DBpedia Schema Queries¶
Overview¶
In this notebook, I begin the process of analyzing the schema of the DBpedia Ontology. This is a local notebook in which I load data from the filesystem into an in-memory graph, thus it is part of the unit tests for gastrodon. This is feasible because the schema is much smaller than DBpedia as a whole.
The following diagram illustrates the relationships between DBpedia Ontology, it's parts, DBpedia, and the world it describes
The numbers above are really rough (by as much as 30 orders of magnitude)), but some key points are:
- The DBpedia Ontology has it's own Ontology, which is a subset of RDFS, OWL, Dublin Core, Prov and similar vocabularies
- The DBpedia Ontology is much smaller (thousands of times) than DBpedia itself
- DBpedia does not directly describe the external universe, but instead describes Wikipedia, which itself describes the universe.
It's important to keep these layers straight, because in this notebook, we are looking at a description of the vocabulary used in DBpedia that uses RDFS, OWL, etc. vocabulary. RDF is unusual among data representations in that schemas in RDF are themselves written in RDF, and can be joined together with the data they describe. In this case, however, I've separated out a small amount of schema data that I intend to use to control operations against a larger database, much like the program of a numerically controlled machine tool or the punched cards that control a Jacquard Loom.
This notebook is part of the test suite for the Gastrodon framework, and a number of bugs were squashed and API improvements made in the process of creating it. It will be of use to anyone who wants to better understand RDF, SPARQL, DBPedia, Pandas, and how to put it all together with Gastrodon.
Setup¶
As always, I import names from the Python packages I use:
%load_ext autotime
import sys
sys.path.append("../..")
%matplotlib inline
from rdflib import Graph,URIRef
from gastrodon import LocalEndpoint,one,QName
import gzip
import pandas as pd
pd.set_option("display.width",100)
pd.set_option("display.max_colwidth",80)
g=Graph()
g.parse(gzip.open("data/dbpedia_2015-10.nt.gz"),format="nt")
Now it is loaded in memory in an RDF graph which I can do SPARQL queries on; think of it as a hashtable on steroids. I can get the size of the graph (number of triples) the same way I would get the size of any Python object:
len(g)
The Graph is supplied by RDFLib, but I wrap it in an Endpoint object supplied by Gastrodon; this provides a bridge between RDFLib and pandas as well as smoothing away the difference between a local endpoint and remote endpoints (a SPARQL database running in another process or on another computer)
e=LocalEndpoint(g)
Counting properties and discovering namespaces¶
Probably the best query to run on an unfamiliar database is to count the properties (predicates) used in it. Note that the predicates that I'm working with in this stage are the predicates that make up the DBpedia Ontology, they are not the predicates that are used in the larger DBpedia Ontology. I'll show you those later.
properties=e.select("""
SELECT ?p (COUNT(*) AS ?cnt) {
?s ?p ?o .
} GROUP BY ?p ORDER BY DESC(?cnt)
""")
properties
Note that the leftmost column is bold; this is because gastrodon recognized that this query groups on the ?p variable and it made this an index of the pandas dataframe. Gastrodon uses the SPARQL parser from RDFLib to understand your queries to support you in writing and displaying them. One advantage of this is that if you want to make a plot from the above data frame (which I'll do in a moment after cleaning the data) the dependent and independent variables will be automatically determined and things will 'just work'.
Another thing to note is that the table shows short names such as rdfs:label as well as full URIs for predicates. The full URIs are tedious to work with, so I add a number of namespace declarations
and make a new LocalEndpoint
g.bind("prov","http://www.w3.org/ns/prov#")
g.bind("owl","http://www.w3.org/2002/07/owl#")
g.bind("cc","http://creativecommons.org/ns#")
g.bind("foaf","http://xmlns.com/foaf/0.1/")
g.bind("dc","http://purl.org/dc/terms/")
g.bind("vann","http://purl.org/vocab/vann/")
e=LocalEndpoint(g)
properties=e.select("""
SELECT ?p (COUNT(*) AS ?cnt) {
?s ?p ?o .
} GROUP BY ?p ORDER BY DESC(?cnt)
""")
properties
Metadata about the DBpedia Ontology¶
I find it suspicious that so many properties occur only once, so I investigate:
single=e.select("""
SELECT ?s {
?s dc:source ?o .
}
""")
single
The one function will extract the single member of any list, iterable, DataFrame, or Series that has just one member.
ontology=one(single)
The select function can see variables in the stack frame that calls it. Simply put, if you use the ?_ontology variable in a SPARQL query, select will look for a Python variable called ontology, and substitute the value of ontology into ?_ontology. The underscore sigil prevents substitutions from happening by accident.
and we see here a number of facts about the DBpedia Ontology, that is, the data set we are working with.
meta=e.select("""
SELECT ?p ?o {
?_ontology ?p ?o .
} ORDER BY ?p
""")
meta
ontology
How Gastrodon handles URI References¶
Gastrodon tries to display things in a simple way while watching your back to prevent mistakes. One potential mistake is that RDF makes a distinction between a literal string such as "http://dbpedia.org/ontology/" and a URI references such as <http://dbpedia.org/ontology/>. Use the wrong one and your queries won't work!
This particularly could be a problem with abbreviated names, for instance, let's look at the first predicate in the meta frame. When displayed in as a result of Jupyter notebook or in a Pandas Dataframe, the short name looks just like a string:
license=meta.at[0,'p']
license
that's because it is a string! It's more than a string, however, it is a class which is a subclass of string:
type(license)
and in fact has the full URI reference hidden away inside of it
meta.at[0,'p'].to_uri_ref()
When you access this value in a SPARQL query, the select function recognizes the type of the variable and automatically inserts the full URI reference
e.select("""
SELECT ?s ?o {
?s ?_license ?o .
}
""")
Counting properties that are not about the Ontology¶
Since the metadata properties that describe this dataset really aren't part of it, it makes sense to remove these from the list so that we don't have so many properties that are used just once
properties=e.select("""
SELECT ?p (COUNT(*) AS ?cnt) {
?s ?p ?o .
FILTER(?s!=?_ontology)
} GROUP BY ?p ORDER BY DESC(?cnt)
""")
properties
At this point it is about as easy to make a pie chart as it is with Excel. A pie chart is a good choice here because each fact has exactly one property in it:
properties["cnt"].plot.pie(figsize=(6,6)).set_ylabel('')
My favorite method for understanding this kind of distribution is to sort the most common properties first and then compute the Cumulative Distribution Function, which is the percentage of facts that have used the predicates we've seen so far.
This is easy to compute with Pandas
100.0*properties["cnt"].cumsum()/properties["cnt"].sum()
Note that this result looks different than the DataFrames you've seen so far because it is not a DataFrame, it is a series, which has just one index column and one data column. It's possible to stick several series together to make a DataFrame, however.
pd.DataFrame.from_items([
('count',properties["cnt"]),
("frequency",100.0*properties["cnt"]/properties["cnt"].sum()),
("distribution",100.0*properties["cnt"].cumsum()/properties["cnt"].sum())
])
Unlike many graphical depictions, the above chart is fair to both highly common and unusually rare predicates.
Languages¶
It makes sense to start with rdf:label, which is the most common property in this database.
Unlike many data formats, RDF supports language tagging for strings. Objects (mainly properties and classes used in DBpedia) that are talked about in the DBpedia Ontology are described in multiple human languages, and counting the language tags involves a query that is very similar to the property counting query:
e.select("""
SELECT (LANG(?label) AS ?lang) (COUNT(*) AS ?cnt) {
?s rdfs:label ?label .
} GROUP BY LANG(?label) ORDER BY DESC(?cnt)
""")
A detail you might notice is that the lang column is not bolded, instead, a sequential numeric index was created when I made the data frame. This is because Gastrodon, at this moment, isn't smart enough to understand a function that appears in the GROUP BY clause.
This is easy to work around by assinging the output of this function to a variable in a BIND clause.
lang=e.select("""
SELECT ?lang (COUNT(*) AS ?cnt) {
?s rdfs:label ?label .
BIND (LANG(?label) AS ?lang)
} GROUP BY ?lang ORDER BY DESC(?cnt)
""")
lang
One key to getting correct results in a data analysis is to test your assumptions. English is the most prevalent language by far, but can we assume that every object has an English name? There are 3593 objects with English labels, but
distinct_s=one(e.select("""
SELECT (COUNT(DISTINCT ?s) AS ?cnt) {
?s rdfs:label ?o .
}
"""))
distinct_s
objects with labels overall, so there must be at least one object without an English label. SPARQL has negation operators so we can find objects like that:
black_sheep=one(e.select("""
SELECT ?s {
?s rdfs:label ?o .
FILTER NOT EXISTS {
?s rdfs:label ?o2 .
FILTER(LANG(?o2)='en')
}
}
"""))
black_sheep
Looking up all the facts for that object (which is a property used in DBpedia) shows that it has a name in greek, but not any other language
meta=e.select("""
SELECT ?p ?o {
?_black_sheep ?p ?o .
} ORDER BY ?p
""")
meta
I guess that's the exception that proves the rule. Everything else has a name in English, about half of the schema objects have a name in German, and the percentage falls off pretty rapidly from there:
lang_coverage=100*lang["cnt"]/distinct_s
lang_coverage
As the percentages add up to more than 100 (an object can have names in many languages), the pie chart would be a wrong choice, but a bar chart is effective.
lang_coverage.plot(kind="barh",figsize=(10,6))
Classes used in the DBpedia Ontology¶
I use another GROUP BY query to count classes used in the DBpedia Ontology. In the name of keeping the levels of abstraction straight, I'll point out that there are eight classes in the DBpedia Ontology, but that the DBpedia Ontology describes 739 classes used in DBpedia.
types=e.select("""
SELECT ?type (COUNT(*) AS ?cnt) {
?s a ?type .
} GROUP BY ?type ORDER BY DESC(?cnt)
""")
types
739 classes are really a lot of classes! You personally might be interested in some particular domain (say Pop Music) but to survey the whole thing, I need some way to pick out classes which are important.
If I had access to the whole DBpedia database, I could count how many instances of these classes occur, and that would be one measure of importance. (I have access to this database, as do you, but I'm not using it for this notebook because I want this notebook to be self-contained)
As it is, one proxy for importance is how many properties apply to a particular class, or, in RDFS speak, how many properties have this class as the domain. The assumption here is that important classes are well documented, and we get a satisfying list of the top 20 classes this way
types=e.select("""
SELECT ?s (COUNT(*) AS ?cnt) {
?s a owl:Class .
?p rdfs:domain ?s .
} GROUP BY ?s ORDER BY DESC(?cnt) LIMIT 20
""")
types
Adding another namespace binding makes sense to make the output more managable
g.bind("dbo","http://dbpedia.org/ontology/")
e=LocalEndpoint(g)
types=e.select("""
SELECT ?s (COUNT(*) AS ?cnt) {
?s a owl:Class .
?p rdfs:domain ?s .
} GROUP BY ?s ORDER BY DESC(?cnt) LIMIT 5
""")
types.head()
Common properties for People¶
To survey some important properties that apply to a dbo:Person I need some estimate of importance. I choose to count how many languages a property is labeled with as a proxy for importance -- after all, if a property is broadly interesting, it will be translated into many languages. The result is pretty satisfying.
person_types=e.select("""
SELECT ?p (COUNT(*) AS ?cnt) {
?p rdfs:domain dbo:Person .
?p rdfs:label ?l .
} GROUP BY ?p ORDER BY DESC(?cnt) LIMIT 30
""")
person_types
To make something that looks like a real report, I reach into my bag of tricks.
Since the predicate URI contains an English name for the predicate, I decide to show a label in German. The OPTIONAL clause is essential so that we don't lose properties that don't have a German label (there is exactly one in the list below). I use
a subquery to compute the language count, and then filter for properties that have more than one language.
e.select("""
SELECT ?p ?range ?label ?cnt {
?p rdfs:range ?range .
OPTIONAL {
?p rdfs:label ?label .
FILTER(LANG(?label)='de')
}
{
SELECT ?p (COUNT(*) AS ?cnt) {
?p rdfs:domain dbo:Person .
?p rdfs:label ?l .
} GROUP BY ?p ORDER BY DESC(?cnt)
}
FILTER(?cnt>4)
} ORDER BY DESC(?cnt)
""")
Towards a simple schema browser¶
You'd probably agree with me that the query above is getting to be a bit much, but now that I have it, I can bake it into a function which makes it easy to ask questions of the schema. The following query lets us make a similar report for any class and any language. (I use the German word for 'class' because the English word class and the synonymous word type are both reserved words in Python.)
def top_properties(klasse='dbo:Person',lang='de',threshold=4):
klasse=QName(klasse)
df=e.select("""
SELECT ?p ?range ?label ?cnt {
?p rdfs:range ?range .
OPTIONAL {
?p rdfs:label ?label .
FILTER(LANG(?label)=?_lang)
}
{
SELECT ?p (COUNT(*) AS ?cnt) {
?p rdfs:domain ?_klasse .
?p rdfs:label ?l .
} GROUP BY ?p ORDER BY DESC(?cnt)
}
FILTER(?cnt>?_threshold)
} ORDER BY DESC(?cnt)
""")
return df.style.highlight_null(null_color='red')
Note that the select here can see variables in the immediately enclosing scope, that is, the function definition. As it is inside a function definition, it does not see variables defined in the Jupyter notebook. The handling of missing values is a big topic in Pandas, so I take the liberty of highlighting the label that is missing in German.
top_properties()
In Japanese, a different set of labels is missing. It's nice to see that Unicode characters outside the latin-1 codepage work just fine.
top_properties(lang='ja')
And of course it can be fun to look at other classes and languages:
top_properties('dbo:SpaceMission',lang='fr',threshold=1)
About "prov:wasDerivedFrom"¶
The prov:wasDerivedFrom property links properties and classes defined in the DBpedia Ontology to the places where they are defined on the mappings web site.
e.select("""
SELECT ?s ?o {
?s prov:wasDerivedFrom ?o .
} LIMIT 10
""")
_.at[0,'o']
Subclasses¶
Subclasses can be queried with queries like the following, which lists direct subtypes of dbo:Person.
e.select("""
SELECT ?type {
?type rdfs:subClassOf dbo:Person .
}
""")
SPARQL 1.1 has property path operators that will make the query engine recurse through multiple rdfs:subClassOf property links.
e.select("""
SELECT ?type {
?type rdfs:subClassOf* dbo:Person .
}
""")
The previous queries work "down" from a higher-level class, but by putting a '^' before the property name, I can reverse the direction of traversal, to find all topics which dbo:Painter is a subclass of.
e.select("""
SELECT ?type {
?type ^rdfs:subClassOf* dbo:Painter .
}
""")
e.select("""
SELECT ?type {
dbo:Painter rdfs:subClassOf* ?type .
}
""")
The same outcome can be had by switching the subject and object positions in the triple:
e.select("""
SELECT ?type {
dbo:City rdfs:subClassOf* ?type .
}
""")
Equivalent Classes¶
The DBpedia Ontology uses owl:equivalentClass to specify equivalency between DBpedia Ontology types and types used in other popular systems such as wikidata and schema.org:
e.select("""
SELECT ?a ?b {
?a owl:equivalentClass ?b .
} LIMIT 10
""")
Here are all of the equivalencies between the DBpedia Ontology and schema.org.
e.select("""
SELECT ?a ?b {
?a owl:equivalentClass ?b .
FILTER(STRSTARTS(STR(?b),"http://schema.org/"))
}
""")
Many of these are as you would expect, but there are some that are not correct, given the definition of owl:equivalentClass from the OWL specification.
9.1.2 Equivalent Classes
An equivalent classes axiom EquivalentClasses( CE1 ... CEn ) states that all of the class expressions CEi, 1 ≤ i ≤ n, are semantically equivalent to each other. This axiom allows one to use each CEi as a synonym for each CEj — that is, in any expression in the ontology containing such an axiom, CEi can be replaced with CEj without affecting the meaning of the ontology. An axiom EquivalentClasses( CE1 CE2 ) is equivalent to the following two axioms:
SubClassOf( CE1 CE2 )
SubClassOf( CE2 CE1 )
Put differently, anything that is a member of one class is a member of the other class and vice versa. That's true for dbo:TelevisionEpisode and schema:TVEpisode, but not true for many cases involving schema:Product
g.bind("schema","http://schema.org/")
e=LocalEndpoint(g)
e.select("""
SELECT ?a ?b {
?a owl:equivalentClass ?b .
FILTER(?b=<http://schema.org/Product>)
}
""")
I think you'd agree that an Automobile is a Product, but that a Product is not necessarily an automobile. In these cases,
dbo:Automobile rdfs:subClassOf schema:Product .is more accurate.
Let's take a look at external classes which aren't part of schema.org or wikidata:
e.select("""
SELECT ?a ?b {
?a owl:equivalentClass ?b .
FILTER(!STRSTARTS(STR(?b),"http://schema.org/"))
FILTER(!STRSTARTS(STR(?b),"http://www.wikidata.org/"))
}
""")
To keep track of them all, I add a few more namespace declarations.
g.bind("dzero","http://www.ontologydesignpatterns.org/ont/d0.owl#")
g.bind("dul","http://www.ontologydesignpatterns.org/ont/dul/DUL.owl#")
g.bind("bibo","http://purl.org/ontology/bibo/")
g.bind("skos","http://www.w3.org/2004/02/skos/core#")
e=LocalEndpoint(g)
e.select("""
SELECT ?a ?b {
?a owl:equivalentClass ?b .
FILTER(!STRSTARTS(STR(?b),"http://schema.org/"))
FILTER(!STRSTARTS(STR(?b),"http://www.wikidata.org/"))
}
""")
The mapping from dbo:Film to <http://dbpedia.org/ontology/Wikidata:Q11424> is almost certainly a typo.
Disjoint Classes¶
Another bit of OWL vocabulary used in the DBpedia Ontology is owl:disjointWith
e.select("""
SELECT ?b ?a {
?a owl:disjointWith ?b .
} ORDER BY ?b
""")
If two classes are disjoint, that means that nothing can be an instance of both things. For instance, a Fish cannot be a Mammal, a Person is not a building, etc. These sort of facts are helpful for validation, but one should resist the impulse to make statements of disjointness which aren't strictly true. (For instance, it would be unlikely, but not impossible, to be the winner of both a Heisman Trophy and a Fields Metal, so these are not disjoint categories.)
Datatypes¶
RDF not only has "types" (classes) that represent named concepts, but it also has literal datatypes. These include the standard datatypes from XML such as xsd:integer and xsd:datetime, but also derived types that specialize those types. This makes it possible to tag quantities in terms of physical units, currency units, etc.
e.select("""
SELECT ?type {
?type a rdfs:Datatype .
} LIMIT 10
""")
g.bind("type","http://dbpedia.org/datatype/")
e=LocalEndpoint(g)
e.select("""
SELECT ?type {
?type a rdfs:Datatype .
} LIMIT 10
""")
The information about these data types are currently sparse: this particular type has just a label and a type.
e.select("""
SELECT ?p ?o {
type:lightYear ?p ?o .
}
""")
These turn out to be the only properties that any datatypes have; pretty clearly, datatypes are not labeled in the rich set of languages that properties and classes are labeled in. (Note that vocabulary exists in RDFS and OWL for doing just that, such as specifying that type:lightYear would be represented as a number, specifying that a particular type of numeric value is in a particular range, etc.)
e.select("""
SELECT ?p (COUNT(*) AS ?cnt) {
?s a rdfs:Datatype .
?s ?p ?o .
} GROUP BY ?p
""")
Another approach is to look at how datatypes get used, that is, how frequently various datatypes are used as the range of a property.
e.select("""
SELECT ?type (COUNT(*) AS ?cnt) {
?p rdfs:range ?type .
?type a rdfs:Datatype .
} GROUP BY ?type ORDER BY DESC(?cnt)
""")
len(_)
Out of 382 properties, only 41 actually appear as the range of the properties in the schema. Here are a few properties that are unused in the schema.
e.select("""
SELECT ?type {
?type a rdfs:Datatype .
MINUS { ?s ?p ?type }
} LIMIT 20
""")
According to the DBpedia Ontology documentation, there are two kinds of datatype declarations in mappings. In some cases the unit is explictly specified in the mapping field (ex. a field that contains a length is specified in meters) and in other cases, a particular datatype is specific to the field.
It turns out most of the knowledge in the DBpedia Ontology system is hard coded into a scala file; this file contains rich information that is not exposed in the RDF form of the Ontology, such as conversion factors, the fact that miles per hour is a speed, etc.
It is quite possible to encode datatypes directly into a fact, for example,
:Iron :meltsAt "1811 K"^^type:kelvin .
It is possible that such facts could be found in DBpedia or some other database, but I'm not going to check for that in this notebook, because this notebook is only considering facts that are in the ontology file supplied with this notebook.
Properties Measured in Kilograms¶
e.select("""
SELECT ?p {
?p rdfs:range type:kilogram
}
""")
One unfortunate thing is that the DBpedia ontology sometimes composes property URIs by putting together the class (ex. "Galaxy") and the property (ex. "mass") with a slash between them. Slash is not allowed in a localname, which means that you can't write ontology:Galaxy/mass. You can write the full URI, or you could define a prefix galaxy such that you can write Galaxy:mass. Yet another approach is to set the base URI to
in which case you could write <Galaxy/mass>. I was tempted to do that for this notebook, but decided against it, because soon I will be joining the schema with more DBpedia data, where I like to set the base to
In a slightly better world, the property might be composed with a period, so that the URI is just "ontology:Galaxy.mass". (Hmm... Could Gastrodon do that for you?)
Datatype properties vs Object Properties¶
RDFS has a single class to represent a property, rdf:Property; OWL makes it a little more complex by defining both owl:DatatypeProperty and owl:ObjectProperty. The difference between these two kinds of property is the range: a Datatype Property has a literal value (object), while an Object Property has a Resource (URI or blank node) as a value.
I'd imagine that every rdf:Property should be either an owl:DatatypeProperty or owl:ObjectProperty, so that the sums would match. I wouldn't take it for granted, so I'll check it:
counts=e.select("""
SELECT ?type (COUNT(*) AS ?cnt) {
?s a ?type .
FILTER (?type IN (rdf:Property,owl:DatatypeProperty,owl:ObjectProperty))
} GROUP BY ?type ORDER BY DESC(?cnt)
""")["cnt"]
counts
counts["rdf:Property"]
counts["owl:DatatypeProperty"]+counts["owl:ObjectProperty"]
The sums don't match.
I'd expect the two kinds of DatatypeProperties to be disjoint; and they are, because I can't find any classes which are an instance of both.
e.select("""
SELECT ?klasse {
?klasse a owl:DatatypeProperty .
?klasse a owl:ObjectProperty .
}
""")
However, there are cases where a property is registered as an OWL property but not as an RDFS property:
e.select("""
SELECT ?klasse {
?klasse a owl:DatatypeProperty .
MINUS {?klasse a rdf:Property}
}
""")
e.select("""
SELECT ?klasse {
?klasse a owl:ObjectProperty .
MINUS {?klasse a rdf:Property}
}
""")
However, there are no properties defined as an RDFS property that are not defined in OWL.
e.select("""
SELECT ?p {
?p a rdf:Property .
MINUS {
{ ?p a owl:DatatypeProperty }
UNION
{ ?p a owl:ObjectProperty }
}
}
""")
Conclusion: to get a complete list of properties defined in the DBpedia Ontology, is necessary and sufficient to use the OWL property declarations. The analysis above that uses rdfs:Property should use the OWL property classes to get complete results.
Subproperties¶
Subproperties are used in RDF to gather together properties that more or less say the same thing.
For instance, the mass of a galaxy is comparable (in principle) to the mass of objects like stars and planets that make it. Thus in a perfect world, the mass of a galaxy would be related to a more general "mass" property that could apply to anything from coins to aircraft carriers.
I go looking for one...
galaxyMass=URIRef("http://dbpedia.org/ontology/Galaxy/mass")
e.select("""
SELECT ?p {
?_galaxyMass rdfs:subPropertyOf ?p .
}
""")
... and don't find it. That's not really a problem, because this I can always add one by adding a few more facts to my copy of the DBpedia Ontology. Let's see what is really there...
e.select("""
SELECT ?from ?to {
?from rdfs:subPropertyOf ?to .
}
""")
It looks like terms on the left are always part of the DBpedia Ontology:
e.select("""
SELECT ?from ?to {
?from rdfs:subPropertyOf ?to .
FILTER(!STRSTARTS(STR(?from),"http://dbpedia.org/ontology/"))
}
""")
Terms on the right are frequently part of the http://ontologydesignpatterns.org/wiki/Ontology:DOLCE%2BDnS_Ultralite (DUL)
ontology and are a way to explain the meaning of DBpedia Ontology terms in terms of DUL. Let's look at superproperties that aren't from the DUL ontology:
e.select("""
SELECT ?from ?to {
?from rdfs:subPropertyOf ?to .
FILTER(!STRSTARTS(STR(?to),"http://www.ontologydesignpatterns.org/ont/dul/DUL.owl#"))
}
""")
Out of those 75 relationships, I bet many of them point to the same superproperties:
e.select("""
SELECT ?to (COUNT(*) AS ?cnt) {
?from rdfs:subPropertyOf ?to .
FILTER(!STRSTARTS(STR(?to),"http://www.ontologydesignpatterns.org/ont/dul/DUL.owl#"))
} GROUP BY ?to ORDER BY DESC(?cnt)
""")
The most common superproperty is dbo:code, which represents identifying codes. For instance,
this could be a postal Code, UPC Code, or a country or regional code. Unfortunately, only a small number of code-containing fields are so identified.
e.select("""
SELECT ?about ?from {
?from
rdfs:subPropertyOf dbo:code ;
rdfs:domain ?about .
}
""")
Looking at the superproperty dbo:closeTo, the subproperties represent (right-hand) locations that are adjacent to (left-hand) locations in the directions of the cardinal and ordinal (definition) directions.
e.select("""
SELECT ?about ?from {
?from
rdfs:subPropertyOf dbo:closeTo ;
rdfs:domain ?about .
}
""")
Looking a the superproperties in DUL, these look much like the kind of properties one would expect to defined in an upper or middle ontology:
e.select("""
SELECT ?to (COUNT(*) AS ?cnt) {
?from rdfs:subPropertyOf ?to .
FILTER(STRSTARTS(STR(?to),"http://www.ontologydesignpatterns.org/ont/dul/DUL.owl#"))
} GROUP BY ?to ORDER BY DESC(?cnt)
""")
A really common kind of property is a "part-of" relationship, known as meronymy if you like greek.
e.select("""
SELECT ?domain ?p ?range {
?p
rdfs:subPropertyOf dul:isPartOf ;
rdfs:domain ?domain ;
rdfs:range ?range .
}
""")
Equivalent Property¶
The case of "part of" properties is a good example of a subproperty relationship in that, say, "Mountain X is a part of Y Mountain range" is clearly a specialization of "X is a part of Y." That's different from the case where two properties mean exactly the same thing.
Let's take a look at equivalent properties defined in the DBpedia Ontology:
e.select("""
SELECT ?a ?b {
?a owl:equivalentProperty ?b
}
""")
Many of these properties are from Wikidata, so it probably makes sense to bind a namespace for Wikidata.
g.bind("wikidata","http://www.wikidata.org/entity/")
e=LocalEndpoint(g)
This kind of equivalency with Wikidata is meaningful precisely because DBpedia and Wikidata are competitive (and cooperative) databases that cover the same domain. Let's take a look at equivalencies to databases other than Wikidata:
e.select("""
SELECT ?a ?b {
?a owl:equivalentProperty ?b
FILTER(!STRSTARTS(STR(?b),"http://www.wikidata.org/entity/"))
}
""")
The vast number of those link to schema.org, except for a handful which link to other DBpedia Ontology properties.
e.select("""
SELECT ?a ?b {
?a owl:equivalentProperty ?b
FILTER(STRSTARTS(STR(?b),"http://dbpedia.org/ontology/"))
}
""")
The quality of these equivalencies are questionable to me; for instance, in geography, people often publish separate "land area" and "water areas" for a region. Still, out of 30,000 facts, I've seen fewer than 30 that looked obviously wrong: an error rate of 0.1% is not bad on some terms, but if we put these facts into a reasoning system, small errors in the schema can result in an avalanche of inferred facts resulting in a disproportionately large impact on results.
Namespaces¶
Rather than starting with a complete list of namespaces used in the DBpedia Ontology, I gradually added them as they turned up in queries. It would be nice to have a tool that automatically generates this kind of list, but for the time being, I am saving this list here for future reference.
e.namespaces()
Conclusion and next steps¶
In this notebook, I've made a quick survey of the contents of the DBpedia Ontology. This data set is useful to build into the "local" tests for Gastrodon because it is small enough to work with in memory, but complex enough to be a real-life example. For other notebooks, I work over the queries and data repeatedly to eliminate imperfections that make the notebooks unclear, but here the data set is a fixed target, which makes it a good shakedown cruise for Gastrodon in which I was able to fix a number of bugs and make a number of improvements.
One longer term goal is to explore data from DBpedia and use it as a basis for visualization and data analysis. The next step towards that is to gather data from the full DBpedia that will help prioritize the exploration (which properties really get used?) and answer some questions that are still unclear (what about the data types which aren't used the schema?)
Another goal is to develop tools to further simplify the exploration of data sets and schemas. The top_properties function defined above is an example of the kind of function that could be built into a function library that would reduce the need to write so many SPARQL queries by hand.