Abstract
In the last years a rapid emergence of lexical resources has evolved in the Semantic Web. Whereas most of the linguistic information is already machine-readable, we found that morphological information is mostly absent or only contained in semi-structured strings. An integration of morphemic data has not yet been undertaken due to the lack of existing domain-specific ontologies and explicit morphemic data. In this paper, we present the Multilingual Morpheme Ontology called MMoOn Core which can be regarded as the first comprehensive ontology for the linguistic domain of morphological language data. It will be described how crucial concepts like morphs, morphemes, word forms and meanings are represented and interrelated and how language-specific morpheme inventories can be created as a new possibility of morphological datasets. The aim of the MMoOn Core ontology is to serve as a shared semantic model for linguists and NLP researchers alike to enable the creation, conversion, exchange, reuse and enrichment of morphological language data across different data-dependent language sciences. Therefore, various use cases are illustrated to draw attention to the cross-disciplinary potential which can be realized with the MMoOn Core ontology in the context of the existing Linguistic Linked Data research landscape.
Keywords
Introduction
Morphological language data (MLD) plays a crucial role across various interdisciplinary research fields. Traditionally, linguists have fundamentally studied morphology on both language-independent and language-specific levels for centuries in order to investigate the underlying mechanisms that a) allow new words to emerge that are not yet recorded in dictionaries (i.e. word formation), b) are required to alter words so that they take the appropriate form within a certain syntactic environment (i.e. inflection) and c) explain to what extent languages structurally differ in encoding lexical or grammatical meanings within words (i.e. comparative linguistics). This work is the basis for the far younger research field of natural language processing (NLP) which strives to apply linguistic knowledge on morphology (in conjunction with other linguistic areas) on large amounts of text in order to automatically analyze, process or create natural language content. While the methods and aims of linguistics and NLP differ, both sciences can highly benefit each other. Within an ideal cycle of interdisciplinary exchange NLP would take the insights on morphology provided by linguists, apply them to large amounts of text and feed back their results to the linguists who could refine their studies on morphology, which in turn would lead to a better research basis that can be taken up by NLP research again.
Both research fields heavily rely on MLD. The realization of the described scientific exchange and advancement is, however, prevented because of the existing data silos on both sides which use many different and non-interoperable data formats, thus, impeding an easy data transfer. Due to the emergence of Semantic Web technologies this state can change. Being based on the principles of Linked Data, they have proven to evoke true data-driven interdisciplinarity for research domains shared by different sciences. This research manifests itself in the area of Linguistic Linked Open Data (LLOD) which was initiated in 2010 with the foundation of the Working Group on Open Data in Linguistics (OWLG) [9,41]. Since then a significant rise of language data on the Semantic Web emerged. Academic, industrial and technological interest into Linguistic Linked Data appeared and materialized in three areas: (1) W3C community groups such as Linked Data for Language Technology (LD4LT)1
A cross-disciplinary usage of LLOD has already been proven to be achievable in the case of the OntoLex-lemon model9
Cf. the emergence and development of the Linguistic Linked Open Data Cloud:
Extracting and explicating the morphological semantics of words, however, requires not only a domain expert with detailed linguistic knowledge about morphology but also close to native-speaker level knowledge about the language. Even though ontologies such as the OntoLex-lemon model [40], LexInfo [11], OLiA [8], or GOLD [18] partially define a minimal RDF vocabulary to describe morphemes and morphological data as such, a dedicated morpheme ontology capturing and formalizing semantics is still missing.
This becomes obvious through the fact that morphological information is predominantly still attached to the lexeme (the unit that carries lexical meaning) or the whole word form (cf. Example 111
This paper follows the generic style rules for linguistic [24]. This means that italics are used for all object-language forms (words and morphs) that are cited within the text or examples and single quotation marks are used for indicating linguistic meanings (morphemes).
Word form: players
Annotation: NNP
Meaning: ‘noun, plural, common’12
Taken from the Lancaster tagset:
Word form: players
Morphs: player-s
Morphemes: ‘player’-PL13
In contrast to digital and Linked Data dictionaries or lexicons, morphemic language resources are mostly available in layout-centric formats, such as HTML website contents, PDF documents, tables or even only in printed media. What is more, the domain of morphology is to a large extent treated by linguists who do not only differ in their understanding of this linguistic area but also compile morphological data with a focus on consumption by humans and not on machine processability. The creation of the MMoOn Core model consequently strives to tackle these challenges and will add the following contributions:
Provide a fine-grained and extensive semantic model for representing MLD suitable for linguistic and NLP tasks.
Publication of MMoOn Core as a language-independent conceptualization of the MLD domain as a freely available, reusable and extensible linguistic resource.
Linking of MMoOn Core to already existing linguistic data models.
First compilation of derivational meanings.
Representation of morphemic glosses as Linked Data.
Usage of MMoOn Core as a unifying building block to compile language-specific morpheme inventories which:
integrate heterogeneous data sources with semantic consistency,
provide resource descriptions for word forms and morphemic language data,
interrelate language elements across the morph, word form and lexeme level,
include direct extensions of the vocabulary with language-specific meanings,
are automatically multilingually interconnected through an underlying shared semantic,
result in a compilation of natural language data in a machine-readable manner by adhering to Linked Data principles and interlinking.
The remainder of the paper is structured as follows: Section 2 states the motivation and background and is followed by an outline of related work in Section 3, also pointing to gaps in existing resources. After introducing a brief domain analysis in Section 4, the main part of the paper – the Multilingual Morpheme Ontology – will be presented in detail in Section 5. This part includes its architectural setup, design principles as well as its basic elements. A more detailed comparison of MMoOn Core to OntoLex-lemon is provided in Section 6 by taking a closer look at the currently developing morphology module. Furthermore, use cases for the application of MMoOn Core for linguistic and NLP research will be outlined in Section 7. Finally, the paper closes with concluding remarks and a prospect of the future work in Sections 8 and 9.
The need for the development of a data model that is able to describe the morphemic inventories of natural languages was expressed by two major research communities. The first one centers around the community groups OWLG,14
The second group of researchers involves linguists whose main subject area is the investigation of natural language per se. Especially linguists who document endangered and under-resourced languages as well as general comparative linguists both produce and rely on adequate linguistic data. A rising awareness of methodological standards in the compilation of language data has emerged in linguistic research “for the sake of [the] speech communities [of languages threatened by extinction] and their interest in their cultural tradition and for the sake of the very database of the discipline itself” [36]. In linguistics the usage of interlinear or morpheme-by-morpheme glosses as a means for the representation of the segments and meanings of text are an established common practice. Due to their widespread application, efforts of standardization have been introduced [13,37]. As a result, a great amount of interlinear-glossed text resources exist in linguistic databases or as text examples in linguistic publications. Unfortunately, this wealth of data is not easily accessible or reusable due to the (1) technical heterogeneity, (2) license restrictions or unavailability of licenses, and (3) nonformal description of linguistic documentation. Here, the field of linguistic documentation is in need of a model that allows for the (automatic) creation, retrieval, processing and publishing of its morphological data in compliance with the granularity of the linguistic representation levels.
In order to fulfill the demands of both research communities just outlined, the MMoOn Core ontology has been created. It presents a new vocabulary which is easily integrable into already existing lexical resources and expressive enough to capture the various correspondences between subword elements and their associated meanings. Hence, all specific MMoOn language inventories will contribute to the development of natural language analyzing methods and tools. At the same time, MMoOn allows linguists to adequately represent their high-quality language data using a vocabulary with well-defined semantics and in a data format that ensures interoperability with a large range of formats and systems. Thus, we believe that, both the NLP research area and linguistics as an empiric discipline will benefit from the reuse of the MMoOn Core vocabulary.
The developmental approach underlying the creation process of the MMoOn Core ontology is grounded in a thorough domain analysis (cf. Section 4) and guided by a defined set of requirements as well as design choices (which are explained in detail in Section 5.3). To this extent, it has been developed from scratch as a standalone ontology without originating from any existing vocabulary or model. On the contrary, the aim of the MMoOn Core ontology is to unite morphological data represented in differing formats or underlying varying linguistic theories and descriptions. Since MMoOn Core further pursues the aim to function as a language-independent domain ontology for MLD, the generalizable elements, relations and characteristics which have been identified for the linguistic research field of morphology [4,25] have been derived and transformed within the semantic modeling of the ontology. These include linguistic concepts such as affix, inflection, derivation, segmentation, meaning or interlinear glossing as described in the foundational linguistic works about morphology and are not only assumed to be applicable to a wide range of languages but also to be familiar concepts to linguists. Under consideration that linguists create the most fine-grained MLD, MMoOn Core is motivated by the provision of as many descriptive domain elements as possible to keep the entry barrier into working with RDF for linguists as low as possible. To conclude, the MMoOn Core ontology can be regarded as the first extensive representation model for MLD to create inventories of the smallest meaningful elements of language similar to dictionaries or lexical databases within the lexical data domain.
An inventory of morphemes requires an appropriate data model on the one side and morphemic data on the other side. In what follows an overview will be given that investigates the applicability of existing linguistic ontologies as well as existing Linked Data morphological resources but also datasets and sources that are based on other formats.
Vocabularies modeling MLD
Within the last few years, ontologies emerged that contain vocabularies partially describing morphological aspects of language. These include the lemon model [39] and the decomp and ontolex submodules of the OntoLex-lemon model [40], LexInfo [11], OLiA [8] and the GOLD [18] ontology. Even though, none of these vocabularies were explicitly designed to capture the domain of MLD, they include conceptual information on the meaning side of morphemes and/or information of morphemic elements. For that reason the MMoOn Core ontology has been interlinked to some of these vocabularies (cf. Section 5 and Section 6) in order to comply to the Semantic Web best practices for reusing existing data models. In this context LexInfo, OLiA and GOLD are mainly reusable as terminological datasets providing the theoretical description of the linguistic concepts involved in lexicography and morphology.
With regard to the representation of subword units lemon and OntoLex-lemon provide elements that belong to the domain of MLD. Lemon was the first model to offer a morphology module17
The latest advancement in modeling MLD is presented in the W3C report of the OntoLex-lemon model specification.19
The recently published Ligt vocabulary has to be mentioned as a possibility for representing morphological data as well [10]. It is specialized to enable the transformation of interlinear glossed text into RDF data. In particular, it can be used to transform resources based on Toolbox, FLEx and Xigt (eXtensible Interlinear Glossed Text) to Ligt-RDF. The main contribution of Ligt is the unification of several heterogeneous interlinear glossed text resources based on different formats within a homogeneous RDF data graph. With respect to its usability for representing MLD, however, the Ligt vocabulary differs fundamentally from MMoOn Core and OntoLex-lemon in that the morphemic elements it describes identify single occurrences of morphs within an interlinear text similar to tokens within a corpus. As a result, the only element relevant for the domain of MLD in Ligt is the class
So far, two datasets have been created and published based on the MMoOn Core model and architecture (cf. Section 5.2), i.e. the Hebrew Morpheme Inventory [34] and the Xhosa RDF dataset [5] together with a dictionary alignment to Kalanga and Ndebele lexical datasets [17].
To the best of our knowledge, all other existing Linked Data resources including MLD are based on the lemon/LIAM model or the OntoLex-lemon model. As a consequence, these datasets contain morphological data only to a limited extent, e.g. the decomposition of compounds or unrelated affix resources (e.g. [16]).
As a specific example for a dataset containing inflectional language data, the Dbnary “morpho” Wiktionary extractions for German, French, English and Serbo-Croation need to be mentioned.20
Due to the fact that the Linked Data paradigm is in comparison to linguistic research and documentation very young, it is not surprising that the majority of MLD exists in non-Linked Data formats. In fact, the largest part of linguistic data is preserved in documents. However, this overview of MLD will not touch upon such data in unstructured formats but focuses on structured data only. Among the datasets which can be found a high variance with regard to aspects like accessibility, data quality, reusability, complexity of morphological data, covered languages and data format can be observed:
a) MLD in linguistic field work data: This kind of data entails fine-grained, complex and segmented MLD documented in interlinear glossed texts that are edited with specific tools like FieldWorks21
Even though the datasets published by Dictionaria are also provided in RDF, this information is omitted here because no standard vocabulary for linguistic Linked Data has been used and only a part of the original data is transformed into RDF, i.e. only the headwords encoded in literals. Instead, very basic vocabularies such as SKOS and DCTERMS have been used. As a consequence, the morphological data that is entailed in the original source dataset is either missing completely or not differentiable from the lexical data within the delivered RDF datasets.
b) MLD as a part of large language databases: For large and well documented languages usually more linguistic data is available to date. Whole research groups and institutes are devoted to collecting and editing resources such as word lists, dictionaries and corpora and also strive to organize and manage all the linguistic data available in large databases. These datasets also cover MLD like word forms, inflection tables and affix lists. These language resources are the outcome of a collaborative work between linguists and computer linguists that merge and structure manually compiled data as well as automatically transformed or created language data. Examples include the Oxford Online Database of Romance Verb Morphology,25
Atelier pour les LEXiques INformatiques et leur Acquisition,
In this context, the Lexical Markup Framework (LMF) [20,21] has to be mentioned as well. It enables the representation of machine-readable dictionaries (MRD) and NLP lexicons and has been applied to create numerous datasets, e.g. ALEXINA, including morphological data based on the morphological extension of the LMF core model. It provides two strategies for representing word forms. The first one applies to an extensional listing of all forms of a lexical entry which are specified for linguistic categories and values. This approach, however, does not explicitly contain morphemes. The second strategy allows for an intensional modeling of so called morphological patterns and inflectional paradigms. These are formalized in detail and specific to lexical entries, however, with no explicit listing of the forms in the lexicon. While the usage of the morphological extension of LMF is very powerful in terms of machine-processing, it is less suitable as a human-understandable basis for a linguistic analysis of the morphology of a language. The lexicon-centric view on morphology additionally reduces morphology to the lexical entry level and impedes the identification of the smallest meaning bearing units of a language on the word form level. Moreover, LMF-based databases are often realized in structured formats such as XML and very customized. As a result, a considerable effort to understand the data is required and a direct data reuse and interoperability is, therefore, reduced.
c) MLD as morphological segmentation tool output: One of the most challenging tasks in computational linguistics is the creation of segmentation tools. Irrespective of the accuracy and quality of the segmentations, such data outputs also create MLD which can be used in several NLP tasks and linguistic research alike. The IDS developed the Morphisto segmentation tool31
The presented overview of Linked and non-Linked Data resources for MLD illustrates two research fields which develop independently from another, even though, both would increase their scientific outcomes by joining their methods and resources as it has been shown for the domain of lexical language data already. In line with the need for lexical data there is also a demand for morphological data that applies both to the language specific morphological domain requirements and to cross-lingual interoperable data modeling. Given the current state of the art, Linked Data vocabularies are not suitable enough to represent the various existing morphological data that will stay isolated and hard to reuse without the unifying RDF data format.
Overview of the linguistic domain of morphology with the English example lexeme play (verb).
The development of MMoOn Core is based on the following domain analysis for MLD. It has been conducted in order to clarify and decide which linguistic elements and relations need to be represented. The linguistic domain of morphology deals with the internal structure of words including the elements and meanings of which they consist, i.e. the morphs and morphemes of a language. In the context of MMoOn Core we define the term morpheme as the smallest component of a word that contributes some sort of meaning, or a grammatical function to the word to which it belongs, whereas the term morph is defined as the perceivable side, i.e. the written or spoken realization, of a single morpheme. Just as other linguistic domains, e.g. syntax or phonology, the study of morphology can either refer to that part of language in general or to the morphological system of a specific language. For the purpose of outlining the domain this section is concerned with the first sense of morphology, although, the second meaning plays a crucial role when it comes to the description and investigation of the MLD of a specific language.
Figure 1 gives a basic overview of the conceptualization of the domain. It depicts a condensed summary based on linguistic works that outline the area and study of morphology in a general way [4,25] and which can be assumed to portray the common agreement among linguists as to what elements and relations are part of morphology. The word level is divided into
In contrast to word formation, inflection does not result in new lexemes. Rather, it involves the morphological modification of a lexeme in order to use the word form of it in a certain syntactic environment. Consequently, word forms consist of a lexical stem and an
Overall, the domain of morphology is mainly concerned with the identification of the smallest meaning bearing units of language and the investigation of their concrete realization, meaning, function, relation to each other and the systematization of the underlying building (ir)regularities.
MMoOn Core – the Multilingual Morpheme Ontology
Everything developed by us around MMoOn Core can be accessed under the following websites:
Overview of the MMoOn Core main classes and properties.
In the following an overview of the eight main classes and central properties provided in MMoOn Core will be given. Due to the size of the ontology vocabulary it is recommended to additionally consult the ontology file to receive more detailed insights into the definitions and interrelations established between the ontology elements.
Main classes
In order to allow for an easy extension of an existing lexical dataset with morphological data,
Efforts to achieve this goal are currently under development within the OntoLex-lemon morphology module [35].
Moreover, with the two classes
As this overview of the eight main classes shows, the class hierarchies in MMoOn Core are very elaborate. Irrespective of the level of granularity of the source data both the very specific subclasses and the more general superclasses enable the representation, identification and classification of the linguistic elements that are involved in the domain of MLD.
A key feature of modeling the domain of MLD constitutes a sufficient set of relations that is able to capture the segmentation of words. Altogether, the MMoOn Core vocabulary provides 37 object properties which can be used to state more or less specific relations for modeling the morphemic elements of the data that should be represented. Figure 3 illustrates a part of the example data that has been introduced in Fig. 1 by using the most specific properties, i.e. the subproperties which are lowest within the hierarchy of an object property.
Modeling of relations between morphological data with the example segmentation of the word form plays.
In practice, datasets containing morphological data highly differ in terms of coverage and granularity. As a result, the variety of the created object properties emerged because of the intention to increase the applicability of the MMoOn Core vocabulary to as many differing kinds of morphological datasets as possible. This aspect is not trivial, since morphological data does not exist to the same extent as lexical language data and ranges from simple tables containing lexemes, stems and affixes over texts with interlinear morphemic glosses to morphological segmentation tool outputs. In what follows, it will be first outlined how morphological data is ideally expressed with the MMoOn Core vocabulary and second, further possibilities for deviating data representations will be motivated.
An ideal MMoOn-based dataset contains instances of the three main classes
The grey areas in Fig. 3 illustrate how the instances of the three main classes in this ideal modeling can be further described to represent word, morph and morpheme data.
On the word level the interrelation between different types of words can be stated. Word form resources are always interconnected to lexemes by using the property
The segmentation of word forms into morphs consists only of stem or root and inflectional morph segments. Derivation and compounding relations are expressed between
The segmentation into derivational affixes takes place on the lexeme level. Therefore, in Fig. 3 the derivational morph
On the morph level the
The introduced one-to-one correspondence between morphs and morphemes enables the identification of allomorphs and homonymous morphs in the data. All
Even though this modeling choice requires a numbering of
On the morpheme level the meanings that are encoded by the morphs are assigned to the
However, comprehensive datasets containing resources that are involved in the blue graph just explained are rather an exception. Especially the
The decision to define not only
The presented overview of the MMoOn Core object properties illustrated the possibilities of their usage for representing MLD of different complexity and coverage. On this basis MLD (if newly created) can be modeled according to the ideal graph just exemplified or (if covering only a part of the domain data) extended later on to include more fine-grained MLD. It shall be noted that datasets containing morpheme, morph, word form and lexeme resources that are interconnected in the most granular way will allow to derive the greatest insights into the morphological elements and structures of a specific language that is represented with the MMoOn Core vocabulary.
Given the complexity of the MMoOn Core ontology the question arises how language-specific MMoOn morpheme inventories are meant to be built. Therefore, an integrational architectural setup (cf. Fig. 4) has been developed which interconnects the language data of each morpheme inventory with MMoOn Core and, thus, ensures the multilinguality of all MMoOn datasets. The architectural setup comprises three data layers that serve to cover the following two aspects of linguistic data, i.e. 1) the difference between primary and secondary language data and 2) their description by assuming either language-independent or language-specific linguistic categories. The first aspect is based on the general assumption that most linguistic datasets comprise primary as well as secondary language data [36]. The former data type is defined here as language data which originates from a certain text compilation or could be applied to any text or token in order to identify the word forms, morphs and morphemes of the morpheme inventory. The latter is then defined as the kind of data which enables the description of the primary language data. E.g. the German plural suffix
Architectural setup of MMoOn morpheme inventories exemplified with morphological German and Hebrew data.
In what follows, the three data layers of MMoOn morpheme inventories will be described and how they allow to model primary and secondary language data simultaneously in the context of a language-independent data model that subsumes and interrelates language-specific data.
The first layer builds the MMoOn Core ontology as the underlying formal and conceptual model shared by all morpheme inventories. Since it models the domain of morphology as a subfield of the study of language it functions at the
The second layer in the architectural setup builds the
The template file can be downloaded here:
The largest part of each dataset constitutes the primary language data. Within the architectural setup it is realized by instances on the
In sum, the aim of this architectural setup is to create a unified multilingual data graph of all MMoOn morpheme inventories to come. The presented layers correspond in practice to three RDF files, i.e. mmoon.ttl, schema.ttl and inventory.ttl.38
Usually the schema and inventory files are specified for the language of the morpheme inventory, e.g. deu_schema, heb_schema and deu_inventory, heb_inventory in Fig. 3.
1) Facilitated multilingual Linked Data usage: Due to the unifying function of the MMoOn Core model language-specific instance data of different languages can be cross-linguistically traversed through a single data graph.
2) Exploitation of linguistic data in NLP tasks for linguistics and vice versa: The rather flat structured language data NLP systems rely on could be supported and extended by also taking fine-grained linguistic data into account to arrive at more stable data-driven approaches. Conversely, empirical linguistic research benefits from vast amounts of language data that can be collected in a structured way with NLP methods, which in turn, can serve as a starting point to create more accurate and interrelated linguistic datasets.39
For more details on how the MMoOn dataset creation setup is involved here, cf. Section 7.2.
3) Enable onomasiological and semasiological data retrieval: Most linguistic datasets only allow for unidirectional data retrieval. A MMoOn morpheme inventory, however, is more flexible in this respect. Because it provides the means to represent the association of a linguistic meaning with its language-specific expression within the same model, the meanings a certain morph or word form encodes as well as the kind of morphemic expressions that are used to encode a certain meaning can be retrieved simultaneously.
4) Development of a meta-collection of linguistic concepts: Every MMoOn Core based language-specific schema ontology automatically adds to the extension of the MMoOn Core
The creation of a domain ontology is guided by several influencing aspects ranging from the granularity of the domain representation, the intended usage of the resulting datasets and possible user groups to the choice of the vocabulary as well as the technical possibilities and limitations of the data format. Thus, modeling the MMoOn Core ontology entailed several design decisions. In order to comprehend the motivations that accompanied the development of MMoOn Core, the design principles and determining domain requirements will be outlined in what follows.
Design principles for the domain of MLD
Domain delimitation: The elements and relations of the domain of MLD in MMoOn Core are based on the domain analysis as outlined in Section 4. However, some of the included linguistic elements such as lexemes, word forms and morphs overlap with other linguistic domains, e.g. lexicography, phonology and syntax. Study areas like morphophonology and morphosyntax indicate that basic linguistic concepts are considered to be part of several linguistic domains depending on their defined characteristics and functions. As a consequence, the domain of morphology can be either described in a very strict way, ignoring possible domain interrelations or in a broader way which would result in an overlap with other domains. The MMoOn Core model takes up the strict approach and, thus, provides anything that is necessary to describe words and the meaningful segmentable subword elements of which they consist. Accordingly, the mentioned overlapping elements are not further specified for postulated functions and usages in other linguistic domains. In that respect, the model strives to be as detailed as possible (on a language-comparative level) and as broad as necessary at the same time. Therefore, MMoOn Core constitutes a quite narrow and fine-grained vocabulary for the domain of MLD but also provides prominent classes, such as
Framework neutrality: Even though no model comes without any predisposition, MMoOn Core aims at completeness and a comprehensive application rather than fitting the descriptive needs of a certain linguistic framework, model or theory of morphology. It is a first proposal of modeling MLD comprising the relevant categories and relations in order to extend and integrate morphemic data into already existing linguistic datasets which are mainly framework neutral models as well. However, if required, the MMoOn Core vocabulary is easily adjustable so that the data that shall be represented is integrable according to strict theoretical descriptive needs.
Modeling of linguistic concepts and categories: One of the main challenges when it comes to the description of language data is the choice and modeling of the concepts for linguistic categories. A highly controversial debate exists among the linguistic research community about the treatment of concepts such as ‘case’, ‘gender’ or ‘noun’ as being interlingual comparative or language-specific descriptive categories (cf. for example [23] and [38]). Given that MMoOn Core serves as an upper ontology to create language-specific morpheme inventories both kinds of concepts needed to be considered. Due to the RDF format this particular issue could be solved by adhering to the Semantic Web’s standard which already entails the representation of commonality and variability through the hierarchy of classes [1]. In line with this, MMoOn Core classes are regarded as prototypical interlingual concepts and consequently function as the least common denominator for a linguistic category. Every instance of the classes is then a language-specific concept of the upper interlingual MMoOn class concept as described in the setup of the language-specific schema file. According to this, MLD of different MMoOn morpheme inventories can be described with all language-specific features while staying comparable because of the shared MMoOn class membership. As a result, all MMoOn Core based datasets will contribute to a multilingual data graph of interconnected MLD of specific languages.
Coverage: The MMoOn Core model covers concepts and relations that are necessary for synchronic language description, i.e. the representations and meanings of the words, morphs and morphemes are given according to a certain point in time (present or past). Thus, etymological and historical information is not considered in the class or property modeling. As Section 5.1 outlined, MMoOn Core encompasses a fine-grained vocabulary that enables the identification and description of linguistic elements that are necessary for representing MLD. Also, a considerable set of object properties allows for a detailed specification of the relations that hold among the words and the morphemes and morphs of which they consist. As mentioned before, the morphological rules underlying the data are not considered explicitly and need to be inferred indirectly from the data or have to be described by using another vocabulary along with MMoOn Core. The main approach pursued provides granular descriptive means for the morph and morpheme elements and their interrelations to word elements by outsourcing granular phonological, lexicographic or syntactic concepts at the same time. This is not seen as a disadvantage because including them would entail the preference of some theoretical framework which is meant to be avoided.
Target user groups: The use of the MMoOn Core model is directed towards linguists, computational linguists, NLP researchers, lexicographers and anyone who has an interest in compiling and managing MLD. It is anticipated that MMoOn language inventories will be set up by data compilers of the various user groups mentioned. That way synergies can evolve between the smaller but high-quality and mainly manually compiled datasets that are expected from the linguists and the large but not as fine-grained data produced by users with an interest in the machine-processable aspect of linguistic data. The emergence of these cross-disciplinary synergies are assumed to advance the whole LLOD community in general.
Data modeling requirements
Linked Data principles: The choice to model MMoOn Core in the RDF format is motivated by the underlying Linked Data principles [3] which promote the creation of structurally and semantically interoperable datasets. This aspect adheres to the aim of providing a data-unifying domain modeling that is based on technical integrability. Furthermore, due to the creation of unique resources as URIs, the ontology is easily accessible on the Web. Consequently, all emerging MMoOn-based datasets will, therefore, contribute to a growing interconnected data graph and, thus, not join the ranks of the already existing morpheme data silos on the Web.
Reuse: In general it is understood as a good practice to reuse existing vocabularies when creating a new ontology. Since the largest part of the MMoOn Core vocabulary aims at representing meanings, we decided to create a new taxonomy within the
Extensibility: Finally, a data compilation is rarely ever complete and a single domain model can never capture all practical and theoretical aspects of MLD in general and even less the aspects of MLD of single languages. Given these circumstances, the MMoOn Core model serves as a starting point for morphological data description that might be sufficient for a considerable number of datasets, but must be also prepared to allow for necessary extensions and/or adjustments. This requirement is also assured by the Linked Data format meeting these needs by taking up the assumption of an open world [1]. Consequently, the RDF format allows for a liberate reuse of all classes and properties as well as for an unrestricted extension of the model with new classes and properties. It is, however, assumed that the central comprehensive elements are provided by MMoOn Core and shared by the majority of the emerging MMoOn-based datasets.
URI design: As outlined in Section 5.2 every MMoOn morpheme inventory consists of three files with the MMoOn Core ontology being shared by all datasets. In order to facilitate the identification of and navigation through a dataset, the following URI scheme is implemented for all MMoOn datasets created by the authors:
In sum, it appears that the data modeling requirements posed by the morphology domain are very well accomplished by the underlying Linked Data format. The MMoOn Core model as a proposal to start with a homogeneous morphemic data compilation fulfils the needs of a specified linguistic data description model and integrates the resulting data into the Semantic Web environment, thus, benefiting from all of its advantages.
MMoOn and OntoLex-lemon
In contrast to existing ontologies for describing language data, linguistic datasets rarely contain linguistic information that neatly corresponds to one single linguistic domain. The OntoLex-lemon model [40] being a W3C community group specification tackled this issue by covering the domain of lexicology by enabling the representation of related linguistic domains via dedicated submodules. With this modular extensible approach the representation of a wide range of the existing linguistic data can be already realized. Consequently, an all-encompassing vocabulary covering any potential or existing kind of linguistic data point is neither feasible nor desirable. Rather, the development and usage of more fine-grained and specific vocabularies that are interconnected with a commonly shared ontological basis, i.e. OntoLex-lemon, will provide the necessary means to enable an appropriate modeling of existing or future linguistic data as Linked Data.
This holds true especially for the domain of MLD, which tends to include lexical as well as morphological data. Depending on the use case and dataset, OntoLex-lemon, i.e. the ontolex and decomp submodules in particular, may be used for describing MLD. This has been, for instance, done for representing the components of compound words [16]. Nonetheless, as already mentioned in Section 3.1 for linguistic data corresponding to the domain analysis of MLD (cf. Section 4), the ontolex and decomp modules are mostly limited to compositional morphology and, hence, leave the larger part of the MLD domain to be non-expressible with the provided vocabulary.
A comparative overview based on detailed examples that shows how data on the lexeme, word form, morph and morpheme levels can be described by using either OntoLex-lemon or MMoOn Core can be consulted in Klimek 2017 [32]. Here, a list shall suffice that summarizes the main results, i.e. aspects that reach representability through the MMoOn Core vocabulary and which are not covered in OntoLex-lemon respectively:
1) Inflectional affixes: Since inflectional information is usually no central part of lexical data, means to represent inflectional affixes are not part of OntoLex-lemon. In fact, even consistently collected number information for nouns by providing the respective morph together with the lexical entry, is not describable with it. Instances that are allowed within the
2) Stems and roots: Those two elements are crucial for describing MLD, not only for decomposing word forms but also lexical entries. While OntoLex-lemon provides the possibility to identify the underlying stems in compound words only (which are not termed as stems but widely included within the class
3) Morphemic interrelations: Part of the description of morphemic elements is also the representation of their relation to other morphs. Therefore, stating the allomorphs and homonyms of a morph is important for their identification, function and the combinatoric rules that apply to them. While the MMoOn Core vocabulary contains two object properties to specify allomorphy and homonymy between morphs, these relations are not part of the lexical domain and, hence, not expressible with OntoLex-lemon.
4) Morphemes and meanings: Also not part of the lexical domain is the representation of morphemes. Meanings, i.e. lexical senses in OntoLex-lemon, differ largely from the grammatical and derivational meanings that are necessary for describing MLD. The 300 meaning classes provided in MMoOn Core are far from being extensive with regard to the large variety across languages. However, they are a first step towards collecting and documenting meanings that are encoded by morphs and constitute a useful starting point for representing morpheme resources.
As a result of the introduced suggestion to create an interconnection between OntoLex-lemon and MMoOn Core in Klimek 2017 [32] both domain ontologies have been aligned, as already mentioned in Section 5.3.2, with the established subclass relation between
As a result, the creators of MMoOn Core have been invited to lead the development of a new OntoLex-lemon morphology module which is currently under development.41
In what follows, possible usages of the MMoOn Core ontology will be outlined. This serves to exemplarily indicate the research potential it entails for the two application areas of linguistics and NLP it has been designed for. It shall be noted that all mentioned usages are equally realizable with the commonly applied methods of language representation and analysis in these fields. However, special awareness should be given to this Linked Data-based approach of MLD representation by using MMoOn Core (alone or in conjunction with other ontologies) because it yields the benefit of interdisciplinary reuse, extension and application as an opportunity to overcome the current limitations of scientific progress caused by data silos and heterogeneous formats.
Use cases for linguistic research
Enhancement of morphological data in dictionaries
Dictionaries and lexical datasets contain a considerable amount of MLD. This includes derivational morphs and the lexical entries they can be productively combined with but also elements and building patterns of inflectional paradigms that vary in the degree of their descriptive granularity across dictionaries of different languages. In dictionaries of Semitic languages, for instance, headwords are collected around roots which are followed by the full list of word forms but also lexemes which can be derived from them. For the description of such fine-grained morphological data, the creation of MMoOn morpheme inventories enables the representation of this data in an appropriate manner which can serve as an addition to vocabularies that are usually used for representing lexical data. The Hebrew Morpheme Inventory can be seen as a proof for this application of the MMoOn Core ontology [34].
Language acquisition
With the availability of more and more language data the applied linguistic research area of (second) language acquisition is provided with new possibilities for creating language learning materials and tools. Within this setting morphological data plays a significant role for the acquisition of inflection and formation patterns of words. The future morphological datasets, therefore, have the potential to broaden and complement already existing data-driven learning tools and techniques for corpus linguistics [22] with valuable morphological data. Provided by MMoOn morpheme inventories, inflection tables, word families and the grammatical as well as lexical morphs with their usage restrictions can be obtained. In this respect single MMoOn-based datasets can be already regarded as source data for language learning and teaching materials. The created Xhosa RDF dataset [5] is an example for a MMoOn-based dataset with an intended usage for language revitalization efforts for Bantu languages by using the MMoOn Core ontology as the uniting model for collecting interoperable data of multiple Bantu languages [17] to develop various learning materials.
Language documentation
The area of language documentation has the intention to “to provide a comprehensive record of the linguistic practices characteristic of a given speech community” [28]. Since the publication of this paper in 1998, this area has sparked a community which aims to create linguistic resources for endangered and minority languages. As mentioned in Section 3.3, due to the work of the language documentation community, a great amount of interlinear-glossed text resources exist in linguistic databases or as text examples in linguistic publications. However, these linguistic resources do not use the same representation format. Hence, sharing it within and especially outside of this community is difficult. If a language was documented using the MMoOn Core ontology, it would be possible to create other output formats such as tables, dictionaries, etc. That way the resulting language resource could not only be shared with the language documentation community but, moreover, this data would become usable by the NLP and Semantic Web communities to create tools supporting minority languages.
Representation of morphemic glosses in linguistic literature
Morphemic glosses are part of many linguistic publications and usually used in given examples. A standardized set for interlinear morphemic glosses does not exist and each publication is accompanied with a customized list of glosses. Nonetheless, an adoption of the proposed standardized application within the Leipzig Glossing Rules [13] as well as the reuse of the therein provided set of glosses can be observed. However, the majority of glosses being used is still heterogeneous in that different glosses are used for the same morphemic concepts across the literature. The morphemic glosses provided in MMoOn Core can be regarded as a reference set of glosses since MMoOn Core already reused the existing glosses provided within the Leipzig Glossing Rules which are already widely accepted and applied by linguists. Given that the links between all
Comparative linguistics
The internally provided links between the
Use cases for NLP research
Conversion of Wiktionary datasets
The already mentioned MLD provided by Wiktionary (cf. Section 3.2) is one of the largest openly available datasets. In the context of Linked Data-based NLP research it is desirable to create an RDF version of this data. The existing Dbnary morpho dataset is, however, not appropriate for NLP tasks because it covers only four languages, uses an outdated lemon vocabulary and contains only a morphological annotation of the grammatical meanings of the word forms given in the Wiktionary inflection tables. Instead, it seems promising to convert existing data provided by UniMorph [30,31] and paradigm extractions42
Morphological annotation tools could be created with a data-driven approach based on MMoOn datasets similar to the task of part-of-speech tagging. The initially required RDF representation of corpora can be provided by using the Natural Language Processing Interchange Format (NIF) [26,44]. The resulting NIF corpus can be then extended with several layers of annotations depending on the granularity of the interconnected MMoOn dataset. This could range from the identification of lexemes, stems, morphosyntactic meanings and also part-of-speech data on the word form level of the tokens up to the segmentation into their morphs together with the underlying inflectional and derivational meanings on the morph level of the tokens. In any case, the
Named entity recognition
Recent work in the field of named entity recognition (NER) in German has revealed that the complexity of morphology is rarely considered in existing NER tools, even though considering it could lead to improved results [33]. This holds true especially for the identification of NEs (or linguistically termed: proper nouns) which have undergone several morphological transformations and appear within complex lexemes. E.g. in order to retrieve the NE Alpen (engl. ‘the Alps’) within the inflected German noun Skilalpinistinnen (engl. ‘female ski alpinists’) all compositional, derivational and inflectional transformations that have been applied to Alpen have to be deconstructed. But also nontransformed proper nouns that are only obligatory affected by inflectional marking can already pose a challenge for NER tools. Within a German MMoOn morpheme inventory the involved morphs -en, -in(1), -ist, Ski, alp and -in(2) would be available and could help to identify the NE within the common noun. A very elaborate MMoOn dataset could also contain the complete token with its full segmentation, which allows for a direct retrieval of the underlying NE from within the data graph. Since the MMoOn Core ontology enables a comprehensive explication of morphological data, the lack of appropriate morphological data can be overcome. Consequently, future morpheme inventories could be a promising consideration in the development of NER tools and systems.
Machine translation
Machine translation belongs to one of the most complex and challenging tasks in NLP. Dictionaries and lexical data play a crucial role as one of the sources that are utilized for identifying the sense of a word in a text in one language and the respective expressions used for this sense in another language. However, depending on the morphological type of the languages that are to be translated this task is getting increasingly difficult the more the word-to-morpheme ratio deviates from one-to-one correspondences. Machine translation systems that would be complemented by MMoOn-based datasets could rely on the more fine-grained morphological data. This might be especially improving when translating from analytical languages, e.g. Vietnamese, to polysynthetic languages (marking the extremes of the typological continuum) or vice versa. A lexical approach only will not be able to capture for instance sentences like angya-ghlla-ng-yug-tuq, ‘I have a fierce headache’ (Siberian Yupik) [12] because it consists of a single word. Within the MMoOn representation, however, the individual morphs are explicated and could be translated into an isolating or agglutinative language through the senses and grammatical meanings they consist of. Since all MMoOn datasets share the MMoOn Core ontology within the unified graph of a multilingual dataset the atomic morphemes of isolating languages and the fusional morphemes of polysynthetic languages can be identified and translated in an onomasiological way (in contrast to the semasiological approach of lexical data).
Sentiment analysis
Comprehensive MLD also has the potential to contribute to the NLP research field of sentiment analysis. Subjective information about topics within texts is not only encoded lexically but also by morphological means. E.g. the detection of negation, being one of the main issues for sentiment analysis [47], could benefit from a morphological data source such as a MMoOn morpheme inventory because negation can be very productively expressed by using prefixes like un- for English together with adjectives. Furthermore, bound morphemes for comparative, superlative or intensification can be easily retrieved from such a dataset and also identified even if the lexemes they are attached to are unknown. In general, MLD represented with MMoOn can explicitly describe obligatory grammatical and highly productive lexical morphemes that express various concepts relevant for sentiment analysis. Consequently, an integration of MLD in the form of MMoOn morpheme inventories poses a promising application case for extending existing resources, algorithms, models and frameworks in the field of sentiment analysis.
Concluding remarks
The development of the MMoOn Core ontology started in 2015. Since then, the ontology has been evaluated for its applicability resulting in the Hebrew Morpheme Inventory [34] as proof of concept. Simultaneously, the architectural setup has been developed, morphemic glosses and meanings have been extended and refined. The interim status of the ontology has been presented at various scientific events to gain feedback from the target user groups which has been considered and integrated into the final publication state of MMoOn Core as well. Despite this longstanding process from conceptualizing to actually publishing this accompanying article for the MMoOn Core ontology, no comparable advancement in creating a domain ontology for representing MLD is recorded [6].
As far as the vocabulary use of MMoOn Core is concerned, it achieves a four out of the five star ranking of Linked Data vocabulary use [29]. According to this, MMoOn Core contains dereferencable human-readable information about the used vocabulary (1 star), available information as machine-readable explicit axiomatization of the vocabulary (2 stars), a linking to other vocabularies, i.e. OntoLex-lemon (3 stars) and provides metadata about the vocabulary (4 stars). At the current state the fifth star, i.e. vocabularies that link to MMoOn Core, is not achieved. With the awareness that exists already for this domain ontology, however, it is very likely that other vocabularies, e.g. OntoLex-lemon or Ligt will create links to MMoOn Core in the future.
In summary, the presentation of the MMoOn Core ontology in this paper has explained how this model will enable the conversion of existing as well as the creation of new morphological datasets and, thus, reaches its aim of contributing to a rising number of homogenized, interoperable linguistic datasets. This result is mainly based on two characteristics of the ontology. First, the rather unusual granularity of the provided meaning classes and their interlinkings with their respective glosses reduce the time for mapping source data of different formats with the ontology and enhances the consistency across datasets. Being embedded within the whole MMoOn Core ontology, these concepts explicate the large part of the linguistic domain of morphology and, therefore, enable the creation, transformation and semantic enrichment of the of MLD that was hitherto inaccessible for machine-processing, e.g. inflection tables, interlinear glossed text, morphological data accompanying lexical databases and dictionaries. The second crucial characteristic of MMoOn Core is its capacity to strengthen the interdisciplinary reuse of MLD originating from the linguistic, NLP and Semantic Web communities. Due to the architectural setup that is based on MMoOn Core, both, language-independent as well as language-specific representations of MLD can be realized. Therefore, depending on the use case and the intended application of the MLD that shall be described as Linked Data either the MMoOn Core ontology can be used to create a very generic and language-independent morpheme inventory or a language-specific schema file that enables specific extensions. Due to the fact that all emerging MMoOn-based datasets are inherently interconnected through the MMoOn Core ontology, datasets that had been of potential interest for a specific user group but have been eventually rejected for an actual reuse (because they were considered too general or too specific in their description) can be now directly adjusted to the required granularity of the representation needs. In this respect it is through the architectural setup of MMoOn Core that the creation of MLD is enabled not only for different user groups and usages but also that all resulting morpheme inventories are semantically unified, thus, leading to an enhanced interoperability and reusability. To conclude, it could be shown that the MMoOn Core ontology contributes to a facilitated and flexible cross-disciplinarity MLD data generation and exchange.
Future work
Even though the MMoOn Core ontology as it is published now can be regarded as a ready to use domain ontology, it is intended to evolve in the future. Collecting and representing all concepts that can be morphologically expressed across the word’s languages can not be achieved by a few scientists. Therefore, the meanings provided in MMoOn Core can be regarded as a starting point of the ontology which shall be constantly adapted and extended according to emerging MMoOn morpheme inventories and their schema files. Especially the list of derivational meanings is envisioned to be enlarged and integrated into MMoOn Core from the language-specific datasets.
Another prospective step entails to outreach to other LLOD communities in order to strengthen collaborative research. This is desirable in order to reach the most consistent usage of existing linguistic domain models and data since the considerable overlap of linguistic data compilations of different research areas can not be avoided. Given that MMoOn Core presents a further addition to existing ontologies for the representation of linguistic domains it is advisable to reach a shared agreement on aligning phonological, morphological and lexical data by interconnecting PHOIBLE [42], MMoOn Core and OntoLex-lemon respectively.
Similarly, the connection of MMoOn Core and the Ligt ontology will be promoted. In doing so, a higher number of semantically richer morphological datasets from interlinear glossed text sources, especially for less-resourced languages, can be expected in the future.
Finally, work on Linked Data-based solutions for an integration and the transformation of non-RDF resources such as the Typecraft or UniMorph datasets into LLOD based on MMoOn is planned.
Footnotes
Acknowledgement
We acknowledge support from Leipzig University for Open Access Publishing.