W3C home > Mailing lists > Public > public-ontolex@w3.org > May 2015

Re: synsem module ready

From: John P. McCrae <jmccrae@cit-ec.uni-bielefeld.de>
Date: Thu, 21 May 2015 18:29:08 +0200
Message-ID: <CAC5njqqW3Y4jbmTQSATyu6nZMtmJDKWhS_gsFhtakO5WEV0GMg@mail.gmail.com>
To: Fahad Khan <anasfkhan81@gmail.com>
Cc: Manuel Fiorelli <manuel.fiorelli@gmail.com>, Philipp Cimiano <cimiano@cit-ec.uni-bielefeld.de>, "public-ontolex@w3.org" <public-ontolex@w3.org>
Hi Fahad,

This has been bugging me too but for quite different reasons... I will
start by trying to explain the history and reason that we ended up with the
current model of SynSem.

On Thu, May 21, 2015 at 3:52 PM, Fahad Khan <anasfkhan81@gmail.com> wrote:

> Hi everyone,
>
>
> Here are some points for discussion that I've come up with after a wee bit
> of discussion with Francesca. Firstly, I think that to avoid confusion, the
> Class Frame should be renamed Syntactic Frame, since the term "Frame" is
> used in a lot of different contexts -- and is to used to mean different
> things even just within Computational Linguistics -- and so it would pay to
> be about what it is that is being referred to in this case.
>

> Also I feel it would be better to make a distinction between Syntactic and
> Semantic arguments by making them two distinct classes. There are so many
> instances where making this distinction is useful in lexical semantics that
> I think it’s worth the extra complexity to include it in the definition of
> Ontolex itself. For example by making this distinction we can easily
> describe verbal alternations such as occur with verbs such as "to give"as
> in "Alice gave the gift to Bob" and "Alice gave Bob the gift" where two
> different kinds of syntactic structure seem to map to the same semantic
> frame. Moreover, syntactic arguments are usually associated with syntactic
> markers or morphological markings, which is not the case with semantic
> arguments which are instead usually marked with different kinds of
> properties, e.g., semantic roles such as Agent or Theme. Additionally in
> morphologically rich languages like Latin there might not even be a fixed
> syntactic order to the arguments of a verb, whereas in terms of a
> predicative semantic representation the order of the arguments does matter.
> In other words we tend to associate syntactic and semantic arguments with
> different properties, properties which are usually included in
> lexico-semantic representations of lexical entries. ( see also
> http://en.wikipedia.org/wiki/Argument_(linguistics)#Syntactic_vs._semantic_arguments)
>
>
The big problem here is that actually *we don't have semantic arguments and
semantic frames in the model* in anything like the sense of, say, FrameNet.
This is because we have an *ontology* that describes its *semantic frame*,
e.g.,

GivingEvent ≡ ∃ agent.Person ⊓ ∃ theme.Thing ⊓ ∃ recipient.Person

This works fine up until the point that we wish to connect this to a
(syntactic) frame, as *we don't have an identifier for 'the object of the
property agent' to link to*. Let's say we have a simple frame

:give a ontolex:LexicalEntry ;
  synsem:synBehavior [ a synsem:Frame ;
    lexinfo:subject :arg1 ;
    lexinfo:directObject :arg2 ;
    lexinfo:indirectObject :arg3 ] ;
  ontolex:sense :give_sense .
:give_sense ontolex:reference :GivingEvent  .

Now how do we go about representing 'the object of the property *agent', *we
could simple create a structure

:semarg1 a notontolex:SemanticArgument ;
  notontolex:subjOfProp :agent .

And link

:arg1 notontolex:mapsTo :semarg1 .  # Model A

So this 'maps to' property is difficult to name and *doesn't do anything
more than connect two objects*. If all we need is to link the (syntactic)
frame in the lexicon to the semantic frame in the ontology we could easily
do it by attaching the subjOfProp property directly to the (syntactic)
argument. This is fine as the 'semantic argument' is only a proxy for the
argument slot in the ontology... any properties of this slot should be
captured by the ontology's axioms.

:arg1 notontolex:subjOfProp :agent .

There is a hitch here, in that we are now mapping not to a 'giving event'
but to any semantic frame with an agent, theme and recipient. So we can
easily fix this by associate the argument to a sense, e.g., (* Frame to a
sense also works)

:arg1 notontolex:subjOfProp :agent ;
  notontolex:isPartOf :give_sense . # Model B

Now, when we drew up Monnet *lemon* we were mostly interested in the case
where the semantic frame is a single property, e.g.,

:knows a lemon:LexicalEntry ;
  lemon:synBehavior [
    lexinfo:subject :arg1 ;
    lexinfo:directObject : arg2 ] ;
  lemon:sense :knows_sense .

:knows_sense lemon:reference foaf:knows .

:arg1 notlemon:subjOfProp foaf:knows ;
  notlemon:isPartOf :knows_sense .

:arg2 notlemon:objOfProp foaf:knows ;
  notlemon:isPartOf :knows_sense .

We quickly noticed that we are now referring to foaf:knows three times so
we could rationalize this by adding 'subjOfProp' and 'objOfProp' between
the sense and the (syntactic) argument instead of directly linking to the
ontology. This can be thought of as using the sense as a proxy for the true
semantic frame in the ontology. This design decision has carried over to
OntoLex but perhaps it is too much of a trick in the representation. There
are thus three options:

1. Have an explicit semantic argument (Model A). This is certainly in
violation of our principle R3 of 'semantics by reference' and is
considerably more bulky (in the sense of more triples to represent the same
thing).
2. Directly link arguments to ontology and sense (Model B). I do like this
but it is radically different to what we currently have.
3. Use the current proposal (Model C). This is conceptually most different
from other models but also the least bulky.
4. Model D ;).... No seriously... Model D was to add subjOfProp to the
ontology class and thus have SemanticFrame ⊑ owl:Class. It had few fans in
Monnet and probably fewer now.

*I have drawn the four models in the attached picture, that should
hopefully convince you of the wisdom of model C.*

---

Also two things to directly address in your arguments:

1. Arguments can be members of multiple (syntactic) frames, so you can have
the same argument as both a prepositional object of one frame and an
indirect object of another. Technically it should make no difference
whether one URI or two are used. [Actually we have previously said that
your example of give has a special syntactic role *indirectObject_to* that
is either realized with the indirect object or a PP with 'to']
2. The synArg/semArg properties do not have order... they work just as well
for fixed and non-fixed word order languages.


> Synsem/Example 4 is unclear/a bad example since the phrase "opening film"
> can either be associated with a prepositional phrase argument headed by
> "in" (as in "the opening film in our programme tonight is...") or "of" (as
> in "the opening film of the Cannes film festival was terrible this year")
>  -- or even no prepositional phrase argument (as in "the opening film is
> terrible")  -- and so creating a lexical entry specifically based on
> "opening film at" seems inefficient (why create a separate lexical entry
> for each of the cases, and not associate an opening_film lexical entry with
> a number of prefered prepositions?). Maybe it would be better to have a
> simpler example of a relational noun like "brother" or "uncle" where the
> noun is strongly associated with a certain preposition  (in these examples
> the preposition  “of”) rather than a number of different prepositions, when
> it is expressing a related ontological Object Property.
>
> Also associating the concept Alma mater in Ex synsem/example5 with a
> lexical entry other than the lexical entry "alma mater" comes across as a
> little strange -- especially since it's not at all clear that the
> expression alma mater always and unproblematically refers to an institution
> that someone has graduated from (wikipedia, different dictionaries disagree
> on this) rather than just attended.
>
> Overall, these examples where an ontological concept is associated with a
> related lexical entry not because there’s a straightforward
> reference/extensional relation between them but because there is some kind
> of lexical entailment relation involved (e.g., the ontological concept
> refers to the result of the action described by a word/lexical entry) are
> problematic because they tend to make the "meaning" of ontolex:reference a
> bit obscure: do we use the reference relation to say that the extension of
> the word is such and such, or that a certain expression is often used in a
> language to refer to an important aspect of an event or a concept?
>
Mapping between words and ontology entities is not one-to-one, an ontology
predicate can be expressed by many different terms and a word has different
senses which resolve to different ontology predicates. Other frames or
other meanings of a term do not stop a single sense (mapping) being valid.

However, if you have better suggestions for examples, I could see there
being words that express the points more clearly.

>
> Also, I agree with John the gestalt stuff in the definition of Semantic
> Frame is a bit puzzling.
>
:)

Regards,
John

>
> Cheers,
> Fahad
>
> On 20 May 2015 at 16:57, John P. McCrae <jmccrae@cit-ec.uni-bielefeld.de>
> wrote:
>
>> Hi,
>>
>> I read through the spec and there are a few major issues I detected in
>> the first couple of sections sections (ontolex + synsem)
>>
>> 1. The definition of 'other form' still says '[Other form] should be ....
>> an abbreviation, short form or acronym'. This is incorrect and contradicts
>> the definition of lexical entry. Can we add an example clarifying the
>> representation of abbreviations?
>> 2. ontolex/example10 now doesn't make any sense... 'bank' is just two
>> words each with a different meaning. Can we change this to a word with
>> genuine polysemy... I suggest 'troll' (1
>> <https://en.wikipedia.org/wiki/Troll>, 2
>> <https://en.wikipedia.org/wiki/Internet_troll>).
>> 3. ontolex/example17 doesn't really show a lot and for some reason refers
>> to IATE for 'cat'!? (this is probably my fault...). Could we switch it to
>> 'spouse'/'marry' showing that these two lexical entries have two different
>> concepts but the same reference dbpedia:spouse
>> 4. The definition of semantic frame is at best confusing, I really don't
>> think we need to bring Gestalt Theory into this as well. My attempt would
>> be:
>>
>> *Semantic Frames* are the meaning of a word (and hence are also lexical
>> senses) but expressed by one or more ontological predicates and their
>> arguments. This sense of the word can only be understood when all of its
>> required arguments are realized.
>>
>> Similarly we need to change subframe to
>>
>> *Subframe *relates a complex semantic frame to frames for each of the
>> individual ontological predicates that form the complex semantic frame.
>>
>> 5. synsem/example5 and example6 are essentially the same as example4 but
>> they connect an eventive verb ('graduate' or 'die') with a consequential
>> fact ('almaMater' or 'deathYear'). This is of questionable soundness
>> although we have argued in papers it is valid when the event and the
>> consequence are in a strict bijection... still, I would prefer to drop this
>> for the spec as it adds a lot of unnecessary complexity.
>>
>> There are a lot of other minor issues I will change directly in the spec.
>>
>> Regards,
>> John
>>
>> On Mon, May 18, 2015 at 10:41 AM, Manuel Fiorelli <
>> manuel.fiorelli@gmail.com> wrote:
>>
>>> Dear Philipp, All
>>>
>>> you can find my comments on the synsem module below.
>>>
>>> In Example synsem/example2, the resource :own_frame_transitive is
>>> wrongly written :own_form_transitive.  Additionally, there are two
>>> usages of owl:subPropertyOf, which instead  should be rdfs:subPropertyOf
>>> .
>>>
>>> The class synsem:SemanticFrame is declared to be subclass of
>>> ontolex:LexicalSense; however, in the picture representing the synsem
>>> module, the arrow representing this axiom is oriented in the opposite
>>> direction.
>>>
>>> In the paragraph "Semantic Frames", there is a table headed "Type",
>>> "Predicate", "Example", whose first row contains *City(x)*, ?x rdf:type
>>> ontology:Person: should it be ?x rdf:type dbpedia-owl:City?
>>>
>>> There is no example (just below the definition of synsem:isA)about the
>>> representation of unary predicates. Nor is there any example about the
>>> representation of individuals.
>>>
>>> The definitions of synsem:{subj|obj}OfProp use the following wording:
>>> "...property represents the semantic argument with represents"
>>>
>>> I would avoid a sequence of two "represents". Moreover, I think that
>>> "with" should be "that".
>>>
>>> In Example synsem/example3, there is again owl:subPropertyOf.
>>>
>>> Also, In Example synsem/example4, there is again owl:subPropertyOf.
>>>
>>> In the section "Complex Senses / Semantic Frames", there is the
>>> definition of synsem:subframe, while in the figure there is the
>>> property synsem:subsense.
>>>
>>> Just below Example synsem/example7, there is an example involving the
>>> property father: the property should point to the child; however, the
>>> name of the property suggests to me that the object is the father (just in
>>> the same manner skos:broader points to the broader of a given concept).
>>>
>>> I think that Example synsem/example9  should be explained in more detail.
>>>
>>> I didn't find the definition of synsem:propertyDomain and
>>> synsem:propertyRange; then, I realized that they were moved to the core
>>> module. The diagram of the core module must be updated to include these
>>> properties, as well as the diagram of the synsem module to remove them.
>>>
>>> I noticed that in the infobox providing the definition of propertyRange
>>> and propertyDomain, the URI still uses the synsem namespace instead of the
>>> core ontolex namespace.
>>>
>>> Finally, I noticed a typo in the definition of ontolex:LexicalEntry:
>>> "The class lexical entry represents a unit of analysis of the lexicon
>>> that consist of a set of forms that are grammatically ... "
>>>
>>> It should be "that consists" with an append "s".
>>>
>>> Best Regards
>>>
>>> Manuel Fiorelli
>>>
>>>
>>> 2015-05-13 21:47 GMT+02:00 Philipp Cimiano <
>>> cimiano@cit-ec.uni-bielefeld.de>:
>>>
>>>> Dear all,
>>>>
>>>>  I have been working on finalizing the synsem module, please check:
>>>>
>>>>
>>>> https://www.w3.org/community/ontolex/wiki/Final_Model_Specification#Syntax_and_Semantics_.28synsem.29
>>>>
>>>> The next telco to discuss the synsem module will be on Friday the 22nd
>>>> of Mai, 16:00 CET.
>>>>
>>>> Please send me any issues to discuss or comments on the synsem module
>>>> by Thurday 21st of Mai at the very latest.
>>>>
>>>> Thanks and best regards,
>>>>
>>>> Philipp.
>>>>
>>>> --
>>>> --
>>>> Prof. Dr. Philipp Cimiano
>>>> AG Semantic Computing
>>>> Exzellenzcluster für Cognitive Interaction Technology (CITEC)
>>>> Universität Bielefeld
>>>>
>>>> Tel: +49 521 106 12249
>>>> Fax: +49 521 106 6560
>>>> Mail: cimiano@cit-ec.uni-bielefeld.de
>>>>
>>>> Office CITEC-2.307
>>>> Universitätsstr. 21-25
>>>> 33615 Bielefeld, NRW
>>>> Germany
>>>>
>>>>
>>>>
>>>
>>
>

SynSem_models.png
(image/png attachment: SynSem_models.png)

Received on Thursday, 21 May 2015 16:29:38 UTC

This archive was generated by hypermail 2.4.0 : Friday, 17 January 2020 16:36:49 UTC