The goal of our research is to predict a relation (predicate) of two given RDF entities (subject and object). Link prediction between entities is important for developing large-scale ontologies and for knowledge graph completion. TransE and TransR have been proposed as the methods for such a prediction. However, TransE and TransR embed both entities and relations in the same (or different) semantic space(s). Since entity embedding is enough for predicting relations, we propose a method for predicting a predicate from a subject and an object by using a Deep Neural Network (DNN), and developed RDFDNN. RDFDNN embeds entities only; given subject and object are embedded and concatenated to predict probability distribution of predicates. Experimental results showed that predictions by RDFDNN are more accurate than those by TransE and TransR. Although RDFDNN learns from RDF triples only, its accuracy is comparable to that of DKRL which uses both RDF triples and entity descriptions for learning.
Ontology learning is one of the important topics for developing the Semantic
Web. In general, there are many entity pairs where the relations between them
are unknown [
The goal of our research is to predict relations between two given entities in Resource Description Framework (RDF) accurately. RDF is the framework for representing Web resources, and each triple in RDF is composed of three entities (subject, predicate, and object). Subject and object are entities, and predicate is the relation between the entities. Suppose (Tokyo, is-capital-of, Japan) is an example of such a triple. We would like to predict \is-capitalof" when \Tokyo" and \Japan" are given. For this purpose, we propose a method for predicting a predicate from a subject and an object by using a Deep Neural Network (DNN), and developed RDFDNN. The code is available at https://github.com/yo0826jp/RDFDNN.
RDFDNN predicts the relation between two entities represented as a RDF triple. When h and t of a RDF triple (h; l; t) are given as inputs, RDFDNN will output l. Relational prediction is similar to classi cation, and DNN is good at classi cation. Therefore we predict the relation between two entities by DNN. entity voc and relation voc are the numbers of entities and relations, respectively. entity dim and relation dim are the dimensions of weight matrices. Embedding is the transformation from RDF entities to their vector representations. The length of the transformed vector is called embedding dimension. The vector representation of entities are learned by RDFDNN. We employ simple concatenation of two embedding vectors for the concat. It is better than element wise multiplication and element wise addition in RDFDNN.
Since the output of RDFDNN is the probability distribution of one-hot representation of relation l, the following cross entropy is used as the objective function for training RDFDNN:
E = ∑
∑ (h;t;l)2S k2relation voc lklogP (h; t)k; (1) where S is the set of triples in training data, P (h; t) the output of RDFDNN when h and t are given as its inputs, k is the integer index that satis es 0 k < relation voc. As the optimizer of the above objective function, Adam is used.
Since entity embedding is enough for predicting relations, RDFDNN focuses on entity embedding. Although RDFDNN cannot predict an object from a predicate and a subject, it has abilities of predicting a predicate accurately from a subject and an object.
In our experiments, we have used the FB15k and WN18. The datasets are the
same as the ones used in the experiments of previous research [
1.0 0.9 0.8 k 0.7 @ s it0.6 H 0.5 0.4 RDFDNN TransR TransE DKRL 0.31 2 3 4 5 6 7 8 9 10 k 1.00 0.95 0.90 0.85 k @0.80 its0.75 H0.70 00..6650 TTRrrDaaFnnDssERNN 0.551 2 3 4 5 6 7 8 9 10 k 3.2
For this failure analysis, 100 triples of RDFDNN failures are randomly sampled. Then the triples are manually evaluated and classi ed into the above four categories.
The most frequent failure is deceived by majority cases. As an example, RDFDNN's prediction of relation between \Leslie Dilley" and \Raiders of the Lost Ark" is \performer", while its correct answer is \art director". This is because the relation \performer" is the most frequent one for the relation between people and movies. The second most frequent failure is complete failure. The third most frequent failure is too abstract or too concrete compared with correct answers. As an example of this type, RDFDNN predicts the relation between \Park Chu-yong" and \South Korea" as \citizenship", while its correct answer is \Olympic representative". The least frequent failure is structurally similar, but this failure means that RDFDNN recognizes structural similarity between relations. As an example of this type, RDFDNN predicts the relation between \Washington Wizards" and \Michael Jordan" as \belonging states", while its correct answer is \team member". From the above failure analysis, we can say that even when RDFDNN failed, more than half of its failed prediction are valid in some sense. 4
In this paper, we propose RDFDNN for predicting relations of RDF from two given entities. RDFDNN is more accurate compared with TransE and TransR. RDFDNN is comparable with DKRL which uses both RDF triples and entity descriptions for learning.
This work was supported by Tokyo Tech - Fuji Xerox Cooperative Research (Project Code KY260195), JSPS Grant-in-Aid for Scienti c Research(B) (Grant Number 17H01785) and JST CREST (Grant Number JPMJCR1687).