ircs workshop on linguistic databases 11-13 december 2001 philadelphia standards for language...

IRCS Workshop on Linguistic Databases • 11-13 December 2001 • Philadelphia

Standards for Language Resources

Nancy IDEDepartment of Computer Science Vassar College

Laurent ROMARYEquipe Langue et Dialogue

LORIA/INRIA


Goals

• present an abstract data model for linguistic annotations and its implementation using XML, RDF and related standards

• outline work of newly formed ISO committee: TC 37/SC 4 Language Resource Management– Using the work described as its starting point– Solicit the participation of members of the

research community


Goals of ISO TC 37/SC 4

• prepare international standards/guidelines for effective language resource (LR) management in mono- and multi-lingual applications

• develop principles and methods for creating, coding, processing and managing LR– written corpora, lexical corpora, speech corpora,

dictionary compiling and classification schemes

• Focus : – data modeling– data exchange, evaluation


Standardization Process

• Two-phases: 1. Develop basic architecture to support

wide-range of applications 2. Use as basis for building more precise

standards for LR management

• Liaison with ISLE – Incorporate existing standards where

possible– Broaden by including additional

languages (e.g. Asian)


Standardization is Tricky

• Skepticism within the community

• Arguments against LR standardization: 1. diversity of theoretical approaches makes

standardization impractical or impossible

2. vast amounts of existing data and processing software will be rendered obsolete by the acceptance of new standards


SC4 Approach• Efforts geared toward defining abstract

models and general frameworks for creation and representation of language resources– In principle, abstract enough to accommodate

diverse theoretical approaches

• Situate development squarely in the framework of XML and related standards – Ensure compatibility with established and widely

accepted web-based technologies– Ensure feasibility of transduction from legacy

formats into newly defined formats


Call for Participation

• Success of the committee depends on community’s awareness of its activity, in order to ensure widespread adoption

• Involve from the outset broad range of potential users of the standards


The General Framework

• Model for linguistic annotation that can– be instantiated in a standard

representational format– serve as a pivot format into and out of

which proprietary formats may be transduced to enable• comparison• merging• manipulation via common tools


Overall Plan

FormatA

FormatC

Abstract Format

Operation via common tools, merging, etc

FormatB

Annotation Format Tower of Babel


DialectSpecification

DATACATEGORY REGISTRY

VirtualAML

ConcreteAML

Data CategorySpecification

STRUCTURAL SKELETON

Abstract XML

encoding

ConcreteXML

encoding

Non-XML Encoding

Universal Resources

Project Specific Resources

XSLT Script

Overall Architecture


N.B.

• We do not expect XML to necessarily serve as the internal format used by tools etc.

• We do not care about creating yet another “standard” format

• We do not care (for this work) about designing specific annotation formats


Data Model• Identify a consistent underlying data

model for data and its annotations– Formalized description of data objects

• Composition• Attributes• Class membership• Applicable procedures, etc

– Formalized description of relations among data objects

– Independent of instantiation in any particular form


(Most) Abstract Model

• An annotation is a set of data or information associated with some other data

• More precise: an annotation is a one- or two-way link between – an annotation object, and – a point or span (or a list/set of points or spans)

within a “base” data set

• Links may or may not have a semantics• Points and spans may be objects, or sets/lists

of objects


PRIMARY DATA

[

[ [

ANNOTATIONOBJECT

ANNOTATIONOBJECT

ANNOTATIONOBJECT

ANNOTATIONOBJECT


Observations

Granularity of the data representation and encoding is critical

Must be possible to represent objects and relations in some form that prevents information loss


Representing Annotation Objects

• Annotation objects may be relatively complex• Abstract representation

– graph of elementary structural nodes to which one or more information units are attached

– distinction between structure and information units is critical to the design of a truly general model

• Annotations may be structured in several ways– Most common: hierarchical

• phrase structure analyses of syntax • lexical and terminological information • etc.


Relations Among Annotations

1. Parallelism– two or more annotations refer to the same

data object

2. Alternatives– two or more annotations comprise a set of

mutually exclusive alternatives

3. Aggregation– two or more annotations comprise a list or

set that should be taken as a unit


Information Units• Also called data categories

– provide the semantics of the annotation– most theory and application-specific part of an

annotation scheme

• No attempt to define data categories – Proposal : development of a Data Category Registry – Define data categories with RDF schemas – Formalize properties and relations – Templates that describe how objects are instantiated– Inheritance of appropriate properties


Data Category Registry

• Several functions1. provide a precise semantics for annotation

categories• can be used “off the shelf” or modified

2. provide a set of reference categories onto which scheme-specific names can be mapped

3. provide a point of departure for definition of variant or more precise categories

• Overall goal– Ensure that semantics of data categories are

well-defined and understood


Generic Mapping Tool (GMT)

• Instantiation of abstract format in XML• Why XML?

– Supported standard– Built-in representation for hierarchies

(nested tags)– Sophisticated linking mechanisms

• Can link to points, spans, use explicit locations or tags

– XSLT for transduction, XML Schemas for validation, etc.


A Few Simple Tags• <struct>

• represents a structural node in the annotation• may be recursively nested at any level

• <feat> – provides information attached to the node

represented by the enclosing <struct> – type attribute identifies data category – Contents:

• string providing a value for the data category • recursively nested <feat> elements (for complex structures)• empty--points via a target attribute to an object in another

document


Other Tags• <alt>

– brackets alternative annotations• <rel>

– points to a non-contiguous related element• <seg>

– points to the data to which the annotation applies

– assume the use of stand-off annotation– target attribute uses XML Pointers

• <brack> – groups information to be regarded as a unit


• Tag names etc. unimportant– It is the underlying data model that counts– Essentially uses feature structures

• GMT sufficiently powerful to represent information across annotation types

• Demonstrated applicability to – terminological and lexical information (Ide, et al.,

2000) – syntactic annotation (Ide and Romary, 2001)

• Existing formats (XML or other) mapped to the GMT for merging, manipulation via common tools, etc.; then re-map to original formats for use in in-house tools and applications. etc.


Examples

• Morpho-syntactic annotation– involves the identification of word classes

over a continuous stream of word tokens– may refer to the segmentation of the input

stream into word tokens– may also involve grouping together

sequences of tokens or identifying sub-token units (or morphemes

– description of word classes may include one or several features

• syntactic category, lemma, gender, number,…


Representation in GMT

• Single type of structural node – represents a word-level structure

unit

• One or several information units associated with each structural node


Simple Case<struct> <struct type=”W-level”>

<feat type=”lemma”>Paul</feat><feat type=”pos”>PNOUN</feat><seg target=”#w1”/>

</struct> <struct type=”W-level”>

<feat type=”lemma”>aimer</feat><feat type=”pos”>VERB</feat><feat type=”tense”>present</feat><feat type=”person”>3</feat><seg target=”#w2”/>


<feat type=”lemma”>le</feat><feat type=”pos”>DET</feat><feat type=”number”>plural</feat><seg target=”#w3”/>


<feat type=”lemma”>croissant</feat><feat type=”pos”>NOUN</feat><feat type=”number”>plural</feat><seg target=”#w4”/>

</struct></struct>

“Paul aime les croissants”

Pointers to data in primary document


Representing More Complex Cases

<struct type=”W-level”><seg target=”#w1”/><struct type=”W-level”> <feat type=”lemma”>de</feat> <feat type=”pos”>PREP</feat></struct><struct type=”W-level”> <feat type=”lemma”>le</feat> <feat type=”pos”>DET</feat></struct>

</struct>

Example: “du” = “de” + “le” in French

Points to “du” in text

Gives the structure of the “word” underlying the word


GMT as a Tree Structure

….………..…….du….…………….………………………..…………

Primary Document

Lemma : de

Pos : prep

seg :

Lemma : le

Pos : det


Compound Words

<struct type=”W-level”><feat type=”lemma”>pomme_de_terre</feat><feat type=”pos”>NOUN</feat>

<struct type=”W-level”> <seg target=”#w1”/> <feat type=”lemma”>pomme</feat> <feat type=”pos”>NOUN</feat>


<seg target=”#w2”/> <feat type=”lemma”>de</feat> <feat type=”pos”>PREP</feat>


<seg target=”#w3”/> <feat type=”lemma”>terre</feat> <feat type=”pos”>NOUN</feat>

</struct></struct>

Example: “pomme de terre”

Componentlemmas

Primarylemma


Tree

….………..……………Pomme de terre

………………………..…………

Primary Document

Seg :

Lemma : pomme

Pos : noun

lemma : pomme_de_terre

Seg :

Lemma : de

Pos : prep

Seg :

Lemma : terre

Pos : noun


Advantages

• Enables specification of the required level of granularity – granularity of the segmentation in (or associated

with) primary data may not correspond to that required for the annotation

• Can define relations over the tree independently– Compositional for morpho-syntax, syntax, etc.– Partitions in lexical data– …


Orth : overdress

Pron : [ ]

Pos : verbDef : To dress (oneself or another) too elaborately or finely

Pron : [ ]Pos : nounDef : A dress that may be worn over a jumper, blouse, etc.

<struct> <feat type=“orth”>overdress</feat> <struct> <feat type=“pos”>verb</feat> <feat type=“pron”>[ ]</pron> <feat type=“def”> To dress (oneself or another) too elaborately or finely </feat> </struct> <struct> <feat type=“pos”>noun</feat> <feat type=“pron”> [ ]</pron> <feat type=“def”> A dress that may be worn over a jumper, blouse, etc.</feat> </struct></struct>


Alternatives<struct type=”W-level”> <seg target=”#w1”/> <brack>

<alt> <feat type=”lemma”>boucher</feat> <feat type=”pos”>VERB</feat> <feat type=”tense”>present</feat> <feat type=”confidence”>0.4</feat>

</alt><alt> <feat type=”lemma”>bouche</feat> <feat type=”pos”>NOUN</feat> <feat type=”confidence”>0.6</feat>

</alt> </brack></struct>


Relating Annotation Levels

• Three ways:1. Temporal anchoring

• associates positional information with each structural level

2. Event-based anchoring• introduces a structural node to represent a

location in the text to which all annotations can refer

3. Object-based anchoring• enables pointing from a given level to one or

several structural nodes at another level


Temporal Anchoring• Positional information

– Usually, a pair of numbers expressing the starting and ending point of segment

• Attributes for <seg>:• /startPosition/: the temporal or offset position of the

beginning of the current structural node;• /endPosition/: the temporal or offset position of the

end of the current structural node.

• Example:<struct type=”phonetic”> <seg startsAt=”2300” endsAt=”3200”/> <feat type=”phone”>iy</feat></struct>


Event-based Anchoring

• Useful when:– Not possible/desirable to modify the primary

data by inserting markup to identify specific objects or points in the data

– Primary data is marked with “milestones” (e.g., time stamps in speech data), where spans across the various milestones must be identified• Here,<struct> elements represent markup for

segmentation (e.g., segmentation into words, sentences, etc.).


GMT Rendering

• Structural node (landmark) referred to by annotations for the defined span

<struct type=”landmark”>

<seg startsAt=”2300” endsAt=”3200”/>

</struct>

• Annotation graph formalism explicitly designed for this


GMT Advantages• AG formalism reifies the “arc” vs.

identification via XML tags• GMT : the two methods are analogous

– annotator can use either method

• AG not well-suited to hierarchically organized annotations– requires special mechanisms

• GMT: exploits the hierarchical structure built in to XML – “flat” and hierarchical annotations treated using the

same mechanisms


Object-based Anchoring

• Useful to make dependencies between two or more annotation levels explicit – Example: syntactic annotation can

refer directly to the relevant nodes in a morpho-syntactically annotated corpus

<struct type=”W-level”> <seg target=”#w3”> <struct type=”W-level”> <seg target=”#w3.1”> <feat type=”lemma”>de</feat> <feat type=”pos”>PREP</feat> </struct> <struct type=”W-level”> <seg target=”#w3.2”> <feat type=”lemma”>le</feat> <feat type=”pos”>DET</feat> <feat type=”gender”>masc</feat> </struct> </struct> <struct type=”W-level”> <seg target=”#w4”> <feat type=”lemma>chat</feat> <feat type=”pos”>NOUN</feat> </struct></struct>

<struct> <feat type=”synCat”>NP</feat> <seg targets=”w3.2 w4”/></struct>

Representation for “du chat”


GMT as a Modeling Tool• Rendering various formats into GMT

representation has revealed some problems, inconsistencies in existing formats– Penn Treebank : inconsistent indication of

relations (see Ide and Romary, ACL 2001 or Abeillé Treebank book, forthcoming)

– NOMLEX lexicon : no (automatically perceivable) distinction between lists and alternatives

• The abstract format serves the unexpected purpose of providing a “template” for fundamental annotation properties


Jumping Ahead…• Is XML distracting us from our real work?

– YES, because• Focus on details of using XML and related standards can

obscure the real work of data modeling– BUT

• Datas models are no use only in the abstract - need means to implement

• XML, schemas, RDF, etc. are powerful data modeling tools based on years of research in this area

• Need to know how to best exploit them for our purposes

• Need a synergy between modeling efforts and implementation in XML, RDF, etc.

• Need to remember that using XML is just a vehicle to ensure flexibility, convertability, and compatibility with evolving technologies


Conclusion

• ISO committee – Work is continually evolving

• Try to stay at the leading edge of data representation

– We are only at the “assembly language” level

– We need to do this right to enable a “web of databases”

• Call for participation!!!


Thank You

Contacts

US Expert, ISO TC37 SC4Nancy [email protected]

Chairman, ISO TC37 SC4Laurent [email protected]

ircs workshop on linguistic databases 11-13 december 2001 philadelphia standards for language...

Documents

linguistic databases

ircs workshop

philadelphia data model

philadelphia standards

linguistic annotations

data representation

abstract data model

set of data