This article could be split quite neatly in two articles. One is an excellent article that begins about a third of the way through the full piece. It covers the technical aspects of ontology building: subsumption hierarchies; necessary vs necessary and sufficient conditions for class membership; disjointness; relations; upper ontologies and their usefulness in restricting the choice of appropriate relations. It draws heavily on upper ontologies developed by philosophers (at least some of them realists) and shows why they are useful. It concludes with a clear and strong case for why good ontologies are needed in the biosciences. I have no argument with this article.

The other is, to me at least, a rather confusing attempt to argue that ontologies consist of concepts, as opposed to statements about reality. I find these arguments difficult to square with some of the statements made in the rest of the article. However, I’m also not convinced that there is much difference between the author’s position and a realist stance. There argument hinges on the subtle issue of the reality of classes and they don’t make other arguments commonly made against a realist stance – for example, the ‘argument from intellectual modesty’ (Smith et al., 2006), or the belief that ontology terms should simply follow the use of terms in language. In fact, they clearly argue for ontology as a means to overcome the latter:

“Ontology should be distinguished from thesauri…”

“Human beings can give multiple labels to … categories. This habit of giving multiple labels to the same category and the same label to different categories (polysemy) leads to grave problems…”

Their argument begins with what strikes me as a cheap rhetorical trick designed to close down debate:

“The definition here will not suit a lot of people and upset many (especially use of the word “concept”); We make no apology for this situation, only noting that the argument can take up resources better used in helping biologists describe and use their data more effectively.”

If the authors think this discussion is a waste of resources, then why bother spending a few paragraphs making their case? I suspect that they do actually care about the argument because they worry about the implications of taking a realist stance. If so, it would have been interesting to hear some of those concerns (on the realist status of maths for example) made more explicit.

There is also a notable lack of reference to any sources of opposing argument. For those who wish to pursue this argument further, waste of time though it might be, some references to counter arguments would be good. Either of these references (or both) would do nicely:

Smith, et al., 2006. Towards a reference terminology for ontology research and development in the biomedical domain. Proceedings of KR-MED 2006

Smith, 2004. Beyond Concepts: Ontology as Reality Representation. Proceedings of FOIS 2004

Of the arguments against a realist stance, the weakest uses a straw man:

“… with a computer science ontology … there is less concern with a true account of reality as it is information that is being processed, not reality.”

Who could argue? Surely the question is whether the information being processed is making assertions about reality or not? The authors case would be stronger if this line were deleted.

The heart of their argument is stated here:

“As human beings, we put these objects into categories or classes. These categories are a description of that which is described in a body of data. The categories themselves are a human conception. We live in a world of objects, but the categories into which humans put them are merely a way of describing the world; they do not themselves exist.”

A perhaps pedantic point: are classes “described in a body of data”? I would have thought it more likely they are assertions about reality that are a reasonable scientific interpretation of a body of data. This confusion of data and its interpretation as assertions occurs consistently throughout the article.

More importantly, what might it mean to state that a class is real? Even the authors seem to agree that there is regularity in the universe, whether we observe it or not. For example, later in the article, they state that:

“Each instance of a ‘Helium’ object was not discovered in 1903; most helium atoms existed prior to that date, but humans discovered and labelled that category at that date.”

In 1903, humans discovered something that already existed: a class of atoms that share specific properties. Surely this means that a definition of the class ‘helium atom’ is making assertions about reality. Is this not different from some arbitrary class defined as including say: all helium atoms, horses, unicorns and two bedroom flats in North London?

This is not to say that there is only one true way to categorise any one object, or that there is a clean dividing line between classes we might be happy to define as Universals (Smith ), such as Helium atoms, and more contingent classes.

In the abstract, the authors argue that the debate over a realist vs a conceptual stance is a distraction that “… can take up resources better used in helping biologists describe and use their data more effectively.” Why might a realist stance be useful in helping scientists?

I believe that a realist stance is useful for ontologies made up of scientific assertions (for example, about chemistry, anatomy, or physiology), because it gives us a way to judge the quality of an ontology. If the ontology makes assertions that run counter to what we have good reason to believe is true, then it is misleading as a knowledge-base about science and its use in inference and grouping of annotations will produce results that we have good reason to believe are incorrect. Surely such an ontology would be bad – even when judged purely in practical terms.

Having said all this, I’m happy for this article to pass review for the Ontogenesis Knowledge Blog as long as the authors add references to opposing arguments. The authors may wish to consider taking into account my points with regard to the abstract and the apparent use of a straw-man argument. The article already provides an excellent introduction to the basic technical aspects of ontology building. With the addition of references to opposing arguments, the article and this review should provide a good starting point for those interested in exploring the realism vs conceptual(ism?) debate further.

Minor corrections:

foundary -> foundry

polysemy – missing initial bracket

License

This paper is an open access work distributed under the terms of the Creative Commons Attribution License 3.0, which permits unrestricted use, distribution, and reproduction in any medium, provided that the original author and source are attributed.

“Back to books: Researchers should be recognized for writing books to convey and develop science.”

References

1. Nature, Vol. 463, No. 7281. (03 February 2010), pp. 588-588. DOI:10.1038/463588a

Unfortunate, but not surprising. Pingbacks still work, so commenting without registration can still happen this way.

The hit count is still at around 3000 hits per day, which works out at around 150 article views a day.

Apologies to those who noticed the outage from the server this afternoon. Naturally, I chose the period of initial internet exposure to fiddle with the server and, so, took apache down for the duration. Possibly not the greatest decision ever; sometimes I amaze even myself. There will be a second outage later in the week for further work.

Are we nearly there yet?

Process

Short or Long?

Wiki or Blog?

Communication

Technologies

Conclusions

Comments

This is a clear, precise and well-written article describing the topic. It covers the motivation, process and technology of this approach for the development of a polyhierarchy while being clear that the axiomatisation necessary can have upfront costs. The article also makes clear use of figures demonstrating the use, as well as examples clarifying the more technical statements.

The latter half of the article addresses the issue of normalisation which is a development methodology that enables automatic maintenance of multiple inheritance ontologies. Well this is appropriate material for the article, it is not strictly necessary to completely normalize an ontology to use reasoning. The conclusions, in particular, could be improved in structure by saying that automatic maintaince can be supported by reasoning, and that normalisation exploits this fully.

Otherwise, an excellent article.

Minor Corrections

General

The section titles should use HTML headers, rather than the typographic markup used. Authors email should be hyperlinked.

Multiple Inheritance Ontologies

The effort is considerable, but worthy, as the automated reasoner is able to maintain the whole structure, avoiding human errors. Also, using such expressive axiomisation enables richer queries and other advantages.

Change to

is considerable requiring a richer axiomatisiation but worthwhile as the…

The more expressive axiomatisation also enables richer…

para 2 The difficult on maintaining -→ of maintaining

What is normalisation

Therefore, adequete and precise -→ However, adequete

OWL provides -→ Languages such as OWL provide

(e.g. part_of some (nucleus and (has_function only photosynthesis))

Would be nice to describe what this means in English.

“Normalised CL there” -→ what does CL mean in this context?

Advantages of …

Such modelling dynamic also results in a modular ontology -→ This modelling process

as such relation is the result of both having a common condition -→ as this relation is the result….

Conclusions

long term, a manual maintenance -→ long term, manual maintenance.

This paper describes the three main components that can be found in an ontology (classes, instances, relations). Therefore it should be of interest to any newcomer to ontology development, as the confusion of which entities to use (especially instance vs. class) is a major problem when learning how to develop an ontology.

There is a difficult sentence at the end of introduction: “These components can be separated into two kinds; those that describe the Entities of the domain being described, and those which either enable the use of the ontology or describe the ontology itself.” There is no more references to such distinction in the rest of the paper. Also, it is not clear what are the components that describe the entities of the domain (axioms, individuals, classes?), the components that enable the use of the ontology (editors, APIs?) and the components that describe the ontology (metadata, metamodelling?).

The whole paper, although it describes general ontological components, has an OWL flavour, and the author should be explicit about it.

I would change the example for illustrating existential/universal, to an example where the class and the filler are different entities (not person/person), to make it more understandable.

I recommend accepting this paper.

Introduction

The Need for ontology learning

Ontologies are the fundamental form of knowledge representation in contemporary Artificial Intelligence systems, especially systems of used in the Life Sciences (link to What is an Ontology article). The vast majority of currently used ontologies have been built entirely by hand, including the major of Life Science ontologies such GO and those in OBO Foundry. This manual development process represents a major knowledge acquisition bottleneck as sometimes hundreds of hours of effort have been involved and there are ongoing teams of people in place to keep the ontologies up to date (1). One consequence of this has been a series of ongoing efforts largely led by members of the Natural Language Processing (NLP) and Text Mining communities to automate, or semi-automate the ontology construction process. Ontology Learning is motivated by a number of factors:

• The high manual cost of ontology construction
• The continuous change in science and knowledge in general
• The very large amount of existing text with numbers growing exponentially (PubMed has over 2000 papers added to it per day)
• The extensive need for a variety of ontology type structures ranging from the relatively informal such as structured vocabularies, through somewhat more formal taxonomies, to fully rigorous ontologies expressed in OWL.

This article will provide a brief introduction to some of the key concepts, approaches and challenges.

Underlying Assumptions

The work undertaken on ontology learning from texts is dependent on a number of core assumptions. All knowledge representation models including ontologies are linguistic artefacts, for two reasons (2). Firstly, ontologies function as a means for human beings to interact with machines, even if in a highly structured manner. Secondly, nearly all ontologies use linguistic terms (or quasi-linguistic terms) as labels for concepts. The fundamental assumption made by the NLP community is that there is an identity between terms in an ontology and in a text which look the same. This perception is further re-enforced by the extensive use of ontologies (especially in the Life Sciences) to annotate data including textual data.

On the NLP technology side, two other assumptions need to be identified. The distributional hypothesis is a very important working assumption which states that words which appear in similar contexts must be semantically identical or closely related (3), (4). All efforts to organise and cluster terms in texts, and to identify semantic similarity (synonymy) depend on the distributional hypothesis, and the identification of semantic difference or identity.

A further assumption is that texts actually contain the necessary information to derive ontologically relevant definitions. A common ontology learning model involves taking a set of domain specific texts and deriving (as far as possible) a corresponding ontology. The inadequacies which this approach has (5) have led to attempts to use multiple knowledge sources for the ontology learning process (6).

Current Approaches

Ontology Learning remains an active area of research with considerable potential for facilitating the work of ontology engineers. Further progress in Natural Language Processing and the integration of multiple sources of knowledge have the promise to substantially reduce the burden of formalisation by the ontology engineer. Concurrently, the outputs of ontology learning systems are of great utility in less formal context wheren taxonomies or formal vocabularies are required.

The Basics of Ontology Learning

At its most simplistic an ontology learning system (or workflow) allows the input of one or more texts and the output of some form of taxonomy. At a more detailed level, we can analyse the ontology learning from text process as follows (adapted from (7)):

1. Text Collection. A corpus of texts has to be identified, collected and preprocessed.
2. Term recognition or keyword extraction. Stopwords, named entities and other noise need to be removed and a subset of the vocabulary in the text identified.
3. Relation Extraction: Synonymy and Clustering. Terms which do have a relation need to be identified. This step also identifies terms or expressions such as abbreviations which are considered to be synonymous.
4. Relation Labelling. Relations which have been identified as existing need to be labelled. (Steps 3 and 4, often occur together).
5. Hierarchy Construction. Most ontologies contain a subsumption hierarchies if classes, however, it may also be desirable to extract different types of axioms from the text, including disjointness and equivalence.

Each step can be undertaken in a number of different ways, some of which we will briefly consider below.

Step 1 Text Collection: Text collections or corpora are usually “convenience corpora” i.e. based on a convenient collection of texts readily accessible to the researcher. Some authors have selected texts from the web, or extracted abstracts or complete texts from sources such as PubMed using specific key words.

Step 2 Term Recognition and keyword extraction: There are a number of standard approaches in the literature including (8), (9), (10), (11), (12). A comparative evaluation and integration of different methodologies relevant to ontology learning is presented in (13).

Step 3 Synonymy and Clustering: There s a large literature on different approaches cf. (14) Chap. 14. Classic systems for the creation of synonym sets include Sextant (15), and the work of the speech technology community in language models (16).

Step 4 Relation Labelling: Automated approaches to ontology learning have tended to restrict themselves mostly to learning ISA hierarchies (hyponomy and hyperonomy relations). There exist some limited efforts to learn other types of relations (17). There are two standard approaches to relation labelling:

• String Inclusion: This approach is used with multi-word terms with a structure ABC, the string inclusion approach assumes ABC ISA BC ISA C. There is a strong argument to interpret this relation not as a hyponym relation but (as defined in SKOS) as a “broader than/narrower than” relation. This is a high recall but low precision approach (1).
• Lexico-syntactic Patterns: Ever since Hearst’s work (18), there has been a strong tradition of using lexico-syntactic patterns to identify a variety of ontological/semantic relations between terms. These patterns are usually of the form “NP is a type of NP” or “NP and other NPs”. This approach has a tendency to have nigh precision but low recall.

Step 5 Hierarchy Construction: Hierarchy construction usually falls out of the relation labelling step and may be combined. There are other approaches which focus first of hierarchy construction and ignore the relation labelling stage (19).

Other Knowledge Sources

It is common practice in ontology learning approaches to use multiple sources of knowledge apart from textual ones. Thus a variety of existing knowledge structures have been used to build ontologies semi-automatically. These include WordNet (20), existing thesauri such as Rogets, and in the Life Science domain such structures as UMLS. There is an inherent contradiction in using existing structures as part of the point is build up to date representations of knowledge as reflected in corpus of texts. However, ever since the work of Agirre et al. (21), there have been many attempts to use a starting ontology and augment or update the ontology using corpora or the web.

Specific Systems and Tools

• Text2Onto: Combines machine learning approaches with basic linguistic approaches such as tokenisation and part-of-speech (POS) tagging. Text2Onto is built upon the General Architecture for Text Engineering (GATE) that allows for flexible integration of natural language processing components. It also allows Java Annotation Pattern Engine (JAPE) rules to be written that can be used to facilitate recognition of ontological primitives.
• Abraxas: This is a system based on an iterative open ended approach to the ontology learning process. A seed corpus is used, terms are identified, ontological knowledge in the form of triples are extracted using lexico-syntactic patterns, gaps in ontologically explicit knowledge are identified and suitable texts to “cover” the gaps are identified from an external repository such as the web (22).
• KnowItAll: While strictly not an ontology learning system, the KnowItAll system is a good example of large scale knowledge acquisition from the web. It uses most of the component techniques widely used in ontology learning except it outputs knowledge fragments, rather than complete ontologies (23).
• OntoLearn: This is a system (24) which uses extensively WordNet as an external resource. A corpus has term extraction applied (the OntoLearn team have a high quality term recognition component (25)), and then knowledge about terms from WordNet is integrated. Ontolearn have undertaken the most extensive end user evaluations of their system (26).

Challenges and Future Directions

There are two major challenges for automated ontology learning. Both have practical and philosophical implications.

The Nature of Text

There is a significant gap between the vagueness, fluidity and ambiguity of natural language and the logical rigour of fully formal ontologies. One of the key challenges in the ontology learning field remains the extent to which techniques can be found to bridge this gap. The current approaches tend to either ignore this gap or make explicit the need for human post editing of the output of the ontology learning system. Domain specific text collections do not contain sufficient instances of use of terms, and sufficient variety of contexts to enable the accurate determination of ontological relations and features (5). In effect a large proportion of the knowledge that needs to be formalised is not present in most corpora, especially domain specific collections. It is in this light that the wider use of the web as a corpus for knowledge extraction/collection is highly attractive (cf. especially the KnowItAll system (23), (27)) but this will tend to increase the problems of ambiguity and domain specific senses.

Evaluation

The evaluation of the output of ontology learning systems remains a major challenge. The usual approaches to ontology evalauation have largely been based on quality control of the ontology building process and ensuring the ontology abides by certain principles (28), (29). From an ontology learning from text perspective, these approaches are inapplicable. The usual approach is to use a Gold Standard with which to compare the automatically generated output (6), sometimes using WordNet as the Gold Standard. The main problem with a Gold Standard is that if the ontology learning system is supposed to discover new knowledge then a GS based evaluation may penalise a system for discovering knowledge absent from the Standard. Also as systems scale up, it becomes harder and harder t get an overall assessment of the quality of the resulting output in a given case.

Conclusion

Ontology learning is an active field of research with considerable potential to make the task of ontology engineering substantially easier. Another important role for OL is to allow ontologies to be kept up to date more effectively. Finally, one should mention that in many contexts where less formal ontologies are needed (taxonomies or structured vocabularies), they output of OL systems is very close to that required.

Acknowledgements

This paper is an open access work distributed under the terms of the Creative Commons Attribution License 3.0 (http://creativecommons.org/licenses/by/3.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided that the original author and source are attributed.

The paper and its publication environment form part of the work of the Ontogenesis Network, supported by EPSRC grant EP/E021352/1.