Using Linked Data to Create a Typological Knowledge...

OverviewInvestigationChallenges

Using Linked Data to Create a TypologicalKnowledge Base

Steven MoranLMU Munchen

[email protected]

March 8, 2012

Steven Moran, LMU Linked Data in Linguistics


Talk Map

� Overview

� Investigation

� Challenges

� Conclusion



Where?What?How?

Where are we in linguistic structure?

!

"

µ

# #

"

µ

f o n ə l ɑ dʒ ə k ɫ Θ i ə ɹ i

$ $ $

" " " " " "

µ µ µ µ µ µ µ

← phonological phrase

← phonological words

← metrical feet

← syllables

← moras

← segments

← distinctive features-voice+labiodental+continuent

+voice+back-low-high

+voice+nasal+coronal

...



Where?What?How?

PHOnetics Information Base and LExicon (PHOIBLE)

� A typological data set of segment inventories with linguisticand non-linguistic information

� Linguistic info� segment inventories� distinctive features� genealogical data (language stock and genus)

� Non-linguistic info� population figures� geographic location (geo-coordinates, country and region)� per-capita GDP, etc.



Where?What?How?

Segment inventory databases

� Stanford Phonology Archive (SPA; Crothers et al 1979)

� UCLA Phonological Segment Inventory Database (UPSID;Maddieson 1984; Maddieson & Precoda 1990)

� Systemes alphabetiques des langues africaines (AA; Hartell1993; Chanard 2006)

� PHOIBLE data (Moran 2009-2012)



Where?What?How?

PHOIBLE resources

SPA (197)phonemes, allophones, phono rules

UPSID (451)phonemes

PHOIBLE (738)

phonemes,allophones,phono rules,

citations,squibs AA (203)

phonemes, graphemes

Metadata and Unicode

IPA

PHOIBLE combined

inventories(1298/1589)

UnicodeNormalization



Where?What?How?

How was PHOIBLE developed?

PHOIBLE relational database

PHOIBLE flat files:

1. Aggregated2. Phoneme level

Data warehouseprocedure

RDF graph ofsegments

Python script

RDF graph offeatures

Python script

PHOIBLERDF graph

(segments and features)

Merge graph Merge graph

Data sources:

PHOIBLE

AA

UPSID

SPAETL

processesData sources:

Unicode IPA

WALS

Multitree

Ethnologue

CIA

Flat files



Where?What?How?

PHOIBLE data warehouse flat file table – phoneme level



Where?What?How?

Building the RDF graph

Subject



Where?What?How?


Subject Object



Where?What?How?


Subject

u!

Object



Where?What?How?


Subject

u!

Object

hasSegment



Where?What?How?


tha

Subject

u!

ObjecthasSegment

khm hasSegment



Where?What?How?


tha

Subject

u!

ObjecthasSegment

khm hasSegment

"#$

kor

lbe

kat

bsk

hasSegment

hasSegment

hasSegment

hasSegment



Where?What?How?


thau!

khm

"#$

kor

lbe

kat

bsk

hasSegment

asisncau

ausa

taik

Korean

Lak-Dargwa Kartvelian

BurushaskiKhmer

Kam-Tai

hasEthnologueLanguageFamilyStock

hasWalsGenus

20,200,000

hasPopulation

37:30

128:0

hasLatitude

hasLongitude

kart



Where?What?How?

Knowledge base with segments

Sisaala, Western

[ssl]p

b

kp

hasSegment

hasSegment

hasSegment



Where?What?How?

Knowledge base with segments and features

� Features added to the graph by linking them from eachsegment

ssl

p b kp

voice plosive velar

hasSegment

hasFeature



Where?What?How?

Knowledge base with segments, features and feature sets

ssl

p b kp

voice plosive velar

hasSegment

Hayes 2009

Maddieson1984

consonantallabial dorsal

hasFeature

hasFeature

bilabial



Where?What?How?

Mapping distinctive features to segment types

� Features are the atoms that combine compositionally to forma segment

� Query features for natural classes of sounds

� Model in RDF/OWL to hierarchically organize features into afeature geometry



Where?What?How?

Simple and complex segment feature resolution



Where?What?How?

Feature specifications for natural classes of sounds

Class of sounds Feature specificationVowels [+syllabic] [-consonantal]Vowels & Syllabic Consonants [+syllabic]Glides [-syllabic] [-consonantal]Liquids [+consonantal] [+approximant]Nasals [+sonorant] [-approximant]Fricatives [-sonorant] [+continuant]Affricates [-continuant] [+delayed release]Stops [-delayed release]Stops & Affricates [-continuant]Liquids & Glides [-syllabic] [+approximant]Liquids, Glides, & Nasals [-syllabic] [+sonorant]



Where?What?How?

Feature geometry



Where?What?How?

Feature geometry

root

supralaryngeallaryngeal

approximant

anteriorback

consonantal

constricted glottis

continuant

coronal

delayed release

distributed

dorsal

fortis

front high

labial

labiodental

lateral

long

low

nasal

round

sonorant

spread glottis

stress

strident

syllabictap

tense

tone

trill

voice

place



BasicsInvestigating patterns

SPARQL

� SPARQL Protocol And RDF Query Language

� RDF query language (Prud’Hommeaux and Seaborne, 2006)

� consist of triple patterns that match concepts and theirrelations by binding variables to match graph patterns




Select segments of a particular language

SELECT ?segmentsWHERE {ssl hasSegment ?segments}

hasSegmentSisaala, Western

[ssl]

p

b

kp

English [eng]

Sisaala, Tumulung

[sil]

gb

German [deu]

v

f




Select languages that have a particular segment

SELECT ?languagesWHERE { ?languages hasSegment gb }


[ssl]

p

b

kp

English [eng]

Sisaala, Tumulung

[sil]

gb

German [deu]

v

f




Select languages that have a class of segments

SELECT ?languagesWHERE { ?languages hasSegment ?segments .?segments hasFeature DELAYED RELEASE }

hasSegment

Sisaala, Western

[ssl]

p

b

kp

Sisaala, Tumulung

[sil]

gb

v

f

labial

delayed release

syllabic

hasFeature

Hayes2009




Select languages that have a class of segments

SELECT ?languagesWHERE { ?languages hasSegment ?segments .?segments hasFeature DELAYED RELEASE }


[ssl]

p

b

kp

Sisaala, Tumulung

[sil]

gb

v

f

labial

delayed release

syllabic

hasFeature

Hayes2009




Investigating phonological universals

� Hyman (2008): Every phonological system has...� stops� at least one unrounded vowel� at least one front vowel or the palatal glide /j/� coronal phoneme(s)




Every phonological system has stops

SELECT ?languagesWHERE {?languages phoible:hasSegment ?segments .?segments phoible:notHasFeature feature:DELAYED RELEASE}




Phonological system with at least one unrounded vowel –and what are those languages and their segments?

SELECT ?languages ?segmentsWHERE {?languages phoible:hasSegment ?segments .?segments phoible:hasFeature feature:SYLLABIC .?segments phoible:notHasFeature feature:CONSONANTAL .?segments phoible:notHasFeature feature:ROUND}




Every phonological system has at least one front vowel orthe palatal glide /j/

SELECT ?languagesWHERE {?languages phoible:hasSegment ?segments .?segments phoible:hasFeature feature:SYLLABIC .?segments phoible:notHasFeature feature:CONSONANTAL .?segments phoible:hasFeature feature:FRONT .UNION { ?languages phoible:hasSegment segment:j }}




Every phonological system has coronal phonemes... nope!

SELECT ?languagesWHERE {?languages phoible:hasSegment ?segments .?segments phoible:hasFeature feature:CORONAL}

� “Another Universal Bites the Dust” (Blevins, 2009)

� Northwest Mekeo [mek] /p, B, m, w, g, N, j, i, e, a, o, u/



Linguistic challengesComputational challengesMetadata and discovery

Linguistic challenges

� Which language is this? (“A Grammar of Haida” - Northern?Southern?)

� Different theoretical models are used in describing languages� Diacritic ordering

� creaky voiced syllabic dental nasal: n”"� labialized aspirated long alveolar plosive: twh:




Different analyses – same language




Using inference

� RDF model we can add additional knowledge to the model

� Change knowledge base without changing our query

� Logically-defined properties in OWL; merge OWL and RDFgraphs

� Establish relationships between resources as inferred by areasoner

� Reasoner evaluates logic statements in graph and addsinferred triples

� Ability to manipulate the ontology and to specify how toderive logical consequences and to create new entailments




Using OWL logic to extend the knowledge base

http://phoible.org/segment/uhttp://phoible.org/id/iso639-3/

amp

http://phoible.org/id/iso639-3/ant http://phoible.org/segment/!

http://phoible.org/segment/"

http://phoible.org/hasSegment

http://phoible.org/id/iso639-3/apn

http://phoible.org/segment/u#




Using OWL logic to extend the knowledge base

http://phoible.org/segment/uhttp://phoible.org/id/iso639-3/

amp

http://phoible.org/id/iso639-3/ant http://phoible.org/segment/!

http://phoible.org/segment/"

http://phoible.org/hasSegment

http://phoible.org/id/iso639-3/apn

http://phoible.org/segment/u#

owl:sameAs




Computational challenges

� Adherence to Unicode IPA:� g/g, !/!, a/A, p/p

� Rendering sequences of Unicode characters as the samesegment

a˜

a˜U+0061 + U+0330 + U+0303 U+0061 + U+0303 + U+0330

latin small letter a + latin small letter a +combining tilde below + combining tilde +combining tilde combining tilde below




Metadata

� DCMI RDF gets you most of OLAC� Two big things missing: resource type and language

identification� http://www.sil.org/iso639-3/documentation.asp?id=aar� http://www.ethnologue.com/show language.asp?code=aar� http://lexvo.org/id/iso639-3/aar� http://wals.info/languoid/by code/iso 639 3 aar� http://phoible.org/id/iso639-3/aar� http://resource dc:subject GOLD:Language� http://glottolog.livingsources.org/languoid/id/25785.xhtml




Poornima & Good 2010




Other questions moving forward...

� How do we access the Linguistic Linked Open Data cloud?� Download the RDF/OWL files and run locally?� In our code and point to the files online?

� What about a publicly accessible interface?

� SPARQL endpoint to make the LLOD data widely available




Summary

� PHOIBLE provides a large sample of segment inventories andadditional phonological information from the world’s languages

� It uses RDF and OWL graph data structures to captureknowledge about segments and distinctive features

� Can be used to ask questions of phonological systems... andmore




Many thanks to the participants and organizers of LDL

And... Emily Bender, Michael Cysouw, Morgana Davids, ScottDrellishak, Shauna Eggers, David Ellison, Scott Farrar, ChristopherGreen, Sharon Hargus, Richard John Harvey, Jeff Good, KelleyKilanski, William Lewis, Michael McAuliffe, Dan McCloy, BrandonPlasters, Tristan Purvis, Cameron Rule, Daniel Smith, Daniel Veja& Richard Wright


Using Linked Data to Create a Typological Knowledge...

Documents

Transcript of Using Linked Data to Create a Typological Knowledge...