- From: Dave Lewis <dave.lewis@cs.tcd.ie>
- Date: Tue, 27 May 2014 02:23:04 +0100
- To: Jorge Gracia <jgracia@fi.upm.es>
- CC: Roberto Navigli <navigli@di.uniroma1.it>, Tiziano Flati <tiziano.flati@gmail.com>, lider <lider@delicias.dia.fi.upm.es>, "public-bpmlod@w3.org" <public-bpmlod@w3.org>, public-ontolex@w3.org
- Message-ID: <5383E8F8.7090208@cs.tcd.ie>
Hi Jorge, guys,
Thanks for these pointers, I had not been following this as closely as I
should, so I have some comment below that are relevant to both the
meta-share RDF model and your translation model in ontolex, so I've
copied them also.
You are quite correct to reify the translation relationship. Deriving an
authoritative translation is rarely straighforward and may involve
different inputs at different times from different sources, e.g.
babelnet has professionally curated translation, translations from
wikipedia and MT oututs.
So in many cases you are dealing with the current status of a
provisional translations rather than 'final' authoritative.
Also, there is some potential confusion in naming the reifying class
'Translation' since in many situations this refers to the string in the
targt language rather than the entity linking a target language string
to a source language string.
In [1] we proposed an approach to handle this by specilising from the
W3C Provenance vocubulary [2].
This means treating the source and targets of translation (LexicalEntry,
LexicalSense) as prov:Entity classes so that their provenance can be
tracked using other classes and proerties from that model.
Specifically we propose specialising the provenance property:
http://www.w3.org/TR/prov-o/#wasDerivedFrom
i.e.
ontolexTrans:wasTranslatedFrom rdfs:subPropertyOf
prov:wasDerivedFrom.
PROV-O also enables reification by defining a class:
http://www.w3.org/TR/prov-o/#Derivation
which is in the range of:
http://www.w3.org/TR/prov-o/#qualifiedDerivation
So similarly we can define
ontolexTrans:Translation rdfs:subClassOf prov:Derivation.
and
ontolexTrans:qualifiedTranslation rdfs:subPropertyOf
prov:qualifiedDerivation.
To flesh this out with an example:
ex:34678en a lemon:LexicalEntry;
a prov:Entity;
lemon:form [ lemon:writtenRep "house"@en ] .
ex:34678es a lemon:LexicalEntry;
a prov:Entity;
lemon:form [ lemon:writtenRep "casa"@es ];
ex:34678es ontolexTrans:wasTranslatedFrom ex:34678en;
its:mtConfidence "0.5";
ontolexTrans:qualifiedTranslation [
a ontolex:Translation;
prov:hadActivity ex:ExMachineTranslation;
].
Note in the above the its:mtConfidence is more accurately used to
annotate the LexicalEntry rather than the Translation, as it is a
property of the text resulting from the translation, rather than a
reification of the translation.
Thoughts welcome.
cheers,
Dave
[1] http://www.lrec-conf.org/proceedings/lrec2012/pdf/636_Paper.pdf
[2] http://www.w3.org/TR/prov-o/
On 23/05/2014 14:48, Jorge Gracia wrote:
> Dear Tiziano, Roberto
>
> You could also consider using the lemon translation module to
> represent explicit translations as linked data. This is currently
> under development in the ONTOLEX group but there is a lemon-based
> version already available, that I will present at LREC next week [1].
> The idea is reifying the translation relation so you can attach
> additional information to it (source, target, confidence, provenance,
> etc.) [2]
>
> Regards,
>
> Jorge
>
> [1]
> http://ra.cps.unizar.es:8080/PUBLICATIONS/attachedFiles/document/LREC2014_translations_V11.pdf
> [2] http://purl.org/net/translation#
>
>
>
>
> 2014-05-23 11:58 GMT+02:00 Dave Lewis <dave.lewis@cs.tcd.ie
> <mailto:dave.lewis@cs.tcd.ie>>:
>
> Roberto, Tiziano,
> Thanks for that.
>
> Have you considered already how you might allow third parties to
> QA and perhaps correct those translations? That is, some sort of
> process by which proposed MT translations between senses can be
> promoted to more authoritative, human checked translations, and
> marked as such?
>
> The ITS text analytics and/or terminology data categories, which
> also have confidence scores could be useful for annotating such a
> process:
> http://www.w3.org/TR/its20/#textanalysis
> http://www.w3.org/TR/its20/#terminology
>
> To enable such checking and progression in the authoritativeness
> of senses in different languages, it is important that you record
> what senses are a translation of what other senses.
>
> In relation to the senses that are extracted from Wikipedia
> interlanguage links. Do you consider those 'translations', and in
> particular can you tell from those which is the source and which
> is the target?
>
> Interested to hear what you think.
>
> cheers,
> Dave
>
>
>
> On 22/05/2014 17:41, Roberto Navigli wrote:
>> Thanks Felix! To answer Dave's comment: translations come from
>> the automatic translations of semantically annotated corpora, as
>> Tiziano said, and we have a confidence for each of these
>> translations together with the source of the original text.
>>
>> Best,
>> Roberto
>>
>>
>> 2014-05-22 18:35 GMT+02:00 Tiziano Flati <tiziano.flati@gmail.com
>> <mailto:tiziano.flati@gmail.com>>:
>>
>> @Felix:
>>
>> I am wondering if ITS 2.0 properties could help here, see
>> https://www.w3.org/International/its/wiki/ITS-RDF_mapping
>> There is mtConfidence which provides the confidence value
>> for machine translation and mtConfidenceAnnotatorsRef to
>> identify the tool used.
>> Also, there is provenance related properties, starting at
>> :org, until :revToolRef, that could identify the
>> provenance information you need. The underlying
>> definitions for the two ITS data categories are at
>> http://www.w3.org/TR/its20/#provenance
>> http://www.w3.org/TR/its20/#mtconfidence
>>
>> Yes, I think that the ITS 2.0 can definitely be a very good
>> point to explore. At the moment I don't think we need
>> modelling properties more complex than those ones (such as
>> mtConfidenceRule, etc.), so I think this fits well our needs.
>>
>> @Lewis:
>>
>> Do you know currently the provenance of the translation
>> between senses in babelNet. Have you produced any of the
>> translations yourself, or to you just take the links
>> where they are present in the source resources, e.g. DBpedia.
>> What is the policy in Babelnet, is some translation
>> better than none, or is there a translation confidence
>> threshold, e.g. based on human checking, Mt confidence or
>> logical inference etc that you employ?
>>
>> BabelNet translations can come from explicit resource
>> information (e.g., Wikipedia interlanguage links) or as
>> automatic translations supported by millions of sense-tagged
>> sentences coming from Wikipedia and Semcor.
>> In conclusion, AFAIK, BabelNet *does have* translation
>> quality estimation, so I think that indication about
>> confidence could be also provided. (Roberto, correct me if I
>> am wrong)
>>
>> Thank you all for your comments and suggestions :)
>> Tiziano
>>
>> 2014-05-22 16:07 GMT+02:00 Dave Lewis <dave.lewis@cs.tcd.ie
>> <mailto:dave.lewis@cs.tcd.ie>>:
>>
>> Hi Tiziano, Roberto,
>> Do you know currently the provenance of the translation
>> between senses in babelNet. Have you produced any of the
>> translations yourself, or to you just take the links
>> where they are present in the source resources, e.g. DBpedia.
>>
>> In a localization or MT application we look at in CNGL
>> and FALCON, where we may use translation to guide
>> translators or help train MT engines, the provenance is
>> important so some policies can be applied to reduce the
>> propagation of inaccurate translation, or translation
>> that are not appropriate to the context at hand - so
>> those ITS attributes are really important there. To thins
>> extend, when representing this as linked data, we define
>> 'wasTranslatedFrom' as a property of
>> 'prov:wasDerivedFrom' to reify other provenance meta-data
>> - agents, tools, context etc.
>>
>> What is the policy in Babelnet, is some translation
>> better than none, or is there a translation confidence
>> threshold, e.g. based on human checking, Mt confidence or
>> logical inference etc that you employ?
>>
>> many thanks,
>> Dave
>>
>>
>> On 22/05/2014 10:42, Felix Sasaki wrote:
>>> Hi Titziano,
>>>
>>> sorry that I could not make the call due to personal
>>> reasons.
>>>
>>> In the draft I saw under „translation“ this issue:
>>>
>>> „Issues: Information about translation confidence (was
>>> it humanly or automatically produced? if automatic, with
>>> what confidence score?) and translation provenance (what
>>> text(s) does the translation come from? who translated
>>> and with what tool?).
>>> Another issue concerns whether the
>>> relation lexinfo:translation is essential or not: every
>>> sense in a language within a BabelSynset is, in fact, a
>>> translation of any other sense in another language, so
>>> that this information could actually be derived (problem
>>> of redundancy). However, having data linked one to each
>>> other could also be a benefit, since the information is
>>> explicit in the resource.“
>>>
>>> I am wondering if ITS 2.0 properties could help here, see
>>>
>>> https://www.w3.org/International/its/wiki/ITS-RDF_mapping
>>>
>>> There is mtConfidence which provides the confidence
>>> value for machine translation and
>>> mtConfidenceAnnotatorsRef to identify the tool used.
>>>
>>> Also, there is provenance related properties, starting
>>> at :org, until :revToolRef, that could identify the
>>> provenance information you need. The underlying
>>> definitions for the two ITS data categories are at
>>> http://www.w3.org/TR/its20/#provenance
>>> http://www.w3.org/TR/its20/#mtconfidence
>>>
>>> Best,
>>>
>>> Felix
>>>
>>> Am 22.05.2014 um 10:12 schrieb Tiziano Flati
>>> <tiziano.flati@gmail.com <mailto:tiziano.flati@gmail.com>>:
>>>
>>>> Dear all,
>>>>
>>>> we have compiled a first draft of guidelines for the
>>>> conversion of BabelNet as Linguistic Linked Data. The
>>>> initial draft is here
>>>> <https://docs.google.com/document/d/184C_AjY7_PYBSc8SnAFghGLyTo1v312N34dsP9QZokI/edit#>.
>>>>
>>>> We can probably integrate this into the BPMLOD
>>>> community report both as a separate document and in the
>>>> form of all our resource-dependent and independent
>>>> details/comments.
>>>> Any feedback and comment is also very appreciated and
>>>> will help us improving the draft.
>>>>
>>>> Best regards,
>>>> Tiziano Flati and Roberto Navigli
>>>
>>
>>
>>
>>
>>
>> --
>> =====================================
>> Roberto Navigli
>> Dipartimento di Informatica
>> Sapienza University of Rome
>> Viale Regina Elena 295 (second floor)
>> 00161 Roma Italy
>> Phone: +39 0649255161 <tel:%2B39%200649255161> - Fax: +39 06
>> 8541842 <tel:%2B39%2006%208541842>
>> Home Page: http://wwwusers.di.uniroma1.it/~navigli
>> <http://wwwusers.di.uniroma1.it/%7Enavigli>
>> =====================================
>
>
>
>
> --
> Jorge Gracia, PhD
> Ontology Engineering Group
> Artificial Intelligence Department
> Universidad Politécnica de Madrid
> http://delicias.dia.fi.upm.es/~jgracia/
> <http://delicias.dia.fi.upm.es/%7Ejgracia/>
Received on Tuesday, 27 May 2014 01:23:43 UTC