUE Animal)". "broadlyRepresentedBy" can be a super property of representedBy. Of course, with this second pattern, we would lose the sophisticated DL reasoning that one can get with the first. Nonetheless, the second seems more practical and simple to apply for different levels of expertise. 
> 
> Ciao 
> Aldo 
> 
> _____________________________________ 
> 
> Aldo Gangemi 
> Senior Researcher 
> Semantic Technology Lab (STLab) 
> Institute for Cognitive Science and Technology, 
> National Research Council (ISTC-CNR) 
> Via Nomentana 56, 00161, Roma, Italy 
> Tel: +390644161535 
> Fax: +390644161513 
> aldo.gangemi@cnr.it 
> http://www.stlab.istc.cnr.it 
> http://www.istc.cnr.it/people/aldo-gangemi 
> skype aldogangemi 
> okkam ID: http://www.okkam.org/entity/ok200707031186131660596 
> 
> On Oct 12, 2012, at 6:55 PM, John McCrae <jmccrae@cit-ec.uni-bielefeld.de> wrote: 
> 
> On Fri, Oct 12, 2012 at 6:35 PM, Armando Stellato <stellato@info.uniroma2.it> wrote:
> From what I got, and hope not to be wrong (it’s useful also for me to clarify as I missed a couple of calls on September), OntologyEntity is a generic rdf:Resource of one of the main entities in the main vocabularies (aka: OWL and SKOS, thus: property, class, individual, skos concept…).
> 
> Another question to John from my side: from your email it seemed to be against stating the propertyChain axiom on (means, <meaning,representedBy>) implying that the direct Entry ---means--> OntologyEntity from "Lexical Entry -> meaning -> Sense -> representedBy -> OntologyEntity"  but then the sentence: “Here the difference is 1 named elements vs. 3 named elements, but as stated above, at least half of users (data consumers) will have to understand all 4 names...” instilled some doubt in my interpretation…
> 
>   
> Are you voting against the larger structure as a whole (thus keeping only the Entry ---means--> OntologyEntity structure), or against the propertyChain axiom? I really got the second, though I’m not even sure how adding the p.chain axiom (or not doing it) would change anything for the user or consumer. I’m sure I’m missing something, so sorry in advance for my potential misinterpretation.
> 
> Sorry it isn't clear: the long chain is TBMK agreed upon (Lexical Entry -> meaning -> Sense -> representedBy -> OntologyEntity)*... we are questioning whether we need the short chain (Entry ---means--> OntologyEntity) as well. I say it is not worth it. 
> 
> Regards, 
> John 
> 
> * or (Word -> sense -> Sememe/Acceptation -> characterizes -> rdf:Resource/skos:Concept/owl:Entity) or some combination of these terms. 
>   
> Have a nice we!
> 
>   
> Armando
> 
>   
>   
> From: Guido Vetere [mailto:gvetere@it.ibm.com] 
> Sent: Friday, October 12, 2012 6:08 PM
> To: public-ontolex
> Subject: Re: Why not to shortcut the "sense" object
> 
>   
> All, 
> 
> I apologize for missing the call today. Here just some short remark. 
> 
> "Entry ---means--> OntologyEntity" means that if you want to predicate on the meaning relationship (e.g. to associate some grammatical constraint) you have to resort on a meta predicates (e.g. OWL Annotations). 
> 
> "Lexical Entry -> meaning -> Sense -> representedBy -> OntologyEntity" sounds good, but instead of 'representedBy' I would say 'characterizes' or something alike, meaning that a linguistic sense gives a (cultural) shape to an entity. Moreover, it is not clear to me (maybe you discussed about that) whether OntologyEntity is a first order TOP concept (e.g. equivalent to OWL Thing). In this case, note that in order to tell that the instance of Sense 'cat#1' (i.e. the first sense of the lemma 'cat') represents an Animal, you have to write something like: 
> 
> cat#1 INSTANCEOF (Sense AND characterizes ONLY Animal). 
> 
> Is it correct? 
> 
> If there is something that I can do, please let me know. 
> 
> Regards, 
> 
> Guido Vetere
> Manager, Center for Advanced Studies IBM Italia
> _________________________________________________
> Rome                                     Trento
> Via Sciangai 53                       Via Sommarive 18
> 00144 Roma, Italy                   38123 Povo in Trento, Italy
> +39 (0)6 59662137                 +39 (0)461 312312
> 
> Mobile: +39 3357454658
> _________________________________________________ 
> 
> John McCrae <jmccrae@cit-ec.uni-bielefeld.de> 
> Sent by: johnmccrae@gmail.com
> 12/10/2012 16:35
> 
> 
> To
> public-ontolex <public-ontolex@w3.org>
> cc
> Subject
> Why not to shortcut the "sense" object
>  
> 
> 
> 
> 
> 
> 
> 
> Hi all, 
> 
> As discussed today in the telco there is a proposal to introduce a shortcut replacing "Entry ---sense--> Sense ---representedBy--> OntologyEntity" with "Entry ---means--> OntologyEntity", while this is theory sounds good, I contend that in practice it is not worth the effort. (This is based on practical experience with the lemon model).
> 
> It does not make the model easier to use: It is clear that for data producers this proposal simplifies the matter (as less links and URIs are required), however for data consumers it complicates the models (as they need to understand both methods of linking and be able to infer equivalence between the two methods). Thus, if EaseOfUse = (% of Consumers) × EaseOfUse(Consumer) + (% of Producers) × EaseOfUse(Producer), hence if we assume there will be approx. as many producers as consumer then we need only ask is it worth "is the extra effort for the producer less than that for the consumer", i.e., "would you rather implement a system that infers similarity across multiple representations, or use extra links and URIs"?
> It does not make the model easier to understand: While, I understand that the sense object is nebulous and difficult per se to understand, I would still argue that the clearest measure of how easy to understand a model is, is the number of named elements it has (as many users may not need to deeply understand the meaning of a sense, but be happy to know that "translation", "antonymy" and "register" go there). Here the difference is 1 named elements vs. 3 named elements, but as stated above, at least half of users (data consumers) will have to understand all 4 names... if we assume out of the producers 70% do not need to represent senses (and thus any associated properties, "translation", "antonymy", "register") then the average number of links a user will need to understand is 4 × 0.5 + 3 × 0.5 × 0.3 + 1 × 0.5 × 0.7 = 2.8... so it makes the model all of 7% easier to understand! Worse, this figure is overgenerous as: I expect there to more data consumers than producers and I expect at least 50% of users to require sense modelling.
> Regards, 
> John 
> 
> IBM Italia S.p.A.
> Sede Legale: Circonvallazione Idroscalo - 20090 Segrate (MI) 
> Cap. Soc. euro 347.256.998,80
> C. F. e Reg. Imprese MI 01442240030 - Partita IVA 10914660153
> Società con unico azionista
> Società soggetta all’attività di direzione e coordinamento di International Business Machines Corporation
> 
> (Salvo che sia diversamente indicato sopra / Unless stated otherwise above)
> 
> 
> 
> IBM Italia S.p.A.
> Sede Legale: Circonvallazione Idroscalo - 20090 Segrate (MI) 
> Cap. Soc. euro 347.256.998,80
> C. F. e Reg. Imprese MI 01442240030 - Partita IVA 10914660153
> Società con unico azionista
> Società soggetta all’attività di direzione e coordinamento di International Business Machines Corporation
> 
> (Salvo che sia diversamente indicato sopra / Unless stated otherwise above)
>