Type inference through the analysis of Wikipedia links

Abstract

DBpedia contains millions of untyped entities, either if we consider the native DBpedia ontology, or Yago plus WordNet. Is it possible to automatically classify those entities? Based on previous work on wikilink invariances, we wondered if wikilinks convey a knowledge rich enough for their classi cation. In this paper we give three contributions. Concerning the DBpedia link structure, we describe some measurements and notice both problems (e.g. the bias that could be induced by the incomplete ontological coverage of the DBpedia ontology), and potentials existing in current type coverage. Concerning classi cation, we present two techniques that exploit wikilinks, one based on induction from machine learning techniques, and the other on abduction. Finally, we discuss the limited results of classi cation, which con rmed our fears expressed in the description of general gures from the measurement. We also suggest some new possible directions to entity classi cation that could be taken.

1. INTRODUCTION

DBpedia is the largest RDF data set in Linked Data extracted from the largest existing multi-domain knowledge source edited by the crowds, i.e. Wikipedia. Part of DBpedia entities are explicitly typed with classes of the DBpedia Ontology (DBPO). This huge source of semantic data provides a powerful knowledge base that can be exploited as background knowledge for developing new generation of knowledge extraction and interaction tools. Nevertheless, most of DBpedia entities are still untyped, which limits its exploitation at its full potential. Based on previous work on wikilink invariances [ 13 ], we wondered if wikilinks convey a knowledge rich enough for their automatic classi cation. In this paper, we discuss both problems and potentials deriving from the current type coverage of DBpedia entities. Additionally, we present two automatic classi cation techniques that exploit wikilinks, one based on indiction from machine learning, and the other based on abduction. Both methods showed limited results in terms of performance (i.e., precision and recall), which can be explained by analyzing the general gures from our measurements on the DBpedia link structure. More speci cally, (i) the mapping procedure between Wikipedia infobox templates and DBpedia ontology classes is conducted manually, as described in [ 9 ], resulting in a lack of ontological coverage of DBPO i.e. DBPO classes are insu cient for typing all DBpedia entities, which in turn results in lack of classi cation knowledge that can be used by automatic techniques for identifying type candidates for untyped entities, (ii) the granularity of the assigned types is not homogeneous e.g. some entities are typed with very general classes e.g. Person, while other similar entities have more specialized types e.g. Musician, (iii) most of the entities are untyped, hence making it hard to build a proper training set for inductive learning techniques, (iv) the distribution of link types ingoing to, and outgoing from, DBpedia entities varies between typed and untyped entities, which impacts on the ability of our abductive method to properly classify untyped entities.

After describing the resources that we have used for our research and discussing limitations and potentials emerging from our measurements on the general gures of the DBpedia link structure (Section 2), we present two methods for automatic classi cation of DBpedia entities and the results we have obtained (Section 3). In Section 4 we discuss related work and in Section 5 we conclude by discussing possible new approaches to DBpedia entity classi cation that could be taken, based on more solid grounds.

2. MATERIALS

DBpedia datasets describe about 18 million resources (3 million with an abstract and a label, less than 2 million typed), and include more than 107 million wikilinks.

The classi cation attempted in this paper is based on the assumption that wikilink relations in Wikipedia convey a rich knowledge that can be used to classify untyped entities referenced in those pages. In practice, given a certain entity The reason of this unbalance can be hypothesized to be caused by the high frequency of homotypes, i.e. wikilinks that have the same type on both the subject and the object of the triple. If this hypothesis is con rmed, untyped resources should have a high ratio of untyped outgoing links.

As a matter of fact, homotypes are actually very frequent (usually the most frequent, or in the top 3) wikilink types (this observation has been made in the research reported by [ 13 ]). Therefore, such distribution of wikilinks for typed and untyped resources is unbalanced.

This means that if we use wikilinks and types as the features for training or designing a good inductive model based on the corpus of typed resources, a bias is created on the appropriateness of such a model for classifying untyped resources.

However, we wanted to check anyway: i) what precision/recall can be obtained when using wikilink structures as features for creating a type induction model from typed DBpedia resources, even considering the bias constituted by the 34% untyped wikilinks; and ii) how much the larger bias (63%), constituted by untyped wikilinks on untyped resources, would a ect the precision/recall established on typed resources.

A part of our wikilink analysis for DBpedia entity classi cation made use of 187 Knowledge Patterns that have been extracted from DBpedia wikilink datasets, which are called Encyclopedic Knowledge Patterns (EKPs) 4 [ 13 ]. EKPs allow to fetch most relevant entity types that provide an e ective and intuitive description of entities of a certain type. As discussed in the next section, EKPs provide a background knowledge to our method of abductive classi cation. described in a Wikipedia page, our classi cation grounds on the analysis of incoming and outgoing wikilinks from and to that certain page.

For this analysis we used DBpedia [ 9 ], the RDF [ 10 ] Linked Data [ 1 ] that contains structured information extracted from Wikipedia.

The types used to classify DBpedia resources are the classes of the DBpedia ontology 1 (DBPO). DBPO ontology is represented in OWL [ 11 ] and covers 272 classes.

The rdf:type statements from DBpedia resources to DBPO classes (available in the dataset dbpedia_instance_types_en) are the result of hand-generated mappings of Wikipedia in- 3. METHODS AND RESULTS foboxes to DBPO, and have been generated for 1,668,503 The classi cation of DBpedia entities relies on an ontology DBpedia resources2. mapping task which de nes how Wikipedia infobox temThose statements cover only a subset of the 15,944,381 DB- plates are mapped to classes of the DBpedia ontology. These pedia resources available in the dbpedia_page_links_en dataset. mappings are manually speci ed using the DBpedia MapHence, only 15.52% of resources in the DBpedia data set of ping Language 5. The mapping language makes use of Mewikilinks are typed with DBPO classes. In the work de- diaWiki templates that allow to map infoboxes to DBpedia scribed in this paper we investigate how to assign a DBPO ontology classes. The mappings cover only a small subset of type to the remaining 84.48% untyped DBpedia resources.3. all Wikipedia infoboxes. As a result, so far, only a small subTable 2 shows further details about how resources are orga- set of all DBpedia entities (1,668,503 of 15,944,381) is typed nized in DBpedia. It emerges that the number of wikilinks with a class of the DBpedia ontology. Probably the e ort is much bigger (107,892,317) than the number of rdf:type spent in manually writing mappings for the classi cation triples (6,173,940). When we take into account only DBPO of DBpedia entities with respect to the DBpedia ontology type we can observe how the number of Wikilink axioms is too expensive and the granularity and the appropriatehaving both the subject and the object typed with DBPO ness of obtained typings are not exhaustive. As an example, types is limited to 16,745,830. dbpedia:Walt_Disney 6 is typed as dbpo:Person, which is doubtlessly correct, but also trivial and less appropriate than the existing dbpo:ComicsCreator type.

Our work is based on the intuition that wikilink relations in DBpedia, i.e. instances of the dbpo:wikiPageWikiLink property, convey a rich knowledge that can be used for classifying DBpedia entities, but at the same time it is very di cult to nd a good training set by using only typed wikilinks. A rst reason is that typed resources having wikilinks are only the 15.52% of the total resources used in the wikilink data set. A second one derives from the fact untyped resources, i.e., the resources we are interested to type, have Furthermore, we analyzed the distribution of wikilinks across typed/untyped resources. The average percentage of fully DBPO typed wikilinks (i.e., wikilink axioms having both the subject and the object typed with a DBPO class) per resource is 66% with respect to the total number of wikilink axioms per resource. Instead, the average percentage of wikilink axioms outgoing from untyped to typed resources is 37% per resource with respect to the total number of wikilink axioms per resource. This means that the ratio between typed and untyped outgoing links is 23 for typed resources vs. 13 for untyped ones. 1http://wiki.dbpedia.org/Ontology 2such gure holds for the typed resources in the English version of DBpedia 3.6 3Actually, excluding some entities that are not relevant: images, categories, and disambiguation pages 4The EKP resource is available as a set of OWL ontologies at http://ontologydesignpatterns.org/ekp/owl/ 5http://mappings.dbpedia.org/index.php/Main Page 6The pre xes dbpo and dbpedia stand for http://dbpedia.org/ontology/ and http://dbpedia.org/resource/ respectively. only one third of their total wikilinks typed. In order to test this intuition, we have investigated and compared type induction methods over DBpedia entities based on two inference types:

Inductive classi cation: works moving from speci c observations to broader generalizations and theories, and can be informally de ned as a \bottom-up" approach; Abductive classi cation: works from more general rules (assumed from previous cases), in order to infer presumably speci c facts. In this case we have used EKPs and homotypes as background knowledge.

Considering that DBPO has 272 classes, and that automatic classi cation on 272 classes is very di cult, we have focused mainly on classi cation of entities with respect to the DBPO top-level. The granularity of the classi cation is solved in this case by adopting a hierarchical-iterative type induction derived by the class hierarchy in DBPO. This means that the classi cation starts from the top level of the DBpedia ontology composed by 27 classes, and then iteratively goes on trying to classify an entity with one of the sub-classes of the class selected in the previous iteration.

3.1 Inductive type inference

Hybridization of Machine Learning (ML) techniques with the Semantic Web is quite e ective [ 5 ], therefore we started our investigation trying to use a ML-based approach to the classi cation of DBpedia entities. We have used the kNearest Neighbor (NN) algorithm for classifying DBpedia entities based on the closest training examples in the labeled feature space, and by assigning the most voted class among the training examples.

We have designed two inductive classi cation experiments based on the NN algorithm: (i) a baseline experiment con gured to classify DBpedia individuals based on 272 features, i.e., all the DBPO classes. (ii) An experiment based on the top-level classes in the DBPO taxonomy, i.e., 27 classes, aimed to simplify the classi cation with less features to investigate. In both cases the training sample has been built with the same approach described as follows: (i) the 20% of the individuals of each class has been used for populating the training sample. Table 3.1 shows how many individuals have been chosen with respect to the classes Place, Person, Work, Organisation and Activity. (ii) The NN algorithm has been trained on a labeled feature space model in which the individuals of the training sample and the classes of the

DBPO were the rows and the feature columns of the model respectively. For each individual we labeled the corresponding row with its known DBPO class and we then analyzed its graph of wikilinks in order to ll the matrix resulting from the space-vector model. This was done marking with either 0 or 1 each intersection cell between a row corresponding to an individual and a feature corresponding to a DBPO class with the following criteria: 0 means that no wikilink exists between the selected individual and any other individual typed with the corresponding DBPO class used as feature; 1 means that at least one wikilink exists between the selected individual and any other individual typed with the corresponding DBPO class used as a feature.

As an example, we may want to built the feature model of the wikilinks of the entity dbpedia:Steve_Jobs with respect to the classes dbpo:Mammal, dbpo:Scientist, dbpo:Company, dbpo:Drug, dbpo:City, dbpo:Magazine. We fetch all the types related to the outgoing wikilinks from dbpedia:Steve_Jobs with a simple SPARQL query like the following: PREFIX dbpedia: <http://dbpedia.org/resource/> SELECT ?link ?type WHERE {

GRAPH<dbpedia_page_links_en> {

dbpedia:Steve_Jobs ?prop ?link } .

GRAPH<dbpedia_instance_types_en> {

?link a ?type }

} Supposing that all the wikilinks retrieved (actually, those shown are only a subset) are the ones showed in table 3.1 the resulting row in feature space model deriving from dbpedia:Steve_Jobs will be the following:

Mammal Scientist Company Drug City Magazine Steve_Jobs 0 0 1 0 1 1 Person After the training, for each class the classi cation function generates what is called a mean, i.e., the reference value used to test similarities and then to classify untyped entities. The classi cation has been performed by estimating the euclidean distance of the features of unlabeled individuals with respect to each mean. The lower values of euclidean distance have been chosen to classify individuals.

The performance of this approach has been evaluated on the remaining 80% untyped individuals. We remind that such evaluation is made considering the \best case" of resources with a known type, in which the percentage of typed wikilinks is high (66%). The precision deriving from the baseline experiment has been 31.65%. The precision of the classi cation of individuals with respect to the DBPO top-level has been 40.27%.

3.2 Abductive type inference

Abduction was introduced by Peirce [ 7, 14 ], and refers to a process oriented to an explanatory hypothesis about the precondition P reasoning on the consequence C. Compared to pure deduction, induction and abduction have a lower strength, much are practical when the set of assumptions is not complete with respect to the observed world. Induction cannot be made fully conclusive, since the inference from a set of cases can be only made certain in a closed world (which is not the case with the Web). Abduction, on its turn, is formally equivalent to the logical fallacy a rming the consequent or Post hoc ergo propter hoc, because there are multiple possible explanations for C. In other words, abducing P from C involves determining that C is su cient (or nearly su cient), but not necessary, for P.

We have used an abductive approach to infer the type of DBpedia entities with two classi cation methods: EKP-based. We assumed Encyclopedic Knowledge Patterns (EKPs) extracted from wikilink relations in Wikipedia [ 13 ] as our background knowledge, and as the adbuctive consequence C, on which we infer the type of entities. In this context, entity types are our precondition P. We want to infer types by analysing the similarity between rules derived by EKPs, and the con gurations of wikilink relations obtained from untyped entities in DBpedia.

Homotype-based. We de ne homotypes as wikilinks that have the same type on both the subject and the object of the triple. Since homotypes are usually the most frequent (or in the top 3) wikilink types [ 13 ], we want to detect emerging homotypes for untyped resources by summing the number of outgoing and incoming wikilinks.

EKP-based type adbuction. EKPs are de ned as sets of type paths that occur most often above a certain threshold t [ 13 ]. In this context, a path relates the types of two entities having a wikilink. The frequency of paths composing an EKP is calculated by the path popularity. The path popularity is de ned as the ratio of how many distinct resources of a certain type participate as subject in a path to the total number of resources of that type. We use path popularity values of each EKP in order to estimate the similarity between the wikilinks of an entity and a EKP. For doing that we build a labeled feature space model i j composed by: i rows, each one labeled by a di erent top-level class of the DBPO taxonomy. This means 0 i 27; j features as columns consisting in the number of all the classes in the DBPO. This means 0 j 272; cell values that contain path popularity values occurring between the type in the row and the type in the column.

The following is an example of feature space model built for the class labels Person, Place and Oganisation over the features Event, Work, Organisation, Person, Activity, Place:

Event Work . . . . . . . . .

Person 4.45 8.4 Organisation 3.21 7.78 Place 3.14 1.86 . . . . . . . . .

Organisation Person Activity Place . . . . . . . . . . . . 22.2 18.29 6.8 27.5 24.13 13.23 2.5 31.93 8.71 8.74 1.1 61.15 . . . . . . . . . . . .

In the abductive approach, di erently from the inductive one, the labeled feature space model is not used for training the classi cation function. In fact, it already provides general rules based on path popularity values over the wikilink domain used to infer types.

Instead, entities to be classi ed are represented in the same way as in the inductive approach. Therefore, they are vector models of length j, where j is the number of features used in the model. Values in vectors consists of either 0 or 1, in which, again: 0 means that no wikilink exists between the selected individual and any other individual typed with the corresponding DBPO class used in the feature; 1 means that at least one wikilink exists between the selected individual and any other individual typed with the corresponding DBPO class used in the feature. Individuals are classi ed by nding similarities to EKPs by applying a similarity metric based on the analysis of path popularity values. This suggests what is the closest EKP to the con guration of wikilinks that the individual presents. Given a labeled feature space model M and a vector I dened as described before, the similarity function is de ned as follows:

S(i) =

PF j=0 Mi;j Ij PF j=0 Mi;j where

F is the number of DBPO classes used as features;

Mi;j is the path popularity value between the i-th class used as label and the j-th class used as feature; Ij is the 0 or 1 value that corresponds to the fact that the entity has or has not a wikilink relation with an entity typed with the j-th class used as feature; F is the number of features used in the model and 0 j F ; 0 i L, where L is the number of available labels in the model. (a) Results of the abductive type inference experiment based on DBpedia top-level classes used both as labels and features.

(b) Results of the abductive type inference experiment based on classes Activity, Device, Event, Infrastructure, MeanOfTransportation, Organisation, Person, Place and Work used both as labels and features.

The function S returns a similarity score calculated locally with respect to a type, while we are interested in the score that maximizes the similarity, hence:

T = max S(i); 8 0 i

L Assuming that for a certain entity Y with respect to the types Event, Work, Organisation, Person, Activity, Place has the following wikilink con guration:

Y

Event Work 0 1

Organisation Person Activity Place 0 1 0 0 We obtain that the entity Y is classi ed as Person, since the value S(P erson) = (1 8:4 + 1 18:29)=87:64 = 0:3 is the highest similarity value.

In our rst abductive experiment we have focused only on the top-level classes of the DBpedia taxonomy both for class labels and for features. Both L and F are then equal to the number of the top-level classes in the DBpedia taxonomy, i.e., 27.

The precision and recall of this experiment are both, as shown in gure 1(a), 44.4%. The gure shows, besides the precision and recall relative to single classes, additional metrics, like the number of true positives, false positives, false negatives, etc.

What also emerges from gure 1(a) is that a small subset (Device, Event, Infrastructure, MeanOfTransportation, Organisation, Person, Place and Work) of classes contains the majority of the individuals. For that reason, we have tried to apply the abductive approach only to that subset plus the class Activity, which instead has a very low precision.

In this second experiment we have used those 9 classes as labels for classifying DBpedia entities over the same 9 classes as features. Under these assumptions, the global precision has fallen from 44.4% of the previous experiment down to 36.5, while the recall has grown from 44.4% up to 79.5%.

Figure 1(b) shows the results of this experiment.

Homotype-based abduction. In this case we use abduction in order to guess the type of DBpedia individuals by using homotypes as background knowledge. An homotype is usually the most frequent (or in the top 3) wikilink type.

For that reason, we want to infer the type of a resource by detecting the emerging homotype. If most of the incoming and outgoing wikilinks of a resource R is of a same type T , then we can infer, under the homotype assumption, that the type of R is T .

Given an individual i, we can then de ne the set W of all the classes used to type individuals having an incoming or an outgoing wikilink relation with i, as:

W (i) = fhn; Xijn = #x 2 X s.t i ! x _ i

xg where the notation 2 stands here for the rdf:type property; the ! and stand for outgoing and incoming dbpo:wikiPageWikiLink properties; n is the number of wikilinks typed by a same class X; < n; X > is any distinct couple that states how many instances of the class X occur in the wikilinks of i.

The homotype range H, emerging for an individual i, is formalized as the type X having the highest value n in W (i), i.e.,

H(i) = fX j hn; Xi 2 W (i) =) 8hn0; X0i 2 W (i) ; n

n0g where, the notation 2 has here the classical meaning from set theory. Figure 2 shows how the homotype range is selected grouping the sum of outgoing and incoming wikilinks for the resource dbpedia:Steve_Jobs grouped by their types.

According to the de nition of homotype given, the resource dbpedia:Steve_Jobs as shown in gure 2 should be typed with the class dbpo:Person.

For the classi cation experiment based on the homotype definition, we have de ned a threshold that represents the minimum number of wikilinks that a resource should have in order to be classi able. In fact, below a certain number of wikilinks, the homotype is less distinctive. The threshold has been xed to be = 10 because, as reported by [ 13 ], the average number of outgoing wikilinks per resource in the dbpedia_page_links_en data set of DBpedia is 10.

The homotype-based classi cation of individuals from the control set (with a known type) from the dbpedia_instance_type_en data set produced a global precision of 52.14% and a global recall of 85.87%. Figure 3 shows the details regarding global and local precision, recall and also reports the number of true positives, false positives and false negatives.

The homotype-based approach produces a side-e ect. In fact, more than one class can emerge with the same frequency of wikilinks in H. In some cases, we get ex-aequo classi cations, i.e. multi-typings. Multi-typing introduces noise that produces a higher number of false positives. For example, if an entity e, whose actual type is T1, is classi ed with types T1, T2 and T3, then T1 will be counted as a true positive, but at the same time T2 and T3 will be counted as false negatives. In general multi-typings are not desirable. Figure 4 shows the number of ex-aequo typings for each cluster found, i.e. 88,845 occurrences with 2 ex-aequo classes, 11,572 occurrences with 3 ex-aequo classes, 1,118 occurrences with 4 ex-aequo classes, 44 occurrences with 5 ex-aequo classes and, nally, 1 occurrence with 6 ex-aequo classes, for a total of 101,580 ex-aequo classi cations.

In order to reduce the noise of multi-typings in the evaluation of the homotype-based approach, we have adjusted the performance analysis by applying the following criteria in case of ex-aequo: if among the ex-aequo classes there is the actual type of the entity, then count 1 true positive and 0 false positives; if among the ex-aequo classes there is not the actual type of the entity, then count 0 true positives and 1 false positive;

This increased the precision up to 55.07%.

Figure 5 shows the trend of the precision and recall through the various experiments. Considering the abductive approach has reported the best results, we have decided to run both the EKP-based and homotype-based classi cation on a sample of 1,000 untyped resources, to be manually evaluated. The classi er was again limited to the 9 core top-level classes as described before.

Manual evaluation reported precision and recall with EKPbased classi cation as 5.98% and 7.9% respectively, while with homotype-based classi cation as 13.93% and 65.51% respectively. Figure 6 shows the details of the results of the classi cations performed on the sample of 1,000 untyped resources for the 9 top-level classes.

4. RELATED WORK

There is valuable research on knowledge extraction based on the exploitation of Wikipedia. Ontology mining aims at discovering hidden knowledge from ontological knowledge bases by using data mining and machine learning techniques [ 5 ]. [ 4 ] proposes an extension of the k-Nearest Neighbor algorithm for Description Logic KBs based on the exploitation of an entropy-based dissimilarity measure, while [ 2, 6 ] makes use of Support Vector Machine (SVN) [ 3 ] rather than NN to perform automatic classi cation. SVM performs instance classi cation by implicitly mapping (through a kernel function) training data to the input values in a higher dimensional feature space where instances can be classi ed by means of a linear classi er. The main di erence between these and our approach is that they analyse all semantic properties used for linking individuals, while we have look for invariances on the usage of only one property, i.e., dbpo:wikiPageWikiLink, which attens the reasoning space. The current procedure for assigning type to DBpedia entity is completely manual. As extensively described in [ 9 ], a limited number of infobox templates have been de ned based on empirical observation of invariances in infoboxes of Wikipedia pages having the same ontological type. Based on previous work on wikilink invariances [ 13 ], we investigate the automatic classi cation of DBpedia entities. The ontology that we use for the classi cation is the DBpedia Ontology (DBPO) that results from the manual procedure described above, hence we inherit its limited ontological coverage. [ 16 ] presents YAGO, an ontology extracted from Wikipedia categories and infoboxes that has been combined with taxonomic relations from WordNet. Here the approach can be described as a reengineering task for transforming a thesaurus, i.e. Wikipedia category taxonomy, to an ontology, (a) Results of the EKP-based classi cation of the sample of 1,000 untyped resources (b) Results of the Homotype-based classi cation of the sample of 1,000 untyped resources which required accurate ontological analysis. There is a signi cant overlap between YAGO and DBPO typed entities, and entities that have only YAGO classes cover a small part of the entities untyped with DBPO. Furthermore, there is a lack of mapping between YAGO and DBPO, which makes it di cult to exploit their merged coverage. Yago has a larger coverage than DBPO, but it has only an overlap with DBPO coverage; moreover, the granularity of Yago categories is ner, and not easily reusable, because the top level is very large. The size of the ontology, and the fact that Yago adopts multi-typing, complicate helplessly the task of automatic classi cation of types.

5. DISCUSSION AND CONCLUSION

We want to discover the types of all DBpedia resources. Currently only a subset of them (about 15%, about 22% in recent version 3.7) have explicit types that are coherently organized into an ontology (DBpedia Ontology, DBPO). We investigated type inference in two directions: (i) an inductive approach typical of machine learning; (ii) an abductive approach, in which we rstly used two already available feature sets over the wikilink domain in order to infer types: (a) the EKPs [ 13 ] that have been extracted from Wikipedia with a statistical analysis over type paths generated by wikilinks, and (b) the notion of emerging homotype, i.e. a link that has the same types in the subject and object of the RDF triple. We observed two possible classi cation biases in the DBpedia datasets. The rst bias is the ratio between the number of typed resources having wikilinks and the total number of resources with wikilinks, i.e., 1,518,697 out of more than 15 million. We suspected that this bias may be the result of a partial ontological coverage, which derives from the manual mappings of Wikipedia infoboxes used to extract DBpedia resources to DBPO. The second, related, bias comes from the ratio between the average typed wikilinks owned by typed resources and the average typed wikilinks owned by untyped resources resources, that is 66% for the former and 37% for the latter.

The results seem to con rm the factor caused by those biases over the classi cation. In fact, while the results in classifying a test set of typed DBpedia resources { precisions 44.4% and 55.07% with EKPs and homotypes respectively { are relatively satisfactory (specially considering the amount of classes), we observed a dramatic fall of the precision in classifying untyped DBpedia resources, decreasing to 5.98% with EKPs and 13.93% with homotypes.

This means that, besides being still valid the results about the cognitive soundness of EKPs in providing an e ective and intuitive description of entities of a certain type (as reported in [ 13 ]), our hypothesis about the distinctive capacity of EKPs is weak. This seems due to wide overlaps among EKPs. The same overlaps emerge when applying homotypebased classi cation as well. Figure 7 shows the overlaps among the 4 largest classes of DBpedia, i.e., dbpo:Place, dbpo:Person, dbpo:Organisation and dbpo:Work.

Notice that the decrease in precision from the test set to the untyped resource set spans between -38% and -41% across the di erent approaches, probably revealing an approximate 40% DBpedia resources that are true negatives with reference to the existing DBPO, so providing a rough quanti cation of the missing ontological coverage of the current DBPO.

For instance, a quick exploration of untyped resources immediately evidences the need for types representing important notions such as Plan, Agreement, Scienti cDiscipline, Collection, Concept, etc.. It's quite interesting to remark that these notions are on one hand harder to generalize from infoboxes, but are also relatively established in existing foundational ontologies and ontology design patterns [ 8 ]. An area of improvement for DBPO and DBpedia typing is therefore an extension of the ontology and the ways to use the extension to type untyped resources.

A second area of improvement might be related to the imprecision deriving from the massive overlap among the abovementioned four core classes of DBPO.

However, we might consider a more critical attitude. While we deem important the role of ontologies in accurately distinguishing semantic types of things, specially in domains and tasks requiring ne-grained types, it might be interesting to explore an alternative vision of systematically ambiguous ontologies, where some types tend to merge because of their mutual dependence in the real world. For example, an organization is typically dependent on persons, places, and works, which on their turn can be often dependent on it: social organization is a major example.

Shifting our perspective, the distinctivity weakness of both EKPs and homotypes in classifying DBpedia resources can nd some basis in recent work [ 15 ], which con rms known semiotic assumptions about the centrality of (systematic) ambiguity in language, and poses signi cant challenges to ontological theories assuming semantic \pedantry". An interesting research topic in ontology design may investigate the consequences of assigning high value to \clusterability" of (systematically dependent) meaning dimensions rather than to their distinctivity.

There are many directions that these results open up. Some of them include the update of EKPs to DBpedia 3.7 and a further analysis in order to discover more distinctive features in their extraction. Another direction involves a mixed approach based both on inference and crowdsourcing through, for instance, exploratory search methods on Linked Data following the direction we took with Aemoo [ 12 ]. Other useful directions include use of indexing techniques, deep parsing of natural language, or social network analyses.

References

[1] C. Bizer , T. Heath , and T. Berners-Lee . Linked Data-The story so far . International Journal on Semantic Web and Information Systems , 4 ( 2 ):1{ 22 , 2009 .

[2] S. Bloehdorn and Y. Sure . Kernel Methods for Mining Instance Data in Ontologies . In K. Aberer, K.-S. Choi, N. Noy , D. Allemang , K.-I. Lee , L. J. B. Nixon , J. Golbeck , P. Mika , D. Maynard , G. Schreiber, and P. Cudre-Mauroux, editors, Proceedings of the 6th International Semantic Web Conference and the 2nd Asian Semantic Web Conference (ISWC 2007 + ASWC 2007) , volume 4825 of Lecture Notes in Computer Science, pages 58 { 71 , Busan , Korea, November 2007 . Springer Verlag.

[3] N. Christianini and J. Shawe-Taylor . Support Vector Machines and Other kernel-based Learning Methods . Cambridge University Press, 2000 .

[4] C. d'Amato , N. Fanizzi , and F. Esposito . Query Answering and Ontology Population: an Inductive Approach . In M. Hauswirth, M. Koubarakis , and S. Bechhofer, editors, Proceedings of the 5th European Semantic Web Conference (ESWC 2008 ), volume 5021 of Lecture Notes in Computer Science, Tenerife, Spain, June 2008. Springer Verlag.

[5] C. d'Amato , N. Fanizzi , and F. Esposito . Inductive Learning for the Semantic Web: What does it buy? Semantic Web , 1 ( 1 ): 53 { 59 , 2010 .

[6] N. Fanizzi , C. d'Amato, and F. Esposito . Statistical Learning for Inductive Query Answering on OWL Ontologies . In A. P. Sheth, S. Staab , M. Dean , M. Paolucci , D. Maynard , T. W. Finin , and K. Thirunarayan, editors, Proceedings of the 7th International Semantic Web Conference (ISWC 2008 ), volume 5318 of Lecture Notes in Computer Science, pages 195 { 212 , Karlsruhe , Germany, October 2008 . Springer.

[7] H. Frankfurt . Peirce's notion of abduction . The Journal of Philosophy , 55 ( 14 ): 593 { 597 , 1958 .

[8] A. Gangemi and V. Presutti . Handbook on Ontologies, chapter Ontology Design Patterns. Springer, 2nd edition , 2009 .

[9] J. Lehmann , C. Bizer , G. Kobilarov, S. Auer , C. Becker , R. Cyganiak , and S. Hellmann. DBpedia - A Crystallization Point for the Web of Data . Journal of Web Semantics , 7 ( 3 ): 154 { 165 , 2009 .

[10] M. Miller and M. F. RDF Primer. W3c recommendation, W3C , feb 2004 . http://www.w3.org/TR/2004/REC-rdf-primer20040210/.

[11] B. Motik , B. Parsia , and P. F. Patel-Schneider . OWL 2 web ontology language structural speci cation and functional-style syntax . W3C recommendation, W3C , Oct. 2009 . http://www.w3.org/TR/2009/REC-owl2 - syntax-20091027/.

[12] A. Musetti , A. G. Nuzzolese , F. Draicchio , V. Presutti , E. Blomqvist , A. Gangemi , and P. Ciancarini . Aemoo: Exploratory search based on knowledge patterns over the semantic web . Semantic Web Challenge.

[13] A. G. Nuzzolese , A. Gangemi , V. Presutti , and P. Ciancarini . Encyclopedic Knowledge Patterns from Wikipedia Links . In L. Aroyo, N. Noy , and C. Welty, editors, Proceedings of the 10th International Semantic Web Conference (ISWC2011) , pages 520 { 536 . Springer, 2011 .

[14] C. Peirce . Pragmatism as a principle and method of right thinking: The 1903 Harvard lectures on pragmatism . State Univ of New York Pr , 1997 .

[15] S. T. Piantadosi , H. Tily , and E. Gibson. The communicative function of ambiguity in language . Cognition , 122 ( 3 ): 280 { 291 , 2012 .

[16] F. Suchanek , G. Kasneci, and G. Weikum. Yago - A Large Ontology from Wikipedia and WordNet . Elsevier Journal of Web Semantics , 6 ( 3 ): 203 { 217 , 2008 .