Providers of products and services are faced with the dual challenge of supporting the languages and individual needs of the global customer while also accommodating the increasing relevance of user-generated content. As a result, the content and localisation industries must now evolve rapidly from manually processing predicable content which arrives in large jobs to the highly automated processing of streams of fast moving, heterogeneous and unpredictable content. This requires a new generation of digital content management technologies that combine the agile flow of content from developers to localisers and consumers with the data-driven language technologies needed to handle the volume of content required to feed the demands of global markets. Data-driven technologies such as statistical machine translation, cross-lingual information retrieval, sentiment analysis and automatic speech recognition, all rely on high quality training content, which in turn must be continually harvested based on the human quality judgments made across the end-to-end content processing flow. This paper presents the motivation, approach and initial semantic models of a collection of research demonstrators where they represent a part of, or a step towards, documenting in a semantic model the multi-lingual semantic web.
To engage successfully with global markets, enterprises increasingly need to manage fast moving streams of multi-lingual content from both within the enterprise and from the wider user communities with which they wish to engage. For example, modern software markets are increasingly dominated by larger numbers of finegrained applications, e.g. smartphone and social network ‘apps’ or Software-as-aService (SaaS) offerings, that feature high frequency/ “perpetual beta” release cycles. For these products, technical documentation increasingly adopts the form of FAQs, blogs and wikis that grow with significant input from users and customer support staff as features and problems are discovered and solutions documented. Even when technical manuals are provided (and localised into different target markets), users increasingly find more direct solutions to problems on user-generated question & answer sites, which may then themselves, merit localisation based on demand.
Managing this multilingual content stream requires seamless integration of content management systems, the localisation chain and data-driven natural language technologies. Cost, timeliness and quality trade-offs need to be actively managed for different content flows with a far greater level of flexibility and automation than that that has been achieved previously to cover the range of multilingual content generation and consumption paths, e.g. from online documentation, to Q-A fora and micro-blog and RSS feeds.
The Centre for Next Generation Localisation (CNGL) specifically consists of over 100 researchers from academia and industry who conduct research into the integration of natural language technologies, localisation and digital content management required in addressing these challenges. Researchers work collaboratively in CNGL to produce a range of technical demonstrators that integrate multiple forms of multilingual content. A key challenge associated with such large-scale research systems integration is the need for researchers to collaborate and for software to interoperate. However, the components are derived from a number of different research and industrial communities, where either meta-data was not formally defined or was specified from a fragmented set of standards or industry specifications. CNGL therefore established a meta-data group (MDG) to concentrate and integrate the metadata expertise from these different research areas, including statistical machine translation and text analytics research, adaptive content and personalisation research and localisation workflow and interoperability. To address the universal trend of content to be web based and to offer a well-supported, community-neutral approach to semantic modelling, the standardised languages of the W3C Semantic Web initiative were used. This allowed multiple existing meta-data standards and component metadata requirements to be incorporated into a single model thereby demonstrating the interrelation and utility of such interlinked meta-data. This paper presents the initial process for the development of such semantic models for interoperation within a large software research project, such as the CNGL. Future research will see these models are deployed and evaluated in multiple industrial settings to establish their broader applicability in the field of localisation and digital content management. This paper only provides a discussion around forming these initial models. 2.
There is an increasing focus on the interoperability of Content Management Systems (CMS), the tools in the localisation chains and the emerging array of language technologies offered as services. In the localisation industry there are many different platforms, favoured by various Language Service Providers (LSPs) for translation projects. Some are in-house deployments, some open-source (such as GlobalSight) and some purchasable from third parties (such as SDL WorldServer). However each platform comes with its own nuances, meta-data specification and translation workflow management tools. For example within Digital Content Management multiple approaches to storing meta-data apply ranging from storing simple XML content as mark-up (from both authors and consumers) through to complete Ontological models. In Machine Translation meta-data appears in the many forms from terminological mark-up, workflow management meta-data and Translation Memory (TM) that allows previous translations to be recycled in the Machine Translation (MT) process. As has been shown complex meta-data standards occur in many places across both the multi-lingual semantic web and localisation industry.
CNGL develops a large number of distinct “demo” systems integrating different aspects of content management, localization and language technology integration.
Figure 1 summarises how the current range of integrated systems in these categories map onto the NGL Process Map. Semantic models, consisting of content and service models are developed by synthesizing a seed model from an analysis conducted across the range of these integrated systems. This model has then been mapped back onto revisions of the integrated system to assess the degree to which the model can accommodate the natural evolution of these systems as they evolve in response to specific industry requirements and technological advances. The availability of such a large, motivated cohort of collaborating researchers presented a range of distinct views and expertise and allowed us to gather input for and define common models of both the content and services offered across Localisation, Content-management and Integrated Language Technologies. The demonstration systems used address use scenarios in the areas of: Bulk Localisation Workflow (BLW); Personalised Multilingual Customer Care (PMCC) and Personalised Multilingual Social Networking (PMSN). To develop a shared understanding of this modelling, demonstration systems workflows and the components implementing specific activities are overlaid onto a Next Generation Localisation (NGL) Process Map (Figure 1) that identified different stakeholders and abstractions of the major process areas involved in content generation, its preparation for manual and machine processing, the translation, such as translation and personalisation, resulting from this processing and the delivery and consumption (including rating and feedback) by consumers. Semantic models, the content being processed and the services offered by the components doing these processing were then developed by synthesizing a seed model from an analysis conducted across the range of integrated systems and making continual iterations to this model.
However, the breadth of input also put into sharp focus the degree of change these models would be subject to, resulting from continual advances in the research components used, we therefore applied Ontological Evolution techniques to create a common evolving model that is representative of this set of changing viewpoints and that is extensible for use by others in the future. This semantic model has therefore been mapped back onto revisions of the integrated demonstration systems to assess the degree to which the model can accommodate the natural evolution of these systems (which have undergone four coordinated iterations over the last three years) as they change in response to specific industry requirements and technological advances. 3.
The purpose of building common semantic models is driven by the fact that
standardisation efforts in support of content management and localisation processes
and tool interoperability are somewhat fragmented between the Organization for the
Advancement of Structured Information Standards (OASIS), the Open Standards for
Container/Content Allowing Re-use (OSCAR) operated by the (now defunct)
Localisation Industry Standards Association (LISA) and the W3C. Most current
standards are in the form of XML document schema designed to support hand-over of
specific content and meta-data between different tools through import and export
functions using these different file formats. This occurs primarily when a translation
job is handed off from a content developer to an Language Service Provider (LSP) via
their respective Translation Management System (TMS) tools and onwards to an
individual translator’s CAT tool, possibly via a translation agency [
For example LISA’s Translation Memory Exchange (TMX) [
The XML Localization Interchange File Format (XLIFF) [
Another OASIS TC, on an Open Architecture for XML Authoring and
Localization (OAXAL) [
The tendency towards tool vendor lock-in, in terms of both Translation Management Systems and exchange formation, and the associated lack of a strong independent tool vendor market supported by widely adopted common standards.
The need to migrate localisation processes to integrate with increasingly web based, multi-media and multi-modal content that can be personalised to individual user and community needs.
The need to apply localisation and personalisation processes to user generated content and live social media while also actively leveraging the wisdom of the crowd in accelerating those processes. More and more content is originating from rapidly changing sources of content, involving the user in the translation and post-editing of such content is beneficial to the user (being provided access to content in their native tongue) and to the producer in terms of reduced costs associated with translation. Where reduced costs, increased speed, free marketing and delivering of quality assurance from the users themselves form a new approach for companies to translate content.
The lack of interoperability standards that span the Next Generation Localisation (NGL) problem space and the lack of maturity of many of the standards that are presented in various sub-domains.
In order to begin to address these challenges integration in CNGL is based around
the adoption of Service Oriented Architectures (SOA). However, the wide and diverse
scope of CNGL requires strong common models so that meta-data can be easily
shared and services integrated without expensive mismatches in assumptions about
data and state and with clear processing expectations for the use of services.
Although localisation tool vendors are already attempting to support interoperability
through exposing web service APIs to the functionality they offer, these interfaces are
still largely proprietary, though there have been attempts to use XLIFF and TMX file
format as input/output message payloads. Service-oriented integration of language
technology has received some attention also by the research community, including:
the Japanese-funded LanguageGrid project that provided a platform for the
integration of arbitrary compositions in language resources [
However the document-centric nature of existing localisation standards means that their use in web service interfaces for localisation has been ad hoc and lacked any shared conceptual or common conformance framework needed to support seamless integration. In contrast, our semantic model forms a shared conceptual view of services and content that has been derived and validated through a set of software integrations. By providing a core common data type model with a well defined conceptual basis for service, developers can define their interfaces more precisely as contracts, with an easily reached shared understanding of precondition states and processing expectation involved in invoking a given service. At the same time, however, we do not aim to supplant existing interoperability standards where they exist, nor do we intend to be a source of new interoperability standards. Instead we aim to provide a minimal common model for data types that can be exchanged via NGL services, while leaving it to individual service developers to use this type set to define their service interfaces in their particular exchange format. 4.
As the common focus of integration was the processing of digital content, a
content-based semantic model was developed that incorporated a content processing
service model Figure 2 and a processed (i.e. managed) content data model Figure 3.
Both of these models have been developed using the Resource Description
Framework (RDF) [
This approach has provided a more flexible and extensible mechanism for defining
data types that can be exchanged between what we expect to be a large and dynamic
set of NGL services. In addition the semantic models also help to simplify the design
and refinement of content processing workflows in the CNGL research space by
interlinking concepts from existing models and existing standards including Content
management including DocBook [
The seed Semantic Model is broken down into two core parts, these are: Service: High level classifications of the different services covered based on their content processing features and Content: The core processed content model that we use to record the content transformations of various types (including use of content as MT training data) delivered by either activities from the business model or services. These models provide an open mechanism for defining common meta-data from existing software development, modelling tools and persistent data-stores. It therefore offers a practical mechanism for defining a minimal core set of data types that can then be composed or extended as new services are defined and new application requirements arise. A benefit of such an approach towards semantic modelling is that it supports the development of content processing management logs, which due to the emerging nature of the services and applications involved, must be more exploratory in nature, while operational, configuration and tuning requirements are established for datadriven NLP technologies.
During the process of model evolution a parent-child relationship is formed between classes of content, so for example a sub-class of “GenerateContent” is “AuthorText” following a more specific or less specific formation of relationships. Properties are assigned to classes as data-type relationships for example “AuthorText” may have a property “WordCount = int” where users are free in their adaptation of the models to add as many sub-classes, super-classes or class structure changes as they wish. In the Content Model users are able to add as many properties as they wish. All of the changes to all of the models are integrated in one single collaborative session where the meta-data group debates, integrates and records the changes made to the initial seed models. The versions presented in this paper are the initial seed models of both content and services with version 1 soon to be released.
The Service Semantic model, shown in Figure 2, is based around the assumption that all services or process, create or consume content in some form or another. As RDF allows multiple inheritances defining new class types, the schema is defined to allow integration of fundamental aspects of content processing services to define a wide range of services. The core upper level service types include: • • • • •
TransformContent: the transformation of content from one human understandable form to another, including translation, text to speech and content personalised or adaption for delivery to a particular user. AnnotateContent: where additional meta-data is associated with content. ProcessGroupContent: where operations on sets of content are represented. CreateService: allowing the creation of a service to result from a processing chain of other services, therefore allowing the configuration and training of an SMT engine or other data-driven components, or adaptive composition of services to be captured.
Similar to the Service Model, the Content Model, shown in Figure 3, aims to provide an extensible fine-grained schema of types from which the input and output of services that can be modelled. The core type in the content model is ManagedContent, which indicates that we are only interested in content that is subject to some form of management or monitoring. The following key content processing content types are then identified: • GeneratedContent: content produced by a person. • AnalysedContent: content that has been analysed prior to making some decisions on further processing, such as user adaption. • PreparedContent: content that has been altered for further processing. • LocalisedContent: content that has been subject to the localisation processes.
Other seed subclasses have been identified to differentiate content that has been personalised, published or presented to a user or been discovered via information retrieval. Further orthogonal subclasses differentiate: the manual and automatic processing of content, content that is managed as an asset, e.g. a linguistic resource, as well as utility types indicating the content is grouped, time-stamped and serialised as a file, has had its elements counted or some intellectual property right asserted over it. In this seed structure it is important to note that “PreparedForLocalisation” and “PreparedForAdapation” are different in the following way: Localised content is content which is translated and personalised for delivery to a user, Adapted content is only content which is personalised and may not specifically have been translated before delivery to a user. Also important to note is that in terms of both the content and service models these change as the crowd of users adapt them to include the services and content types offered through their research. For example the content model does not include any user generated or live social media content types, however once deployed and adapted by users using a crowd-sourced approach such additions are deemed more likely.
The models presented here have been captured as an ontology using RDF. An
Ontology or semantic model is a specification of a shared conceptualization of a
domain [
In [9 and 10] ontology evolution is defined as: “The timely adaptation of an
ontology to changed patterns of usage in the ontology-based application, as well as
the consistent management and propagation of these changes to dependent elements.”
Ontology evolution takes place when an ontology management system facilitates the
modification of an ontology by applying the proposed changes to the model and by
ensuring its subsequent consistency. The need for ontology evolution is directly
related to the changes that occur in the domain area or in the business environment
and reasons for change include conceptualization, representation or specification of
the domain. State-of-the-art ontology evolution process has six defined phases
[
In a collaborative environment such as CNGL, achieving a common semantic model requires a continuous evolution of models. Users expect the model to support their changing requirements and thus leave the model in a continuous evolution. Researchers applying their technical demonstrators to the current model of services and content adapt the models presented in this paper and these changes are incorporated into the seed models producing iterative versions of common, CNGL wide, agreed semantic models, the first full versions soon to be published. The approach taken, shown in Figure 4, to populating the semantic models, from the collection of demonstrator systems, across the CNGL, involved producing a seed semantic model applying the model to each individual demo team. Incorporating feedback from the collection of demo teams into the seed models and extending the semantic model through a process of continual evolution using a panel of domain experts (the meta-data group) to enforce chances. In this approach, the need for evolving the semantic model comes directly from the demonstrators.
In the process of building the model, new services and content types emerged from the demo teams. To incorporate these new services and content types, it is essential for the models to evolve. This evolution occurs whilst incorporating feedback from seventeen demo teams through an interview approach to data collection. A total of 43 change requests were made to the NGL service model (12 changes) and NGL content model (31 changes). These requests included the addition of new content types which were not captured in the seed model, specialization of existing concepts into two or more types, generalization of categories to a single type and renaming of existing types. There were no requests made for deletion of a given content or service type, however in the process of implementing the above changes, for example renaming, there were deletions introduced. In addition to working on the semantic models of services and content properties were incorporated into the content model 44 of which were identified. Some example of additions made in populating the models include:
• Add subclass (user feedback, managed content) • Add subclass(user opinion, user feedback) • Add subclass(user input, user feedback) Requested service changes: • Add subclass(rank TM, process group content) • Add subclass(compare content, process Group content) • Add subclass( classify audio, annotate content) Requested properties: • Add Property (source Language) • Add Property (input Language) • Add property (number Of Sentences)
Localization is moving more and more towards Content Management in terms of the process pipeline apparent in taking a piece of content and turning it from source language to destination. As the breadth and depth of the content being localized increases there becomes a more apparent need to move towards a common semantic model of the content being processed. This model needs to be both human understandable but also normative, and for this reason it is argued that RDF is a suitable candidate for building models of a rapidly expanding content set. This paper has presented an approach currently being utilized for building common semantic models of the services and content types from a number of demonstrator systems focusing on localization, translation, natural language processing technologies and digital content management. It is argued that such a collaborative approach to gathering and defining semantic models of services and content provision is vital in the future of the multi-lingual semantic web as it attempts to span the different related research and industrial communities.