We present a survey of the usage and style of identi ers and labels of named entities in a corpus of OWL ontologies. We investigated the frequency of use of both labels and meaningful or meaningless identi ers in those ontologies. We also surveyed common practices of lexical encoding styles for identi ers. We found that most ontologies do not use labels for named entities. When they do use labels, those labels are mostly meaningful and most ontologies also used meaningful identi ers. CamelCase style appears to be the most widely used style of lexical encoding for identi ers. We observed, however, that the majority of the ontologies use a mixture of two or more lexical encoding styles. The result of this survey is useful when considering strategies, for example, natural language generation from ontologies or converting artefacts, such as OWL ontologies, into languages like the Simple Knowledge Representation System (SKOS), where the notion of label is important. Given that labels are optional in OWL ontologies, what is the best way to handle the label selection when converting them into SKOS? Merging multiple entities may require selection from labels or identi ers assigned to these entities for skos:prefLabel and skos:altLabel.
In this paper we present a survey of how identi ers and labels are used within
Web Ontology Language (OWL)1 ontologies. We are interested in transforming
such ontologies in to other forms|such as natural language and in to other
Semantic Web representations such as the Simple Knowledge Organisation System
(SKOS)2. In these transformations it is important to be able to deal with both
identi ers and labels in OWL ontologies. In natural language generation, for
example, a human understandable form of the entity needs to be available to place
within a natural language setting [
Ontologies are used to capture knowledge about some domain of interest. An ontology describes the concepts in the domain and also the relationships that hold among these concepts. These concepts can be represented by classes or individuals, and the relationships are represented using properties. In OWL, the concepts and properties can be referred to as entities. A named entity refers to a named class, a named individual or a named property. Each named entity must have a unique identi er, called an IRI. An IRI refers to an object that can act as a reference to something that has identity.
Identi ers are not only used by computers, but also by humans. Humans prefer using meaningful identi ers|the name encapsulates the nature of the entity that it names. An identi er is meaningful if there is a direct relationship between the natural language term used and the characteristics about the entity being identi ed. For example, if the entity is used to represent a concept \dog", then using dog as an identi er for this concept helps to make the identi er meaningful (to an English speaker).
It is also possible, however, to have a meaningless identi er, or also called a \semantic-free" identi er. For example, ABC 20020 is a semantic-free identi er. An identi er is meaningless or semantic free if there is no direct relationship between the natural language term used and the characteristics about the entity being identi ed. In OWL it is possible to separate the IRI for the entity and the label for that entity (usually provided through rdfs:label). When identi ers are meaningless in human terms, the entity needs a label that is a natural language term for that entity. This can have several desirable e ects, including: ability for having di erent language renderings; being able to change the label without having to change the identity of the entity (which is useful when the ontology is being used to encode data); and so on.
As identi ers in OWL can contain no spaces, meaningful identi ers that would normally contain spaces have to be encoded in a way that excludes spaces, but retains the meaningful nature of the identi er for human readers. An identi er can be encoded in various lexical encoding styles. Using internal upper case letters within an identi er to denote word boundaries{camel case style (eg. PetOwner or petOwner); underscore style ( ) (eg. pet owner); and hyphen style (-), are among the styles that are used in meaningful IRIs. This means that a meaningful identi er needs some processing to have it in a conventional form for human reading|that is, containing spaces between words.
We used a survey to determine the current use of labels and identi ers in ontologies including the naming convention of identi ers. The goal was to allow us to answer the following questions: 1. Given that labels are optional, what is the frequency of label use in an ontology? 2. Given that a label should be meaningful, and if labels are used in an ontology, what is the frequency of meaningful labels used? 3. Given that an entity could have multiple labels, what is the frequency of an entity having multiple labels? 4. What is the frequency of meaningful and meaningless identi ers used in an ontology? 5. What is the frequency of the following combinations between identi ers and labels used in an ontology? (a) an entity with a meaningful identi er and has meaningful label(s)? (b) an entity with a meaningful identi er and has meaningless label(s)? (c) an entity with a meaningful identi er and has no label? (d) an entity with a meaningless identi er and has meaningful label(s)? (e) an entity with a meaningless identi er and has meaningless label(s)? (f) an entity with a meaningless identi er and has no label? 6. What is the frequency of camel-case, underscore style and hyphen styles encoding used in an ontology to encode identi ers? 2
An overview of the methods used to answer these questions is: 1. Corpus preparation; 2. Isolation of identi ers and labels; 3. Determination of whether the identi ers and labels are meaningful or meaningless; 4. Result recording; 5. Data analysis. 2.1
In this survey we used ontologies in the TONES repository3. We also searched on Google using filetype:owl for more OWL ontologies to be added to our corpus. From the search result, we looked through each ontology for OWL constructs such as owl:Class, owl:Individual, owl:ObjectProperty or owl:DataProperty to be considered as a \valid" OWL ontology for this survey. All collected ontologies from both sources were compared to eliminate duplication. We utilized the OWL API4 for loading and managing the OWL ontologies. All ontologies were locally stored for future reference.
For each ontology, we isolated the identi ers and labels for each named entity. The identi er was extracted from the IRI for each named entity. If the IRI contained a fragment identi er, then the identi er for this entity is the fragment identi er (the fragment after the (#) character). For example, for the named class http://owl.cs.manchester.ac.uk/2010/people#person, we extracted person as the identi er. Otherwise, we took the last portion of the path component as an identi er (the fragment after the last (/) character5. For example, for the named class http://owl.cs.manchester.ac.uk/2010/pizza/ pizzaTopping, we extracted pizzaTopping as the identi er.
Entity labels were identi ed through the annotation property rdfs:label in the ontology. An entity is considered to have a label if there exist one or more rdfs:label associated with the entity. We also considered labels made through sub-properties of rdfs:label. 2.3
Our aim is to test a label or identi er against the Web to see if it is meaningful. Gaining many pages or `hits' to a query based on a label or identi er would suggest that it is meaningful|based on the assumption that use of the string on Web pages suggests its use in natural language. This is a two-stage process. Normalise the lexical encoding style of identi ers Prior to this test, however, an identi er must be put in to a form suitable for querying as identi ers are formed with no spaces. As described in Section 1, identi ers are normally encoded in various lexical encoding styles. In order to determine the meaningfulness of an identi er, the human brain will apply some cognitive manipulation on the encoded identi er into a form (space-separated form) that could be more readily interpreted. We called this process `normalisation'. To check the meaningfulness of an identi er, we normalised the string used by transforming it in to a space separated form. In order to do the transformation, we rst needed to identify the style of lexical encoding used to encode the identi ers. We have identi ed the following commonly used lexical encoding styles, and limit our categorisation to these, placing any identi ers not using these styles in an \Other style" category. 1. CamelCaseStyle. 2. Underscore style. 3. Hyphen-style. 4. HybridCamelCase underscore style. 5. HybridCamelCase-hyphen-style. 6. Hybrid-hyphen underscore style. 7. Single word
8. Other style|any identi ers that are encoded using other than the styles mentioned above are grouped under this category.
All single word identi ers are grouped in the \single word" category. This
category can be considered as a \wild card" as it is compatible with all other
categories. Therefore, we used to following rules to decide the lexical encoding
style used in the ontologies.
1. If all identi ers in an ontology are encoded using single word, then classify
the ontology as having only single word identi ers.
2. If some identi ers are encoded using the single word style and only one other
lexical encoding style is used for the rest of the identi ers in an ontology, the
single word category can be made compatible with the one lexical encoding
style, and the ontology is classify to have that one lexical encoding style.
For example, the rest of the identi ers in an ontology were encoded with
camelCase style, then, the single word category can be made compatible
with camelCase style and the ontology can be classi ed as encoded using
camelCase style.
3. If some identi ers are encoded using single words and more than one lexical
encoding style is used for the rest of the identi ers in an ontology, then the
single word category cannot be made compatible with any of these styles
and the ontology is classi ed as having a mixture of lexical encoding style.
Once the lexical encoding style has been identi ed, we then normalise the
identi er into a space-separated form to be used in the meaningfulness checking.
Check for meaningfulness For our meaningfulness check we used a Web
search query using the Bing API 6. For each label and normalised identi er, we
sent a Web search query to the World Wide Web (WWW) to search for the
number of websites with the words in the labels and normalised identi ers. We
are interested in the number of results returned from the search query. There
are three options for sending strings to the Web search query. First, using the
quotation (\ ") around the search string. For example, string hello world is
search as "hello world". Since this query searched for exactly the same
occurrence of string in the Web it returned limited number of search results due
to the reason that not all words in an identi er occurred together in natural
language presented in the Web. Second, search for the string without the
quotation. For example, using string as hello world as the search string. This type
of query searched for the words in an identi er that occur anywhere in a Web
page, and not necessarily in the same order. The result return by this type of
query is moderate and acceptable. Third, if a string consists of multiple words,
search for the words as separate string with and/or without quotation [
be chosen. For this survey, we chose the second option to search the terms together without quotation. Additionally, we used the order of words as how it appears in the identi ers and labels. For example, if the normalised identi er is \hello world", then we use a string hello world with the same order for the Web search query. We set a threshold value of 100 hits which is used to determine the meaningfulness of the searched term. A hit result that is below 100 is considered not meaningful. The choice of 100 as the threshold is a heuristic based on running a few ontologies from various domains and simply judging a reasonable threshold. We found that ontologies with medical terminologies, get fewer hits for meaningful identi ers. 2.4
We recorded the results of this survey at various stages. All extracted identi ers and labels were recorded in XML format for future reference and analysis. We also recorded the hit results for each of the identi ers and labels from the Web search query in CSV format for future reference and analysis. 2.5
Based on the recorded results, we calculated and recorded for each ontology and for each named entity, the following: 1. frequency of labels used; 2. frequency of meaningful and meaningless labels; 3. frequency of identi ers with one label and more than one labels; 4. frequency of meaningful and meaningless identi ers; 5. frequency of the combination of identi ers and labels; 6. frequency of lexical encoding styles.
For each entity type in each ontology, we also calculated the proportion of these frequency with respect to its total entity for each criterion listed above in the form of a percentage. Finally, we calculated the mode, mean and median of these percentages for each of the criteria. 3
We used 219 valid ontologies from the TONES repository7, after discarding any URIs that no longer existed or were too big to be loaded in our machine. There were 354 hits returned from the search query8 after all the duplicate results were omitted. After looking at each URLs, only 264 URLs represented valid OWL ontologies, the rest were URLs linked to pages that no longer existed. We also compared the list of URLs with the ontologies from the TONES repository to avoid duplication. Out of the remaining 241 ontologies, we have randomly selected 87 ontologies to be added to our corpus|making a total of 306 ontologies.
Out of 306 ontologies, 5 ontologies contained none of the named entities| leaving 301 ontologies in the corpus. There are 296 ontologies containing named classes; 105 ontologies with named individual; 264 ontologies with named object properties and 138 ontologies with named data properties. The rest of the analyses were performed on ontologies that contained named entities.
Table 1 shows the result summary with the number of ontologies for each criteria surveyed9. The mean shown represents the mean of the proportion of the measured criteria.
Type Total ontologies Ontologies with Labels Meaningful labels Meaningless labels Ontologies with single label Ontologies with multi labels Meaningful identi ers Meaningless identi ers Meaningful identi ers with meaningful labels with meaningless labels with no label Meaningless identi ers with meaningful labels with meaningless labels with no label Lexical Encoding Style of Identi ers CamelCase style Underscore style Hyphen-style CamelCase underscore style CamelCase-hyphen style Hyphen-underscore style Single word Others Mixture
Classes Individuals Object Data
Properties Properties Count Mean Count Mean Count Mean Count Mean 296 105 264 138 122 32.9% 32 20% 82 27.8% 23 14.4% 122 89.4% 32 94.4% 82 95.6% 23 91.7% 70 6.9% 12 5.6% 13 4.4% 6 8.3% 121 93.9% 31 88.6% 81 97.5% 22 94.5% 21 6.1% 8 11.5% 4 2.5% 2 5.5% 286 85.2% 103 90.6% 263 97.8% 137 97.2% 135 14.9% 49 9.4% 37 2.3% 23 3.5% 9 The complete analysis of the result for this survey is made available at http://www.
myexperiment.org/packs/110
Table 1 provides basic answers to the questions raised in Section 1 in numerical terms. Here we present here some observations based on an initial analysis of those results along with closer examination of some of the ontologies.
First, we appreciate that the technique used to determine meaningfulness of both labels and identi ers { using a search with a xed cuto threshold { is rather basic. The threshold value was selected based on some preliminary experiments, but it is likely that the use of a single static threshold value is not appropriate for all domains { see the discussion below. However, for the purpose of this survey, the technique is enough to show some interesting results. We are currently extending and exploring possible mechansisms for the selection of variable threshold values based on the content of each ontology rather than having a single static cut-o value for all ontologies.
Labels are not widely used in all named entity types. However, when labels are used in an ontology, those labels are usually meaningful. In terms of the number of labels per entity, we observed that, for all named entity types, almost all ontologies contained single labels. Where an ontology does contain more than one label per entity, closer investigation revealed that the multiple labels were used to represent labels in di erent languages. Single labels usually represent labels in one language only.
Almost all of the ontologies used meaningful identi ers for named entities, with object property and data property entities showing the highest use of meaningful identi ers. Further analysis, shows that those identi ers for object and data properties that are classi ed as meaningless are actually meaningful, but the meaningfulness test gave a hit below the threshold (as discussed above). As for meaningless identi ers, even though the result shows that quite a number of ontologies used meaningless identi ers, their percentage of usage (in terms of the proportion of entities in the ontologies) is quite small.
For all named entity types, most of the ontologies contained meaningful identi ers with no label. This observation supports our ndings that labels are not widely used in the ontologies and most ontologies do have meaningful identi ers. Interestingly, we observed that there are also a few ontologies that use meaningless identi ers with meaningless labels or no labels. However, their mean percentage of use is rather small. We suspect again that our approach in identifying the meaningfulness of terms is a factor in this abnormality. Having a meaningless identi er and no label makes little sense; it is reasonable to suspect that specialised language will appear meaningless in the face of the simple threshold approach used.
As for lexical encoding styles, the result show that camel case style is the most used lexical encoding style for all named entity type. There are also a signi cant number of ontologies that use a mixture of lexical encoding styles. A small number of ontologies used unidenti ed lexical encoding style under the others category. Further analysis of this category showed the identi ers classi ed into this category used other punctuation symbols such as dot (.) to encode the identi ers. Some example of identi ers in this category are as follows: { E1.CRM Entity (combination of dot (.) and underscore) { E71.Man-Made Thing (combination of dot (.), hyphen and underscore) { erbB-2 Genes (combination of camel case style (erbB), hyphen and underscore) The small numbers obtained for the \others" category suggests that the categories identi ed in Section 2.3 are indeed su cient to characterize the bulk of ontologies in the corpus. 5
There are several surveys that analyse Semantic Web documents especially OWL
ontologies to help understanding of the nature of OWL ontologies. Bechhofer and
Volz [
We found that most ontologies do not use labels for named entities. When they do use labels, these labels are mostly meaningful. Only a few ontologies have more than one label per named entity. We also found that most of the ontologies do use meaningful identi ers and if they do use meaningless identi ers, these identi ers only occupied a small portion of the ontologies. Most ontologies that have meaningful identi ers do not have labels. Interestingly, there are also a few ontologies that used meaningless identi ers with meaningless labels or no label; though this may well be an artefact of our test for meaningfulness.
Camel case style appears to be the most widely used lexical encoding style for identi ers. However, most ontologies are inconsistent in their identi er encoding style, as more than one style is used to encode the identi ers within an ontology.
We hope to extend this survey on a larger corpus of ontologies. For example, collecting for more ontologies from various other sources like the Swoogle 10 and Watson 11. It also might be interesting if we could extend the survey to not only 10 http://swoogle.umbc.edu/ 11 http://kmi-web05.open.ac.uk/WatsonWUI/ investigate the use of labels and identi ers, but also other OWL constructs such as property restrictions, to have a better understanding of the common practice of use of these constructs in the existing OWL ontologies.
We can raise further questions about the e ect of the domain for which an
ontology was built on its style of identi er and label use. For example, the Open
Biomedical Ontologies consortium [
When transforming OWL ontologies into other forms { such as natural language or to other Semantic Web representations such as SKOS, an understanding of the use of labels and identi ers within the ontologies is bene cial. If nothing else, it allows developers to make judgements about situations for which strategies should be developed.
Acknowledgements: This work was funded in part by the SWAT project EP/G032459/1. The authors would like to thank Majlis Amanah Rakyat (MARA), an agency under the Malaysian Government, for funding the student. Many thanks to the reviewers who gave insightful comments and suggestion to improve this paper.