Dataset publication on the Web has been greatly influenced by the Linked Open Data (LOD) project. Many interlinked datasets have become freely available on the Web creating a structured and distributed knowledge representation. Analysis and aligning of concepts and instances in these interconnected datasets have received a lot of attention in the recent past compared to properties. We identify three different categories of property pairs found in the alignment process and study their relative distribution among well known LOD datasets. We also provide comparative analysis of state-of-the-art techniques with regard to different categories, highlighting their capabilities. This could lead to more realistic and useful alignment of properties in LOD and similar datasets.
LOD [
We introduced a novel way of computing property alignment (similarity) between
interconnected datasets by exploring the available links between the datasets and using
statistical measures [
We analyze different types of property pairs found along with the experiments
performed in [
We can categorize the types of property pairs between linked datasets in two orthogonal ways: (1) on the basis of their semantics, and (2) on the basis of the techniques and tools required to determine the inter-relationships or alignment among property pairs. On the basis of semantics, the related property pairs can be classified as (1) equivalent properties or (2) those possessing a property-sub property relationship. On the basis of the techniques used to align properties, we can classify property pairs as follows: 1. Simple property pairs: These have high syntactic similarity in the property names and may have a common prefix, common suffix, adjectives, or different ordering of words, e.g., birthPlace vs placeOfBirth. Here the words “place” and “birth” are in a different order for the two properties. 2. Opaque property pairs: These have the same meaning but use different words. This can be further categorized into two parts. (a) Synonymous property pairs: Similarity of the two properties can be decided by analyzing the meaning of the property names and is intentional. This can be achieved by using an external dictionary or a lexical database like WordNet. If property name is a word phrase, similarity can be checked by removing common words from the property names, e.g., occupation vs profession, city of birth vs place of birth. In the second property pair, the common suffix can be eliminated from the comparison. (b) Complex property pairs: Similarity cannot be determined by considering property names alone, but requires additional information such as extension analysis, and domain and range. These are ambiguous or have multiple meanings but have a specific meaning in a dataset, e.g., battle vs participated in conflict, resting place vs place of burial. The two terms “conflict” and “resting place” have multiple meanings and are used in many contexts. Hence, the similarity is harder to identify.
In this analysis, we highlight the advantages of using property extensions compared
to string based and external dictionary based methods that focus on analyzing property
names in the matching process. We consider only object-type properties for this analysis
in DBpedia, Freebase, LinkedMDB, and DBLP datasets1, taking 5000 instances in each
sample set [
DB-FB DB-FB Film DB-FB Person Software
Sim Comp Syn
DB-LMDB
Film
DBLP Articles
DB-FB DB-FB Film DB-FB Person Software
Exact Non-Exact
DB-LMDB
Film
DBLP Articles (a) Property pair type breakdown (b) Exact matching of property pairs
The correct matches in this analysis were manually identified and categorized by the authors and verified by an external reviewer. Figure 1 shows the breakdown of properties into the three types that we are interested in. According to Figure 1(a), the majority of the property pairs belong to simple property pairs followed by complex and synonymous property pairs. Moreover, some property pairs can be matched using exact property name matching as shown in Figure 1(b), but they account for less. Based on the facts presented in Figure 1, on average, the majority of the matching property pairs are simple, but cannot be matched using exact matching of property names.
There are different approaches for aligning property pairs between datasets
including [
Considering these matchers, Figure 2 shows the percentages of the correctly identified property pairs for the three types of property pairs. It also shows the superiority of the extension based approach over string based and dictionary (WordNet) based approaches. It is clear from Figures 2(a), 2(b), 2(c), and 2(d) that the extension based 2 namely, LCH, RES, HSO, JCN, LESK, PATH, WUP and LIN 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0%
DB-FB DB-FB Film DB-FB Person Software
Sim Comp Syn
DB-LMDB
Film
DBLP Articles
DB-FB DB-FB Film DB-FB Person Software Sim Sim & Exact Comp & Exact Syn
DB-LMDB
Film Comp Syn & Exact
DBLP Articles (a) Extension based algorithm matching
(b) WordNet similarity based matching DB-FB DB-FB Film DB-FB Person Software
Sim Comp Syn
DB-LMDB
Film
No Results
DBLP Articles
DB-FB DB-FB Film DB-FB Person Software
Sim Comp Syn
DB-LMDB
Film
No Results
DBLP
Articles
(c) String similarity based - Jaro Winkler
(d) String similarity based - Dice
approach performed better and achieved the highest results in matching all three types
of property pairs. We added exact matching of property names capability to WordNet
based algorithm and improved its performance as shown in Figure 2(b). This is
because some word phrases cannot be matched (searched) using WordNet but they have
the same or common word phrases in their names. It is also interesting to note that
the WordNet based approach failed to identify any of the synonymous property pairs
in most of the experiments as shown in Figure 2(b). This kind of behavior is expected
for string similarity or syntax based approaches, but not for a lexical database based
approach like WordNet, which is specialized in synonym word categorization. Figures
2(c) and 2(d) present matching performances when the similarity of property names are
considered using string matching algorithms. It is shown that string similarity based
matching missed all synonymous and complex property pairs leaving them unsuitable
for matching property pairs in general. Based on the facts (recall values) represented
in Figure 2, extension based property alignment has the capability to identify many
property pairs including complex and hidden property pairs compared to others.
Furthermore, Table 1 outlines both precision and recall for each matcher for all property
pair types, which also sheds lights on false positives (see [
Extension Based Algorithm Dice Similarity
We provided a breakdown of types of property pairs that can be found on linked datasets in the alignment process. Even though the majority of the property pairs are simple, many cannot be identified using string manipulation techniques. In our sample datasets, 63%, 29%, and 8% of all property pairs are simple, complex, and synonymous, respectively. We have shown that in every category, extension based property pair alignment showed better results. For example, the extension based approach showed an average improvement in the range of 5% - 32% compared to simple syntactic and WordNet based approaches. Hence, we conclude that the extension (or instance) based approach can discover many property pairs that are semantically the same, which cannot be uncovered by purely syntactic means.
Acknowledgement - This work was supported by the NSF award 1143717 “III: EAGER Expressive Scalable Querying over Linked Open Data”. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.