The role of natural language is becoming in these years a more and more acknowledged aspect of the Semantic Web. Not limited to mere modeling and representation proposals, but backed by concrete use-cases and scenarios, the use of natural language is emerging through a plethora of approaches and solutions. Now that we have languages and protocols for modeling and publishing content, for querying, and for efficiently describing datasets and repositories (metadata), it is time for natural language to regain its due space and become a first-class-citizen in the Web of Data. Lexical resources need to comply with standard, unifying vocabularies upon which they can be discovered, chosen, evaluated and ultimately queried upon need. At the same time, NLP systems and components should pull their head out of their esoteric corner and become classifiable, discoverable and interactive elements in the Semantic Web, so that many language related tasks can be carried on more easily thanks to coordinating modules/agents sensible to this information. In this position paper, I will provide by first a quick outlook into the last years of language and ontologies and describe what the community has achieved by the state-of-the-art. I will then discuss open points, and try to draw conclusions, based on my perspective and contributions to this research field, towards the future of a more Language-Aware Semantic Web.
More than a decade has passed since the first OntoLex workshops [
Even though no specific criticisms exists in literature, the objectives of this
community are sometimes being claimed by critics and detractors to be unreasonably
complex, looping in a never-ending lack of consensus and providing no benefit to the
Web of Data. However, the role of natural language has become in recent years a more
and more acknowledged aspect of the Semantic Web. Not limited to mere modeling
and representation proposals, but backed by concrete use-cases and scenarios (e.g.
question answering [
Still the question arises: what can we do with a more linguistically-aware Web? Can it contribute to the Web of Data, which is meant to be food for machines, or will it be doomed to consume itself, in a niche of language-oriented tasks?
As we have learned from the slow evolution of the Semantic Web itself, its vision has not lost any of its original power, yet some of the related innovations and problems took time to be acknowledged, or required foundations to be laid before they even started to make sense. For many years, “language” has always been the seasoning added by experts for improving performances and quality in data-intensive tasks such as ontology alignment, question answering over ontologies and knowledge acquisition, decorating them with evocative terms such as “language-based” or “language-driven”, and still being out of any specification on the Web architecture. Now that we have languages and protocols for modeling and publishing content, for querying, and for efficiently describing datasets and repositories (metadata), it is time for natural language to regain its due space and become a first-class citizen in the Web of Data. lexical resources need not only to be represented as Linked Data, but to comply with standard, unifying vocabularies upon which they can be discovered, chosen, evaluated and finally queried upon need. At the same time, NLP systems and components should pull their head out of their esoteric corner, exploit the stream of innovation brought by software provisioning mechanisms such as Maven Repositories and OBR, breed it with Linked Open Data (LOD) principles, and become classifiable, discoverable and interactive elements in the Semantic Web. With such an elaborated web-of-language architecture, those same tasks mentioned above, which today require lot of preprocessing and fine-tuning, should be carried out more easily (if not on-the-fly) thanks to coordinating modules/agents sensitive to the lexical information present in ontologies, datasets, lexical resources and in NLP software and services.
In this position paper, I will provide by first a quick outlook into the last years of language and ontologies and see what the community has achieved by now. I will then discuss open points, and try to draw conclusions, based on my perspective and contributions to this research field, towards the future of a more Language-Aware Semantic Web. 2
The first events in the area of ontologies and lexical resources were the OntoLex series
of workshops [
With the end of the OntoLex experience (Ontolex 2010, [
In parallel with events, many proposals have come out in these years, in terms of models and vocabularies for representing lexical resources, for a better lexical representation of ontologies, and for trying to standardize the plethora of efforts that have been conducted in the NLP area, in the context of linked data.
A number of models have been proposed to enrich ontologies with information about
how vocabulary elements are expressed in different natural languages, including the
Linguistic Watermark framework [
Due to these efforts, concrete proposals emerged and were refined into new ones through experience and discussion, and community group emerged to concretize forthcoming standards for language-related aspects of the Web of Data.
The OntoLex W3C Community Group1 has the goal of providing an agreed-upon
standard by building on the aforementioned models, the designers of which are all
involved in the community group. Additionally, linguists have acknowledged [
These complementary efforts by Semantic Web practitioners and linguists are in fact
converging: the ontology lexicon model being defined in OntoLex provides a principled
way [
By seeing all of these efforts from the very neutral perspective of someone approaching the world of the Web of Data, one question would naturally arise: if the Semantic Web is all about machine-understandable data, why should we care at all about language? Can a more linguistically-aware Web contribute to the Web of Data, which is meant to be food for machines, or will it be doomed to consume itself, in a niche of languageoriented tasks? What can we do with it?
We could admit that large part of the research done in the area is more biased by the (linguistic) background of the researchers working on it, than driven by the general interest in the problems it tries to tackle. What puts in the same category all of the works and efforts described until now, is the scientific and cultural background of the communities working on them. As “Semantic Web” and “Linked Data” are not properly “foundations of science” but applied sectors of fields of computer science such as knowledge representation, knowledge and data management, and many others.., researchers who contributed to any aspect of natural language in the Semantic Web have their foundational background as well. Specifically, they flowed into the semantic river as Computational Linguists (or pure Linguists), or, in any case, language experts who wanted to give their contribution to the Semantic Web cause.
The risk of proposing solutions to issues which do not characterize the addressed scenario is high, and this has probably occurred more than occasionally. For sure, a lot of effort has been spent in addressing specific issues seen with the glasses of traditional computational linguists. Without mentioning the specific works, any short look on the literature of the Semantic Web area will return a plethora of articles on converting the nth lexical resource to the kth OWL porting, or on leveraging existing “less noble” resources to ontologies. Such is the case of the frenzied plethora of articles on lifting the so-called “folksonomies” to ontologies (“folksonomies” which were strangely more recent than the concept of ontology itself). All efforts that, by applying well known techniques for knowledge synthesis, but lacking an overall perspective on the matter, left neither new theoretical foundations nor concrete results (e.g. in term of resources) for the future to come.
On the other side however, as it often happens in research, problems take their time to be acknowledged, or may require foundations to be laid before they even start to make sense. The efforts described in the previous section all aimed at providing general frameworks for describing lexical resources, common loci for finding their descriptions on the Web and metadata for properly selecting them. As well as much desired results for trying to convey order in the plethora of models and processes which years of NLP research have produced at the expenses of any engineering attempt in the field. Together with the above, a few application scenarios have proven the usefulness of a more aware representation of natural language content in the Web of Data. In the next section I will describe the importance of such scenarios in the context of the whole Web of Data, and how they can be improved by a more systematic inclusion of natural language in the Semantic Web architecture.
I describe here three scenarios that prove the advantages brought by a more languageaware approach to the Web of Data. These are, respectively: Linguistic Enrichment of Ontologies Ontology Alignment Knowledge Acquisition 4.1
It may seem redundant and self-referential to prove the importance of a more linguistically-aware Web, by showing a scenario for linguistic enrichment.
Fact is that the provision of natural language descriptors (for how simple and
destructured these can be) for ontological entities is a primary necessity which comes far
before the more elaborated features that we are trying to highlight. From the dawn of
computer programming, providing “meaningful and evocative names” for functions
(and later for classes/objects and methods) is a primary necessity for, to say the least,
improving the readability of source code. In this sense, ontologies and their RDF code
are no different from computer programs, and they need human understandable
descriptors in order to be maintained and evolved. Furthermore, in the case of
ontologies, the connection between concepts and their referents [
A complete ontology development process should thus consider the formal aspects of conceptual knowledge representation, as well as guarantee that the same knowledge be recognizable amongst its multiple expressions that are available on real data. Thus, the importance of properly enriching ontologies with lexical content does not need further support, while what we need to prove is that more attention to the representation of language can improve the quality (and time to realization) of this necessary practice.
Ontology Development tools should support this task, providing users with dedicated interfaces for efficient carrying it out. It appears obvious that the reuse of linguistic resources creates added value, as they can contribute additional descriptors, as well as suggesting additional knowledge which was not considered in the conceptual development phase. Views over lexical resources must be integrated with classic views over knowledge data such as class trees, property and instance lists, offering a set of functionalities for linguistically enriching concepts and, possibly, for building new ontological knowledge starting from linguistic one.
By considering past experiences [
However, one question emerges: how are these tools supposed to interact with lexical resources? In years of literature in NLP, many systems have adopted resourcebased approaches in order to perform given tasks, however the “knowledge” about these resources, i.e., their semantics, were always “hardwired”, “embedded” within that same systems that were exploiting their content. Besides doing the same work again and again, these systems were lacking at the same time the capability to scale up to a wider scenario including similar resources, or being adaptively capable of reusing different ones…in different ways. In a word, they lacked a common lexical model, upon which to see each (lexical) resource not as a world-per-se, but as the emanation of a more general theory.
In [
(Semi-)Automatic Ontology Alignment
Ontology Matching [
Since 2004, the OAEI4 organizes annual campaigns for the evaluation of automatic ontology matchers. The general scenario for the contests is simple: two datasets, and the objective to align their elements as much as possible. The contest then comprises different challenges, differentiating by the type of dataset (an ontology, a thesaurus instance data), the natural language(s) involved (same language, two different languages, sets of languages available for each dataset), the elements to match etc.
While the participants to the contest are focused on developing more and more complex techniques for alignment (and the results can obviously be considered scientific as they are run on repeatable experimental settings), the question whether it is important to improve performances on worst case scenarios – while the conditions of these scenarios could be improved at the Web level – arises. Currently, there is no widely adopted metadata vocabulary providing resuming information about the available natural languages, their coverage of the analyzed resources and the modeling vocabulary adopted to represent this information (are terms expressed as rdfs:labels, through SKOS, SKOS-XL, or other advanced models such as those presented in section 2). While metadata can obviously be recomputed as a first processing step in a deep alignment procedure, chances that these alignments could be applied “on the fly” are severely reduced (in quality, if not on feasibility at all). If we want the Web of Data to succeed, it is obvious that proper publication of data should be encouraged by providing quality indicators, such as the Five Stars Open Data reference5. Metadata should be a fundamental part of the Web, and vocabularies such as VoID provide useful information that all publishers should provide. While it is important to foster research in robust procedures considering worst cases, it should be also important to understand what would be the benefit of a better acknowledgement of the available information in a repository, and how this can be exploited.
By analyzing the conclusions from each of the workshops and related literature, we
can gather some interesting facts:
─ We are probably close to the achievable limits (discussion in OAEI 2013)
─ There is however still an incremental growth by tuning existing techniques [
In particular, in the library track of OAEI, dealing with the alignment of the thesauri
TheSoz6 and STW7:
6 http://www.gesis.org/en/services/tools-standards/social-science-thesaurus/
7 http://zbw.eu/stw/
─ Most of the participants in OAEI do not understand SKOS. Thus, thesauri need
to be sweeten into ontologies by means of a lossy and incorrect mapping [
It thus seems that the quality of algorithms adopted is an order of magnitude less important than properly filtering, cleaning, managing and, most importantly, being able to exploit the linguistic information available in ontologies and datasets. Simply put, the competing systems are currently not squeezing all the juice out of this very important information, losing much of what would be easily available, while pushing hard on disentangling un-understandable URIs and trying to grasp some knowledge from their structural organization.
Unfortunately, as I remarked, this information is not explicitly available through any dedicated vocabulary and most of these tools require fine-tuning by users, where they should be able to self-configure. In a few words: machines need humans to tell them how to read machine-understandable data!
As I explained in [
In the context of the OntoLex community group, the scenario envisioned in [
A recent attempt at exploiting information from Lime and from other available
metadata (such as the Vocabulary of Interlinked Datasets VoID [
The acquisition of knowledge from unstructured content has maturated dramatically –
especially in the latest years – on the engineering aspects other than on the techniques
for content analysis. The TIPSTER [
Following the advent of RDF, a few attempts at empowering information extraction
architectures with data projection mechanisms have been presented. Some examples
are the RDF UIMA CAS (Common Analysis Structure) Consumer [
Another important achievement has been the proposal for a NLP Interchange Format
(NIF, [
However, again, in the same way as a linked data approach can benefit NLP, the benefit can be reciprocal. Until now, NLP has been a very closed world: even when getting out of the scientific field and moving to industry, NLP has been seen as a sort of highly specialized sector where experts work out of ordinary software performance/quality parameters and interoperability needs. This has explained the limited success that it had until now, especially in smaller organization which are have not the power to deal with the complexity that NLP processes require.
In my vision, those same NLP components the behavior of which is advertised as linked data (in environments such as NIF), should be made available as open, downloadable and pluggable components that can be dynamically agglomerated, in order to assemble complete NLP chains. Much in the spirit of provisioning systems such as Maven9, that are however focused (or, at least, commonly used for the large part) on the provisioning of software libraries and on dependency resolution for Integrated Development Environments, NLP components should be provisioned as interconnectable pieces of a larger puzzle. Their metadata, and metadata coming from linguistic resources and from ontologies as well, should all create a global ecosystem in which it will be possible to target knowledge to be acquired/improved, sources of information, potential supporting resources and the software needed to provide them. 5
In this article, which followed my keynote speech at the 4th Multilingual Semantic Web Workshop (collocated with the 12th edition of the Extended Semantic Web Conference, in Portoroz, Slovenia), I showed a few practical applications of a more linguistically8 Organization for the Advancement of Structured Information Standards 9 https://maven.apache.org/ aware Semantic Web, in terms of enriched possibilities for well known tasks, such as lexical documentation, ontology alignment and content acquisition, which ultimately can contribute to the growth of the Web itself.
The objective is to remark, once more, that despite (to reuse the words of Tim Berners-Lee about his own creation) the “concept of machine-understandable documents does not imply some magical artificial intelligence which allows machines to comprehend human mumblings”, still this machine-understandability has its limits, dictated by open and environments and heterogeneous needs. Allowing machines to trade their knowledge when they have no better agreement on it, or to support humans in extracting information from textual content, brings about a marriage with the sole form of knowledge exchange that has endured hundreds of years of evolution: natural language. UIMA.
Available