Ontology of Linguistic Annotation: Overview (2007)

Ontologies of Linguistic Annotation. Machine-readable tagsets and annotation schemata for more than 100 languages.

Ontology of Linguistic Annotation: Overview (2007)

Note: This document provides overview over and design decisions of the modular architecture of OLiA and the design of the OLiA Reference. However, it has been written in May 2007 and (unless marked as such) it has been marginally updated only. The original is available from the University of Potsdam, Germany, with additions taken from another 2007 overview. Major later additions only partially addressed here include the coverage of syntax and discourse annotations that have been added in 2007, and 2014, respectively.

Background

Concentrating on the more elementary levels of linguistic analysis such as parts of speech and morphology, a generalization over different terminologies applied for the annotation of the corpora hosted by three collaborative research centers SFB 441 (Tübingen), SFB 538 (Hamburg) and SFB 632 (Potsdam/Berlin) was developed, and later extended for NLP tools and corpora beyond these resources. As a result, an ontology was developed which specifies reference terminology, and the tags of the original annotated data are linked with this reference terminology. Besides its function in annotation documentation, the ontology can be applied for the formulation of tag-set neutral corpus queries. For this purpose, I developed the OntoClient, a JAVA-based query pre-processor which translates formal ontology-based specifications into disjunctions of concrete tags. The OntoClient serves as a pre-processor for corpus querying languages such as ANNIS-QL and CQP, furthermore, it was applied in the specification of tag-set independent corpus processing scripts.

The OLiA ontologies were initially developed in the context of the project "Sustainability of Linguistic Resources", a collaborative project between three German Collaborative Research Centers (SFBs), The Collaborative Research Centres involved in the project are the SFB 538 'Multilingualism' at the University of Hamburg, the SFB 632 'Information Structure' at the University of Potsdam and the Humboldt University Berlin, and the SFB 441 'Linguistic Data Structures' at the Eberhard Karls University Tübingen.

The project aimed at preparing language resources to assure an accessible dissemination and sustainable storage of linguistic corpora. One of the main goals of the project was a practical one: resources acquired in long-term projects situated in the three Collaborative Research Centres have to be converted in either one or multiple formats to be sustainably usable by researchers and applications. Furthermore, the project developed unified methods of access for the heterogeneous data acquired in the projects.

Why do we need it?

One of the tasks addressed by the sustainability project was the integration of heterogeneous terminology, especially those applied for the annotation of existing corpora. Examples for such differences range from minor variation in the choice of tag names (which often go unrealized and thus, affect the reliability of broad-scale corpus studies) to fundamental conceptual differences.

Different abbreviations for the same annotations

Same abbreviation for different annotations

Same annotation, but different interpretation

Different granularity of tag sets

Conceptual overlap

All these problems are taken from the seemingly most elementary domain, the domain of part of speech tags, however, more problems arise as soon as morphology, syntax, or discourse phenomena are addressed.

In order to overcome such problems, terminological integration is necessary, i.e.

To provide an integrated access to terminologically heterogeneous resources, it is also necessary to provide an abstract model of linguistic reference terminology to which individual annotations refer, a so-called "terminological backbone".

Classical solutions are the standardization approach and the interlingua approach:

Standardization (cf. the EAGLES recommendations on morphosyntactic annotation)

Interlingua (cf. the AMALGAM project)

Both solutions are limited in flexibility and scalability, and hence, both approaches are applicable only within a limited domain. The standardization approach relies on the existence of common grammatical categories and features found in the languages for which standard-conformant tag sets are to be developed. Otherwise, it results in both reductionism (there are languages for which necessary distinctions cannot be expressed) and complexity projection (there are languages for which a distinction postulated by the standard doesn’t apply). However, even the sheer existence of universal morphosyntactic categories has been questioned in typologic research, and hence, the EAGLES-based standardization approach is unlikely to extend beyond "Standard Average European" languages.

The interlingua approach, however, involves the process to construct an interlingua between existing schemes, and is less statically than the standardization process. However, the complexity of the interlingua grows monotonically with every new language/tag set considered, and, hence, the general applicability of the interlingua approach is restricted by its limited scalability.

For this reason, the project “Sustainability of Linguistic Data” (2005-2008) has been developing an ontology of linguistic annotations as a more flexible representation of a "terminological backbone". Subsequently, this research has been transferred into an open community project by 2011, with major contributions especially provided by the Applied Computational Linguistics (ACoLi) lab of Goethe University Frankfurt, Germany.

Design

So far, we have developed an ontology of linguistic annotations with special consideration of part of speech and morphological annotations existing the participating Collaborative Research Centers (Schmidt et al. 2006, Chiarcos 2006c, Chiarcos 2006d, Chiarcos 2007).

The approach relies on the ontological reconstruction of annotation schemes based on guidelines and additional documentation in so-called "annotation models" (or "domain models").

Every annotation model represents one tag set or annotation scheme, with nonterminal nodes (concepts) representing conceptual categories as mentioned in the documentation or indicated in the document structure of the annotation guidelines, and terminal nodes (instances) representing concrete annotation values, or tags.

As an illustration, prototypes for the following annotation models are available in an HTML serialization:

With respect to morphosyntactic annotations, the OLiA annotation models currently comprise 16 annotation schemes applied to 42 languages (5 annotation models for English, 5 annotation models for German, 2 annotation models for Russian, one annotation model for Tibetan, one for Old High German, the Connexor annotation model for 10 European languages, one annotation model for a typologically-oriented annotation scheme applied to 29 languages). Annotation models for syntax and information structure/anaphora are currently under construction.

The concepts of these annotation models are linked to a common OLiA Reference Model which is based on the EAGLES recommendations for morphosyntax, and extended according to the needs of the participating annotation models.

The annotation models are then mapped onto the categories specified in the reference model by means of conceptual subsumption (rdfs:subClassOf, rdfs:subPropertyOf). This mapping is specified in separate "linking files", thus making both the reference model and the annotation models independent and self-contained ontologies.

The "reference model", however, does not specify authoritative definitions for existing terminology, but only a fairly traditional view on it. Hence, its primary function is not to provide prescriptive definitions of terms, but only to provide a reference point for the participating annotation models. Whenever a more reliable ontology of linguistic terminology will be developed (e.g. revised versions of the General Ontology of Linguistic Description (GOLD) or the grammis ontology), the reference model can be linked with it in the same way as the annotation models are linked with the reference model, and thus mediate between such an external reference model and the annotation models. In this sense, the reference model serves as an interface to the annotation model, and it could be better termed "interface model".

A Structured Ontology

The OLiA ontology consists of three major components, i.e.:

This tripartite structure can be augmented by the optional linking of the Reference Model with additional upper models (External Reference Models). As a result, these upper models can be applied for the formulation of search queries as an alternative to the reference terminology specified in the interface model. Reference definitions retrievable from upper models to domain models are thus mediated by the Reference Model.

We claim that this modular approach is more flexible as it allows alternative specifications of linking and the inclusion of alternative upper models as well as additional Annotation Models. In present-day annotation technology, it finds a close pendant in the standoff paradigm according to which different levels of annotation and the primary data have to be separated from each other in order to allow for distributed maintenance and concurrent modification. Besides these advantages, it allows for user-specific modifications (such as the specification of alternative upper models) without compromising the ontology as a whole.

Drafting the OLiA Reference Model

The OLiA Reference Model provides shared terminology for linguistic annotation of various phenomena, with a focus on morphosyntax and syntax. Semantics has been excluded as we assume this is sufficiently covered by WordNet, FrameNet, etc., and should be modelled in the context of an initiative on lexical resources rather than annotated corpora.

Update 2011: Such an initiative is represented by the W3C Community Group Ontology-Lexica (OntoLex).

Update 2014: We have developed a designated extension for discourse phenomena. It is, however, not directly integrated with the OLiA Reference Model.

Note 2022: The intended function of the OLiA Reference Model is not to provide reference terminology, but to serve as an interface between numerous annotation schemes and numerous community-maintained reference terminology. Originally, the OLiA Reference Model has thus been referred to as Interface Model, but as a result of internal discussions within the Project “Sustainability of Linguistic Data” (2005-2008, U Tübingen, U Hamburg, U Potsdam), within which the OLiA Reference Model had been created, the current name was adopted as a compromise between Interface Model (too vague) and Upper Model (too prescriptive). Note that the original text of this document originates fro m 2007 and that the term Interface Model continues to be occasionally used below.

The core structure of the OLiA Reference Model has been inspired by the specifications of the EAGLES initiative, revised and extended according to the GOLD ontology. The first draft of the OLiA Reference Model has thus also been referred as E(xtended)-EAGLES ontology. By early 2007, the first version of the E-EAGLES ontology has been implemented using OWL/DL with Protege. Currently, it covers all the obligatory and recommended features from the EAGLES recommendations for morpho-syntactic annotation (Wilson and Leech 1996) plus several categories from non-EAGLES conformant tag sets (e.g. noun classifiers).

The classes in the Reference Model are retrieved from the EAGLES recommendations in the following way:

As the project data includes a MULTEXT-East-based annotation scheme for Russian, the Uppsala scheme, the relevant definitions of MULTEXT-East have been integrated as well.

The hierarchy of verbal classes in E-Eagles is given in Fig. 1. Note that compared to the original EAGLES recommendations, AuxillaryVerb and VerbalNoun are redefined in order to account for non-EAGLES conformant tag sets.

Besides this hierarchy of classes, verbs can be further specified by properties such as hasTense, hasAspect, hasPerson, hasNumber, hasVoice, hasSeparability, hasReflexivity and hasGender.

Figure 1: Fragment of upper model: Sub-classification of verbal categories in E-Eagles

An Annotation Model: Uppsala

Annotation Models are built in a similar manner. Usually, annotation guidelines have a document structure which specifies an otherwise implicitly assumed hierarchical organization, thus, a similar hierarchical structuring of concepts can be achieved.

For the tagset applied to the Uppsala corpus, the corresponding structuring of the Annotation Model ontology is given in Fig. 2.

Again, inflectional differentiations are specified by properties in the ontology, i.e. hasGender, hasMood, hasVoice, hasPerson, hasNumber, hasFiniteness, hasAspect and hasTense.

Besides these abstract conceptualizations, concrete tags are integrated as instances into the Annotation Model ontology. In Turtle pseudo-code, the definition of the Uppsala tag verb_finit_prt_0\_sg_neut_refl_pfipf in the ontology can thus be given as:

:verb_finit_prt_0\_sg_neut_refl_pfipf a :VerbFinitPrtType; 
:hasTense :past;  
:hasVoice :reflexive; 
:hasFiniteness :finite;
:hasGender :neuter;
:hasMood :indicative;
:hasNumber :singular.

Figure 2: Fragment of Annotation Model: verbal categories in the Uppsala tag set.

Further examples below adopt a slightly more compact but equivalent syntax that has been used in an ontology-based query processor operating on OLiA:

VerbFinitPrtType and 
hasTense(past) and 
hasVoice(reflexive) and 
hasFiniteness(finite) and 
hasGender(neuter) and 
hasMood(indicative) and 
hasNumber(singular)

Linking Annotation Model and Reference Model

While Annotation Model and Reference Model are specified as self-contained ontologies in individual owl files, the linking between both is implemented in a separate file by the rdf:description mechanism.

Basically, the linking file contains a specification of Annotation Model classes (not instances) in terms of Reference Model classes and properties, making up a complex inheritance structure as in Fig. 3 (restricted to subclass relationships). Note that besides the primary classes of word types, also properties and property values from the Annotation Model are specified as sub-properties, instances or sub-classes of properties and classes in the Reference Model.

Figure 3: Linking Annotation Model and Reference Model. The case of verbal categories in the Uppsala tagset.

A sample query

The linking file imports both the Reference Model and the corresponding Annotation Model, and thus, it represents an integrating ontology comprising both. If multiple Annotation Models (tag sets) are considered, the corresponding linking files (and the ontologies they import) have to be imported by another file, the so-called master file which represents the ontology as a whole.

In the querying scenario, then, expressions based upon classes and properties in the Reference Model are expanded according to the inheritance structure within and between Reference Model and domain models, and then evaluated.

As an example, if we are searching for past-tense reflexive verbs, a specification like Verb and hasTense(Past) and hasVoice(Reflexive) mentions the Reference Model classes e-eagles:Verb, e-eagles:Past and e-eagles:Reflexive and the properties e-eagles:hasTense and e-eagles:hasVoice. According to the interitance structure depicted in Fig. 3, e-eagles:Verb expands to russ:Verb and further to russ:VerbFinitPrtType, etc. Similarly, e-eagles:hasTense expands to russ:hasTense etc. Thus, amongst other instances, the instance verb_finit_prt_0\_sg_neut_refl_pfipf is returned.

The ontology-based query preprocessor, OntoClient, then replaces the ontology-sensitive part of a search query by a disjunction of the tags corresponding to the respective instances, and this modified search query can be further processed by a corpus querying tool.

External Reference Models: Alternative Upper Models

The very same mechanism that was used to link Annotation Model concepts with Reference Model concepts can be employed to establish a linking between the Reference Model and an additional upper model which provides independent conceptualizations of linguistic terms. Candidates for such upper models are the OntoTag ontologies (an EAGLES-based ontology of linguistic terms with a special application to English and Iberian languages, cf. de Cea 2004), the Data Category Registry currently developed in the context of the Linguistic Annotation Framework (Ide et al. 2004), or GOLD.

As illustration, we are concentrating on GOLD here, as it is a freely available and well-recipied ontological resource with a good coverage of non-European languages. As of January 2007, any concept in the E-EAGLES ontology is augmented with a reference to the (E-)GOLD ontology.

Nevertheless, it seems reasonable to keep the Reference Model ontology and the upper model apart. As the development of GOLD is still ongoing, updated versions of GOLD could compromise the linking with the domain models if the Annotation Models are mapped onto the upper model directly. If both upper model and Reference Model are separated, a modification of the upper model might force an adaption of the linking between upper model and Reference Model, but not necessarily between the upper model and any other existing Annotation Models.

As the upper model is linked with the Reference Model in the same way as the Reference Model and Annotation Model, the corresponding upper model expressions can be used for the formulation of ontology-sensitive corpus queries.

Advantages of the structured approach

The crucial advantage of a structured modular ontology is its highly flexible and user-adaptive character. As illustrated in Fig. 4, the different components of the ontology are stored separately from each other, and as the import mechanism relies on rdf mechanisms, the concrete location of the corresponding files does not affect the validity of the references. As an example, a user may prefer to use a local variant of a certain Annotation Model, for example because his version of the underlying annotation scheme had slightly different naming conventions than the "official" Annotation Model for this annotation scheme, for a typical example see the numerous variants of STTS which have different tags for pronominal adverbs, e.g. PAV, PROAV and PROP. In this case, only some instances in the Annotation Model have to be renamed, whereas the linking can stay as it is. However, in this case the user has to use a local copy of the linking as well which does not differ from the "official" linking in any other ways than the source of the Annotation Model to be imported.

A user thus may introduce an external upper model, he may redefine the linking between an existing Annotation Model and the Reference Model without affecting either of them, and he may integrate additional Annotation Models. However, he may not modify the Reference Model. As it is the central reference point for any linking file, this could affect the linking of other Annotation Models and produce inconsistencies.

This modular structure is thus highly flexible and user-adaptive. A user might even decide to disagree with the conceptualizations in the Reference Model and develop his very own alternative, but as long as he provides a linking between his conceptualizations and those of the Reference Model (i.e. he implements his alternative as an upper model in our sense), he does not have to reconsider the linking to all existing Annotation Models.

Especially in the long run, ongoing maintenance of the ontology might require the integration of additional upper models in order to keep touch with the continuous process of terminological evolution, but not the redesign of the Reference Model. The effort to have an intelligible interface to the resources associated with certain Annotation Models is thus reduced to the task to maintain the linking between Reference Model and upper model.

Our implementation provides a modular view on the ontology. The ontology consists of three principal components, the upper model presenting a central registry of relevant terminology, several Annotation Models, each covering the tags of one specific POS tag set, and the respective linking between upper model and Annotation Model, which are each stored in independent files.

To access to the ontology as a whole, an additional "master file" is necessary which provides unified access to the Reference Model, the upper model, the Annotation Models and the linking between them from separate OWL/RDF files. As the Reference Model does not specify the ultimate repository of linguistic terminology, additional upper models can be integrated in this master file. As a user can define own conceptualisations by this mechanism, the main benefit of our approach and the development of the Reference Model lies in the fact that it is no longer necessary to consider every tag set by its own. Instead, later refinements are mediated by the upper model, thus the most important achievement is that the Reference Model provides a unified access to different tag sets for both querying and redefinition.

Besides its function in tag set neutral corpus queries and in the theory-neutral definition of language-, project- or task-specific annotation schemes by linking the corresponding Annotation Model with the Reference Model, the ontology can be practically applied in the design of tag set neutral corpus processing scripts (Krasavina et al. 2007), or, more generally, in the field of Semantic Web applications and ontology-based annotation (for a similar approach on a more restricted set of languages cf. de Cea et al. 2004).

Figure 4. Structured modular ontology.

Example ontology

An example can be downloaded here!

Addenda

Deep and Shallow Interoperability (2022)

Almost a decade after the section above on OLiA - Why do we need it was drafted for the first time (in 2004, during the conceptional phase of OLiA), the Universal Dependencies (UD) initiative was created (in May 2014, as a GitHub organization) in an attempt to provide a “universal” (cross-linguistically applicable) standard for dependency annotations and morphosyntax.

UD clearly stands out from earlier standardization/interlingua approaches in its popularity and its wide, cross-linguistic application. UD provides a middle ground between standardization and interlingua by having core vocabularies which can be extended for non-universal aspects. The UD parts of speech are a (highly reductionist) standard, but non-standardized features of morphosyntax can be preserved in the UD FEATS annotation. UD FEATS implement an interlingua approach for morphosyntactic features that is originally rooted (via MULTEXT-East) in EAGLES, and which provides a standardized core inventory (“universal standard”) that can be extended with language-specific features on demand. UD dependencies are a (reductionist and theory-specific) standard, but they provide :-separated subtypes that are non-standardized and created on demand.

The UD community has been (and continues to be) going through several iterations in the attempt to develop more consistent guidelines and cross-linguistically valid generalizations, e.g., with the recently introduced amendments mechanism, and it provides a standard core vocabulary, but at the moment, neither resource-specific features nor dependency subtypes are properly harmonized with each other nor could be reduced to standard features. Likewise, the “universal” definition of categories sometimes just delegates to language-specific definitions, e.g., VERB:

“Modal verbs may be considered VERB or AUX, depending on their behavior in the given language. Language-specific documentation should specify which verbs are tagged AUX in which contexts. … Depending on language and context, [participles] may be classified as either VERB or ADJ. … Depending on language and context, [some verb forms such as gerunds and infinitives] may be classified as either VERB or NOUN. … Depending on language and context, [verb forms such as converbs (transgressives) or adverbial participles] may be classified as either VERB or ADV. …” (emphasis mine).

The UD community is working hard on reducing such cases and as soon as a consistent axiomatization emerges that is proven to be applicable across languages and theories, OLiA should inherit these axioms by means of equivalence links with UD as an External Reference Model. We provide UD ontologies, and these can be dynamically generated from the UD guidelines, but at the moment, these are not axiomatized and linked as Annotation Models only, i.e., UD concepts are subclasses of OLiA concepts, disjunctions or intersections of OLiA concepts.

Both OLiA and Universal Dependencies (resp., their subsequent sibling projects on morphology, frame semantics, etc.) provide solutions to the interoperability problem of linguistically annotated corpora, but there is a difference in effort and coverage in UD- and OLiA-based interoperability:

In other words, UD and OLiA provide different levels of interoperability that can be described as “shallow” and “deep”:

As UD is an OLiA Annotation Model, all UD-compliant annotations are OLiA-interoperable, but not all OLiA-interoperable annotations are UD-compliant. But even for annotations that do not share the specific theoretical standpoint of UD, these can be compared with (and mapped to or grouped with) those from other corpora if they formalize the same distinction.

Deep (UD) interoperability entails shallow (OLiA) interoperability, or: OLiA extends the (deep) interoperability established between UD resources with a level of interoperability that these have with non-UD resources.

The situation is similar with respect to lexical resources, where LexInfo serves a function for machine-readable dictionaries (published as OntoLex data). LexInfo is a standard for grammatical (and other) features in lexical resources, so that it corresponds to the function of UD for dependency corpora. Like UD, LexInfo is integrated as an annotation model (domain model) into the OLiA architecture. The situation is somewhat different in that

In comparison to UD and LexInfo, a unique contribution of OLiA is that it establishes (shallow) interoperability between annotations and dictionaries, so that a dictionary can be used to produce or verify corpus annotations, or that a corpus can be used to bootstrap a dictionary or quantify the distribution of certain phenomena. In parts, this is inherited from the 1998 Expert Advisory Group on Language Engineering Standards (EAGLES), as these are also (indirectly) underlying UD (FEATS annotations; partially modelled after MULTEXT-East specifications from EAGLES) and LexInfo (from ISOcat/LMF via Parole-Simple from EAGLES), but these hidden ties between UD and LexInfo are neither documented nor even widely known in either of these communities.

Papers and Publications (2007)

Note: to be updated

Author Title Published in Year
Chiarcos, C. An Ontology of Linguistic Annotation: Word Classes and Morphology. In Proceedings DIALOG 2007, Bekasovo/Moscow, May 30 – June 3, 2007, p.630-637. 2007
Lehmberg, T., Chiarcos, C., Hinrichs, E., Rehm, G. & Witt, A. Collecting Legally Relevant Metadata by Means of a Decision-Tree-Based Questionnaire System. In Proceedings of Digital Humanities 2007, University of Illinois, Urbana-Champaign, USA. 2007
Lehmberg, T., Chiarcos, C., Rehm, G. & Witt, A. Rechtsfragen bei der Nutzung und Weitergabe linguistischer Daten. In Georg Rehm, Andreas Witt, Lothar Lemnitzer (eds.), Data Structures for Linguistic Resources and Applications. Proceedings of the Biennial GLDV Conference 2007, Tübingen/Germany, April 11-13, 2007. Gunter Narr: Tübingen, p.93-102. 2007
Krasavina, O., Chiarcos, C. & Zalmanov, D. Aspects of topicality in the use of demonstrative expressions in German, English and Russian. In António Branco, Tony McEnery, Ruslan Mitkov and Fátima Silva (Eds.), Proc. 6th Discourse Anaphora and Anaphor Resolution Colloquium (DAARC-2007), Lagos (Algarve)/Portugal, March 29-30, 2007, p.53-58. 2007
Chiarcos, C. Semimanuelle Generierung und Auswertung von Alternativentexten. In Hardarik Blühdorn, Eva Breindl, Ulrich Waßner (Eds.), Text – Verstehen. Grammatik und darüber hinaus. Institut für Deutsche Sprache. Jahrbuch 2005. De Gruyter, Berlin, New York, 2006, p.406-410. 2006
Chiarcos, C. Sustainability of Linguistic Data. In Proceedings of the 1st International Conference of SFB632: Information structure between linguistic theory and empirical methods. June 6-8, 2006, Potsdam, p. 161-166. 2006
Chiarcos, C. Avoiding Data Graveyards: Deriving an Ontology for Accessing Heterogeneous Data Collections. In Proceedings of the International Workshop „Ontologies in Text Technology (OTT'06). Approaches to Extract Semantic Knowledge from Syntactic Information“. September 28-29, 2006, Osnabrück, Germany, p.113-118. 2006
Chiarcos, C. An Ontology for Heterogeneous Data Collections. In Proceedings of the Int. Conference “Corpus Linguistics 2006”, October 10–14, 2006, St.-Petersburg, St.-Petersburg University Press, p. 373-380. 2006
Schmidt, Th., Chiarcos, C., Lehmberg, T., Rehm, G., Witt, A. & Hinrichs, E. Avoiding Data Graveyards: From Heterogeneous Data Collected in Multiple Research Projects to Sustainable Linguistic Resources. In Proceedings of the E-MELD 2006 Workshop on Digital Language Documentation: Tools and Standards – The State of the Art, Michigan State University in East Lansing, Michigan, June 2006. 2006