1 Introduction
Ontology graphs are a special category of knowledge graphs that support and augment the Semantic Web with comprehensive and transportable machine understanding [23]. They store formal descriptions and specification of human knowledge in forms of relation facts (triples), making it semantically understandable and inferrable for the machine. Unlike other instancelevel knowledge graphs [41] that define simple and casual labeled relations for specified entities, ontology graphs define a fixed set of specialized semantic relations among generalized concepts. Such semantic relations of ontologies are typically very comprehensive in terms of relational properties and form hierarchies, which we are going to discuss shortly.
Populating large ontologies has been a critical challenge to the Semantic Web. In the past decade, several wellknown ontology graphs have been created and widely utilized, including Yago [24], ConceptNet [35], and DBpedia OWL [20]. Although some of these graphs contain millions of relation facts, they still face the coverage and completeness issues that have been the subject of much research [31, 26]. This is because enriching such large structures of expertise knowledge requires levels of intelligence and labor that is hardly affordable to humans. Hence, some works have proposed to mine ontologies from text using parsingbased [13, 25, 14] or fuzzylogicbased [31, 19, 39] techniques. However, in practice, these techniques are often limited by the lack of highquality reference corpora that are required for the harvest of the dedicated domain knowledge. Also, the precise recognition of relation facts for the ontology is another unsolved problem, since these relation facts are very highlevel and are often not explicitly expressed in the corpora [19]. Hence, these methods merely help populate some small ontology graphs in narrow domains such as gene ontologies and scholarly ontologies [11, 31], but they have not been successfully used to improve the completeness of these large crossdomain ontology graphs such as Yago and ConceptNet.
A more practical solution is to use translationbased graph embedding methods, which predict the missing relation facts using vector representations of the graph, without the need of additional information from any text corpus. Specifically, given a triple
such that , denote the source and the target entities (or concepts), and denotes the edge that marks the relation between and , then and are represented as two dimensional vectors and , respectively. An energy function is used to measure the plausibility of the triple, which also implies the transformation that characterizes . Therefore, new triples with high plausibility (or low energy) are often induced. For example, TransE [3] uses the energy function ^{1}^{1}1Hereafter, means or norm unless specified., where is characterized as a translation vector learnt from the latent connectivity patterns in the graph. Other representative works, such as TransH [38], TransR [21], and TransD [17] improve TransE by specializing the encoding process for each relation type using a relationspecific projection on entities.While these methods help enrich instancelevel knowledge graphs, they only focus on capturing the simple relations in instancelevel knowledge graphs, paying less attention to the comprehensive relations in ontology graphs. In fact, relation facts in ontology graphs are often defined with relational properties, such as transitivity and symmetry, as well as form hierarchies. A typical example is provided by IsA, which is both transitive and hierarchical, and is the most frequently appearing semantic relation in ontologies. We find that, in wellknown ontology graphs, comprehensive relations usually comprise the majority: 85% of the triples in Yago, 96% of the triples in ConceptNet, and 47% of the triples in DBpedia OWL enforce relational properties, while 60%, 38%, and 48% of these triples are defined with hierarchical relations. However, existing methods fail to represent these comprehensive relations for several reasons: (i) These methods at most use the same relationspecific projection in the energy function, but fail to differentiate the components of triples. Therefore, they are illposed to characterize triples with relational properties. In fact, the encoding of a concept that serves as different components in such triples, i.e. either or , must be differentiated so as to correctly preserve relational properties in the embedding spaces (as shown in Section 3.2). (ii) These methods also lack a learning phase that is dedicated to hierarchical relations. This also impairs the preciseness of embeddings. We observe in our experiments that, above limitations largely hinder the effectiveness of existing methods for ontology graphs.
Therefore, to support ontology population more effectively, we propose On2Vec, a translationbased graph embedding model that specializes in characterizing the comprehensive semantic relations in ontology graphs. On2Vec adopts two component models: the Componentspecific Model which preserves the relational properties by applying componentspecific projections on source and target concepts respectively, and the Hierarchy Model which performs an attentive learning process on hierarchical relations. We evaluate our model with the tasks of relation prediction and relation verification, which respond respectively to the following two questions: (i) What relation should be added between two concepts? (ii) Is the predicted relation correct? Experimental results on data sets extracted from Yago, ConceptNet, and DBpedia OWL show promising results and significant improvement on related methods.
The rest of the paper is organized as follows. We first discuss the related work, and then introduce our approach in the section that follows. After that we present the experimental evaluation, and conclude the paper in the last section.
2 Related Work
In this section, we discuss three lines of works that are related to our topic.
Ontology Population. Extensive human efforts are often put into the creation of ontology graphs. Thus, ontology population aims at automatically extending the graphs with the missing relation facts. A traditional strategy is to mine those facts from text corpora. Many works rely on parsingbased techniques to harvest relation facts [13, 25, 36, 15, 14]. These approaches often construct hundreds or thousands of rules or parsetrees that are not reusable, and human involvement is indispensable to filter the frequently generated conflict candidates. Other works depend on fuzzy logic [31, 19, 39] to generate relation facts with uncertainty, which is more tractable than parsingbased techniques and do not generate conflict candidates. However, identifying or summarizing the concepts from text still requires human intelligence. Moreover, methods mentioned above suffer from the lack of reference corpora that are closely related to and highly cover the knowledge of the ontology. Moreover, to associate the right contexts of the corpora with corresponding relation facts creates another major challenge, as semantic relations in ontologies are often specialized and are not explicitly expressed in the text. Due to these issues, we have seen few successful applications of these traditional approaches in improving the coverage of large crossdomain ontology graphs like Yago and ConceptNet. These issues motivate us to consider the more flexible “textfree” methods based on translationbased graph embeddings.
Translationbased Graph Embedding Methods. Recently, significant advancements have been made in learning translationbased embeddings for knowledge graphs. To characterize a triple , models of this family follow the common assumption that , where and are either the original vectors of and , or the transformed vectors and under a certain transformation w.r.t. relation . The forerunner TransE [3] sets and as the original and . Later works improve TransE by introducing relationspecific projections on entities to obtain different and
, including projections on relationspecific hyperplanes in TransH
[38], linear transformations to multiple relation spaces in TransR
[21], dynamic matrices in TransD [17], and other forms [18, 28]. These variants of TransE specialize the encoding process for different relations, therefore they often achieve better representations of triples than TransE. Meanwhile translationbased models cooperate well with other models. For example, variants of TransE are trained in joint with word embeddings to enable synthesized word embeddings with relational inferences [37, 43], and are combined with alignment models to help crosslingual knowledge alignment [8, 10]. However, existing translationbased models are not able to preserve triples with relational properties in the embedding spaces, because they do not differentiate the encoding of concepts that serve as different components in these triples. They also fail to provide a proper learning process for hierarchical relations. These are the major limitations we want to overcome.On the other hand, to enrich the knowledge in graphs, translationbased models proceed with entity prediction that predicts missing entities for triples. Since the candidate space of entities is extremely large, all these works seek to rank a set of candidates rather than acquiring the exact answers [38, 21, 17, 18, 28]. We instead proceed with relation prediction, which practically obtains the exact answers, as the relations in ontology graphs are not very diverse.
Other Knowledge Graph Embedding Methods. There are nontranslationbased methods that learn graph embeddings. UM [4], SME [1] are simplified versions of TransE and TransR; LFM [16] learns bilinear transformations among entities; TADW [40] learns contextbased embeddings from randomwalk generated contexts of the graphs (which is very similar to the recently introduced Rdf2Vec [32]). These methods do not explicitly embed relations, thus do not apply to our tasks. Others include neuralbased models SLM [12], and NTN [34] that were outperformed substantially by TransE and other translationbased methods on the tasks for populating instancelevel knowledge graphs [3, 38, 21, 17]. There are some which perform comparably with translationbased methods, but at the cost of much higher parameter complexity, such as RESCAL [30], and HolE [29]. We choose to compare with these two popular methods as well.
3 Embedding Ontology Graphs
In this section, we introduce the proposed method for learning ontology graph embeddings. We begin with the formalization of ontology graphs.
3.1 Preliminary
An ontology is a graph where is the set of concepts, and is the set of semantic relations. denotes a triple that represents a relation fact, for which and . Boldfaced , , respectively represent the embedding vectors of source , relation , and target
. Relations are further classified by
, which respectively denote the sets of transitive, symmetric, hierarchical, and other simple relations. We do not specify reflexive relations here because such relations can be easily model as a zero vector by any translationbased model. and thereof, are not required to be disjoint, while is disjoint with all the rest three. For transitive relations, that is to say, given , and three different concepts , if , then . As for symmetric relations, that is to say, given , and two different concepts , if , then . As for hierarchical relations, we further divide them into where denotes refinement relations that partition coarser concepts into finer ones, and denotes coercion relations that group finer concepts to coarser ones [5, 6, 7].3.2 Modeling
On2Vec adopts two component models that learn on the two facets of the ontology graph: the Componentspecific Model (CSM) which encodes concepts and relations into lowdimensional embedding spaces without the loss of the relational properties, and the Hierarchy Model (HM) which strengthens the learning process on hierarchical relations with an auxiliary energy.
3.2.1 Componentspecific Model.
The reason that previous translationbased models fail to preserve relational properties is because the relationspecific projection place concepts involved in transitive or symmetric relations at conflict positions. Fig. 1 depicts such conflicts, and a brief proof is given below:

[noitemsep]

Case 1. Consider and such that , , , where , , and are projected to , , and respectively by . Then if and hold for the first and second triples, it is impossible for to hold for the third triple, since (otherwise does not provide a valid vector translation).

Case 2. Consider and such that , where and are projected to and respectively by . Then it is not possible for both and to hold, since .
Hence, to solve the conflicts in the above two cases, CSM provides two componentspecific (and also relationspecific) projections to differentiate the encoding of the same concept that serves as different components in triples. The general form of the energy function is given as below,
where and are respectively the componentspecific projections for the source and the target concepts. It is easy to show that the componentspecific projections are able to solve the conflicts in learning the relational properties, as in Case 1 is projected differently when it serves as the source of or the target of , while both and in Case 2 can be learnt to be embedded in opposite positions respectively for and by the two projections. Corresponding conclusion can be easily extended to cases with more than three relation facts via mathematical induction.
Besides measuring the plausibility (or the opposite: dissimilarity) of a given triple, is also the basis for predicting missing relation facts for an ontology. Given two concepts and , we find the which leads to the lowest . The forms of and are decided particularly by the techniques to differentiate the concept encoding under different contexts of relations. In this paper, we adopt the relationspecific linear transformations [21]. Hence, we have and , such that . Other techniques like hyperplane projections, dynamic matrices, and bilinear transformations may also be considered, which we leave as future work.
The objective of CSM is to minimize the total energy of all triples. To achieve more efficient learning, we import negative sampling to the learning process, which is widely applied in previous works [3, 38, 21, 17]. Unlike these works that select negative samples on entities (or concepts), we perform negative sampling on semantic relations to better suit our tasks. Then the complete energy function of CSM is defined as the following hinge loss,
for which is a randomly sampled relation that does not hold between and , is a positive margin, and denotes the positive part of (i.e., ).
3.2.2 Hierarchy Model.
For a hierarchical relation, we often have multiple finer concepts that apply this relation to a coarser one. In this case, we appreciate a good representation where all the embeddings of the finer concepts converge closely in a tight neighborhood, which corresponds to low dissimilarity of the embedded relation. However, it is very likely for the learning process to spread out the embeddings of the finer concepts. Because each of the finer concepts can participate in multiple relation facts, encoding of a concept in one relation fact can be easily interfered by that of many other relation facts. This no doubt indicates low plausibility measures of the triples, and imprecise vector translation for the corresponding relations. Therefore, HM is dedicated to converge closely the projected embeddings of every finer concepts for a hierarchical relation.
To facilitate the definition of the energy function, we first define a refine operator denoted as :

[noitemsep]

Given , , then fetches all the finer concepts that directly apply the refinement relation to the coarser .

Given , , then fetches all the finer concepts that directly apply the coercion relation to the coarser .
The energy function of HM is defined below,
where is a function that monotonically increases w.r.t. the angle or the distance of the two argument vectors. In practice, can be easily implemented as cosine distance.
Negative sampling is imported to rewrite as below,
such that and are negative samples of concepts, and are respectively the hinge loss for refinement and coercion relations defined as below, where is a positive margin.
Model  #Parameters  Complex. rel. predict. 
TransE  
TransH  
TransR  
TransD  
On2Vec 
Data Set  DB3.6k  CN30k  YG15k  YG60k 
#trip.  6,485  286,763  219,472  522,282 
pct. prop.  47.39%  96.89%  45.69%  85.58% 
pct. hier.  47.11%  59.96%  76.80%  59.96% 
#rel.  8  41  17  17 
#con.  3,625  29,564  14,887  56,910 
#train.  5,485  256,762  204,064  472,280 
#valid.  500  10,001  5,000  10,000 
#test.  500  20,000  10,400  40,000 
Table 2 gives the model complexity of On2Vec and some related models in terms of parameter sizes. We also give out the computational complexity of the relation prediction for a pair of concepts, which is the most frequent operation in our tasks. Although On2Vec unavoidably increases the parameter sizes due to additional projections, it keeps the computational complexity of relation prediction at the same magnitude as TransR, which is lower than TransD.
3.3 Learning Process
The objective of learning On2Vec is to minimize the combined energy of and . Meanwhile, norm constraints are enforced on embeddings and projections to prevent training from a trivial solution where vectors collapse to infinitely large [2, 8, 38]. Such constraints are conjuncted below.
In the learning process, these constraints are quantified as soft constraints:
Finally, learning On2Vec
is realized by using batch stochastic gradient descent (SGD)
[27] to minimize the joint energy function given as below,where and are two nonnegative hyperparameters, and is the set of model parameters that include embedding vectors and projection matrices. Empirically (as shown in [38, 21]), is assigned with a small value within (0, 1]. is adjusted in experiments to weigh between the two component models. Instead of directly updating , the learning process optimizes and in separated groups of batches, and the batches from both groups are used to optimize
. We initialize vectors by drawing from a uniform distribution on the unit spherical surface, and initialize matrices using random orthogonal initialization
[33]. The detailed optimization procedure is given in Algorithm 1.4 Experiments
In this section, we evaluate On2Vec on two tasks that answer two important questions for ontology population: (i) Relation prediction: what is the relation to be added between a given pair of concepts? (ii) Relation verification: is a candidate relation fact correct or not?
The baselines that we compare against include the representative translationbased embedding methods TransE, TransH, TransR, and TransD [3, 38, 21, 17], and neural methos RESCAL and HolE [30, 29]. Experimental results are reported on four data sets extracted from DBpedia, ConceptNet, and Yago, for which comprehensive relation types have been predefined. Statistics of the data sets are shown in Table 2. All the meta relations that assign URIs and system timestamps are removed during the preparation of the data sets. To simplify the experiments, transitive relations are limited to fourhops. Relation facts for extra hops are hence discarded. Since DBpedia provides both ontology and instancelevel graphs, we keep only the ontology view to obtain DB3.6k. CN30k and YG15k are extracted from English versions of ConceptNet and Yago respectively. These two graphs match the number of nodes with WN18 and FB15k respectively, which are two commonlyused instancelevel graphs in related works [3, 38, 21, 17, 4, 1, 40]. YG60k is a much larger data set that is about half of the entire Englishversion Yago after data cleaning. Each data set is randomly partitioned into training, validation, and test sets.
Data Sets  DB3.6k  CN30k  YG15k  YG60k  
Rel Type  prop.  hier.  overall  prop.  hier.  overall  prop.  hier.  overall  prop.  hier.  overall 
TransE  8.40  8.71  13.31  5.09  3.21  8.01  2.03  0.56  0.20  0.02  0.00  0.16 
TransH  47.55  47.83  50.80  13.8  7.29  13.66  65.53  61.57  66.27  62.92  43.79  59.78 
TransD  50.40  57.98  80.74  72.34  76.18  77.67  74.42  75.60  77.77  72.39  66.18  73.23 
TransR  68.14  71.72  78.32  79.32  84.37  80.56  79.74  79.56  79.81  77.40  71.19  78.22 
RESCAL  29.70  35.65  36.19  55.39  56.06  54.46  58.88  54.50  59.07  52.36  53.16  58.51 
HolE  82.76  81.68  89.63  79.21  80.99  77.71  76.78  75.20  79.13  73.69  74.47  78.10 
O2V w/ HM  86.46  89.65  93.35  88.99  96.05  89.21  88.88  89.36  88.75  89.09  88.71  88.74 
O2V w/o HM  86.85  86.06  90.69  85.58  95.07  86.01  85.87  83.98  84.29  80.57  75.96  81.47 
4.1 Relation Prediction
This task aims at extending an ontology graph by predicting the missing relations for given concept pairs.
Evaluation Protocol. We evaluate our approach by way of heldout evaluation [37, 22]. Each model is trained on the training set that represents the known ontology. Then, for each case in the test set, given the source and target concepts, the model predicts the relation that leads to the lowest dissimilarity score defined in Section 3.2.1. To evaluate with controlled variables, on each data set, we employ the same configuration for every models. On DB3.6k, we fix dimensionality , margin , learning rate , , and norm. CN30k and YG15k shares the configuration as , , , , and norm. Lastly, we use , , , , and norm. is configured for On2Vec. To test the effect of HM, we also provide two versions of On2Vec. One version (On2Vec w/ HM) is set with , which is empirically decided via the hyperparameter study in Section 4.3. The other version (On2Vec w/o HM) nullifies HM by setting . To enable batch sampling for HM, we implement the function for hierarchical relation facts using hash trees. The learning process is stopped once the accuracy on the validation set stops improving.
Results. The overall accuracy is reported per data set in Table 3. On each data set, we also aggregate respectively the accuracy on the test cases with relational properties, as well as the accuracy on those with hierarchical relations. We discover that, TransE, though has performed well on encoding instancelevel knowledge graphs [3], receives unsatisfactory results on predicting the comprehensive ontology relations. By learning each relation type on a different hyperplane, TransH notably solves the problem of TransE, but appears to fail on CN30k where the candidate space is larger than other graphs. TransR and TransD provide more robust characterization of relations than TransH, especially in TransR where relationspecific projections are implemented as linear transformations. However, the overall performance of both TransR and TransD is impaired by the two types of comprehensive relations. For neural models, HolE adapts better on the smaller DB3.6k data set, while it is at most comparable to TransR and TransD on larger ones, and RESCAL is less successful on all settings. As expected, On2Vec greatly outperforms the above baseline methods, regardless of whether HM is enabled or not. The On2Vec with HM thereof, outperforms the best runnerup baselines respectively in all settings by 3.72%10.52% of overall accuracy, 4.09%11.69% of accuracy on cases with relational properties, and 7.97%14.24% of accuracy on cases with hierarchical relations. We also discover that, when HM is enabled, it leverages the accuracy on hierarchical relations by up to 12.75%, and overall accuracy by up to 7.27%, and does not noticeably cause interference to the prediction for cases with relational properties. Though, the advantage of CSM alone (i.e. On2Vec w/o HM) is still significant over the baselines. Since the relation prediction accuracy of On2Vec is close to 90% on all four data sets, this indicates that On2Vec achieves a promising level of performance in populating ontology graphs, and it is effective on both small and large graphs.
We also perform precisionrecall analysis on the two Yago data sets on translationbased models. To do so, we calculate the dissimilarity scores (Equation 3.2.1) for the possible predictions of each test case, and select those that are not ranked behind the correct prediction. Then a threshold is initiated as the minimum dissimilarity score. The answer set is inserted with predictions for which the dissimilarity scores fall below the threshold, and the answer set grows along with the increasing of the threshold, until all correct predictions are inserted. Therefore, we obtain the precisionrecall curves in Fig. 2, for which the area under curve is reported as: (i) For YG15k, On2Vec w/ HM: 0.9138; On2Vec w/o HM: 0.8938; TransE: 0.0457; TransH: 0.4973; TransD: 0.8386; TransR: 0.8587. (ii) For YG60k, On2Vec w/ HM: 0.9005; On2Vec w/o HM: 0.8703; TransE: 0.0313; TransH: 0.6688; TransD: 0.7275; TransR: 0.8372. This further indicates that On2Vec achieves better performance than other baselines, and HM improves the performance of On2Vec with CSM alone.
4.2 Relation Verification
Data Sets  DB3.6k  CN30k  YG15k  YG60k  
Rel Type  prop.  hier.  overall  prop.  hier.  overall  prop.  hier.  overall  prop.  hier.  overall 
TransE  67.49  71.44  67.57  69.14  18.23  51.85  58.73  62.69  69.09  60.92  61.30  66.89 
TransH  72.88  82.06  69.71  93.40  86.16  94.17  69.24  72.96  89.20  66.47  71.62  88.81 
TransD  76.79  81.11  74.44  91.63  84.20  93.36  65.63  70.58  88.01  61.76  71.08  86.34 
TransR  77.11  86.82  73.76  85.83  52.01  74.73  71.80  72.73  88.63  71.92  71.09  87.77 
RESCAL  75.30  74.61  76.20  70.41  75.64  72.28  68.76  67.30  72.29  69.36  69.16  76.21 
HolE  82.89  79.23  85.90  90.31  91.43  91.18  78.31  77.10  86.88  71.22  70.80  87.67 
O2V w/ HM  95.46  94.97  95.57  97.19  95.54  98.04  80.33  78.39  93.29  74.92  73.36  91.79 
O2V w/o HM  91.94  91.15  91.74  97,99  93.73  96.51  81.01  74.30  91.12  73.72  72.93  90.97 
Relation verification aims at judging whether a relation marked between two concepts is correct or not. It produces a classifier that helps to verify the candidate relation facts.
Evaluation Protocol. Because this is a binary classification problem that needs positive and negative cases, we use a complete data set as the positive cases. Then, following the approach of [34], we corrupt the data set to create negative cases. In detail, a negative case is created by (i) randomly replacing the relation of a positive case with another relation, or (ii) randomly assign a relation to a pair of unrelated concepts. Options (i) and (ii) respectively contribute negative cases that are as many as 100% and 50% of positive cases. We perform a 10fold crossvalidation. Within each fold, embeddings and the classifier are trained on the training data, and the classifier is evaluated on the remaining validation data.
We use a thresholdbased classifier, which is similar to the one for triple alignment verification in [8]. This simple classifier adequately relies on how precisely each model preserves the structure of the ontology graph in the embedding space. In detail, for each case, we calculate its dissimilarity score (Section 3.2.1). The classifier then finds a threshold such that implies positive, otherwise negative. The value of is determined to maximize the accuracy on the training data of each fold.
We carry forward the corresponding configurations from the last experiment, in order to show the performance of each model under controlled variables.
CN30k 
Offer, Entails, Degree 
Offer, Entails, Decide 
State, IsA, Boundary 
National_Capital, IsA, Boundary 
Get_in_line, HasFirstSubevent, Pay 
Convert, SimilarTo, Transform 
Person, ReceivesAction, Hint 
Stock, Entails, Receive 
Evasion, HasContext, Physic 
YG60k 
Luisa_de_Guzmn, isMarriedTo, John_IV_of_Portugal 
Georgetown, isLocatedIn, South_Carolina 
Gmina_pomiechwek, isLocatedIn, Gmina_Konstancin 
rebro_Airport, isLocatedIn, Karlskoga 
Horgen, isLocatedIn, Blach_District 
Luxor_International_Airport, isConnectedTo, 
Daqing_Sartu_Airport 
Akron, isLocatedIn, Ohio 
Curtis_guild_Jr, hasGender, Male 
Aalbach, isLocatedIn, Europe 
Results. We aggregate the mean accuracy for the two categories of comprehensive relation facts as well as the overall accuracy for each setting. The results are shown in Table 4
, which has a maximum standard deviation of 0.005 in crossvalidation for each setting. Thus, the results are statistically sufficient to reflect the performance of classifiers. Both versions of
On2Vec again outperform the other models, especially on comprehensive relation facts. On all four data sets, On2Vec outperforms the best runnerup baselines by 2.98%9.67% of overall accuracy, 2.02%12.57% of accuracy for cases with relational properties, and 1.298.15% of accuracy on hierarchical relations. This indicates that On2Vec precisely encodes the ontology graph structures, and provides much accurate plausibility measurement to decide the correctness of unknown triples. We also discover that, On2Vec trained with HM has a drop of accuracy for up to 0.8% on cases with relational properties from CN30k and YG15k. This is likely due to that the auxiliary learning process for hierarchical relations causes minor interference to the characterization of relational properties, while HM leverages the accuracy on hierarchical relations of these two data sets by at least 1.81%, and the overall accuracy by 0.82%3.83%. This indicates that HM is helpful in relation verification.4.3 Case Study
Lastly, we provide some case studies on hyperparameter values, and some examples of relation prediction.
4.3.1 Hyperparameter study
We examine the hyperparameter , which is the tradeoff between CSM and HM. The result based on relation prediction on YG15k is shown in Fig. 3. As we can see, although enabling HM with even a small value of can noticeably leverage the performance of On2Vec, the influence of different values of is not very notable, and the accuracy does not always go up along with the higher . In practice, may be finetuned for marginal improvement, while can be empirically selected.
4.3.2 Examples of relation prediction
Relation prediction is also performed for the complete data set of CN30k and YG60k. To do so, we randomly select 20 million pairs of unlinked concepts from these two data sets, and rank all the predictions based on the dissimilarity score . Then topranked predictions are selected. Human evaluation is used in this procedure, since there is no ground truth for the relation facts that are not preexisting. Like previous works [22, 42], we aggregate , i.e. the precision on the 200 predictions with highest confidence, which results in 73% and 71% respectively. Some examples of topranked predictions are shown in Table 5.
5 Conclusion
This paper proposes a greatly improved translationbased graph embedding method that helps ontology population by way of relation prediction. The proposed On2Vec model can effectively address the learning issues on the two categories of comprehensive semantic relations in ontology graphs, and improves previous methods using two dedicated component models. Extensive experiments on four data sets show promising capability of On2Vec on predicting and verifying relation facts.
The results here are very encouraging, but we also point out opportunities for further work and improvements. In particular, we should explore the effects of other possible forms of componentspecific projections, such as dynamic mapping matrices and bilinear mappings. Encoding other information such as the domain and range information of concepts may also improve the precision of our tasks. More advanced applications may also be developed using On2Vec such as ontologyboosted question answering. Jointly training On2Vec with alignment models [9] is another meaningful direction since it provides a more generic embedding model that helps populating and aligning multilingual ontology graphs.
References
 [1] A. Bordes, X. Glorot, et al., Joint learning of words and meaning representations for opentext semantic parsing, in AISTATS, 2012.

[2]
A. Bordes, X. Glorot, et al., A semantic matching energy function
for learning with multirelational data
, Machine Learning, 94 (2014).
 [3] A. Bordes, N. Usunier, et al., Translating embeddings for modeling multirelational data, in NIPS, 2013.
 [4] A. Bordes, J. Weston, et al., Learning structured embeddings of knowledge bases, in AAAI, 2011.
 [5] E. Camossi, M. Bertolotto, et al., A multigranular objectoriented framework supporting spatiotemporal granularity conversions, IJGIS, 20 (2006), pp. 511–534.
 [6] M. Chen, S. Gao, et al., Converting spatiotemporal data among heterogeneous granularity systems, in FUZZIEEE, 2016.
 [7] M. Chen, S. Gao, et al., Converting spatiotemporal data among multiple granularity systems, in SAC, 2016.
 [8] M. Chen, Y. Tian, et al., Multilingual knowledge graph embeddings for crosslingual knowledge alignment, IJCAI, (2017).

[9]
M. Chen and C. Zaniolo,
Learning multifaceted knowledge graph embeddings for natural language processing
, IJCAI, (2017).  [10] M. Chen, T. Zhou, et al., Multigraph affinity embeddings for multilingual knowledge graphs, in AKBC, 2017.
 [11] J. Cheng, S. Sun, et al., Netaffx gene ontology mining tool: a visual approach for microarray data analysis, Bioinformatics, 20 (2004).

[12]
R. Collobert and J. Weston,
A unified architecture for natural language processing: Deep neural networks with multitask learning
, in ICML, 2008.  [13] A. Culotta and J. Sorensen, Dependency tree kernels for relation extraction, in ACL, 2004.
 [14] K. Fundel, R. Küffner, et al., RelEx–Relation extraction using dependency parse trees, Bioinformatics, 23 (2007).
 [15] C. Giuliano and A. Gliozzo, Instancebased ontology population exploiting namedentity substitution, in COLING, 2008.
 [16] R. Jenatton, N. L. Roux, et al., A latent factor model for highly multirelational data, in NIPS, 2012.
 [17] G. Ji, S. He, et al., Knowledge graph embedding via dynamic mapping matrix, in ACL, 2015.
 [18] Y. Jia, Y. Wang, et al., Locally adaptive translation for knowledge graph embedding, in AAAI, 2016.
 [19] R. Y. Lau, D. Song, et al., Toward a fuzzy domain ontology extraction method for adaptive elearning, TKDE, 21 (2009).
 [20] J. Lehmann, R. Isele, et al., Dbpedia–a largescale, multilingual knowledge base extracted from Wikipedia, Semantic Web, 6 (2015).
 [21] Y. Lin, Z. Liu, et al., Learning entity and relation embeddings for knowledge graph completion., in AAAI, 2015.
 [22] Y. Lin, S. Shen, et al., Neural relation extraction with selective attention over instances, in ACL, 2016.
 [23] A. Maedche and S. Staab, Learning ontologies for the semantic web, in ISWC, 2001.
 [24] F. Mahdisoltani, J. Biega, et al., Yago3: A knowledge base from multilingual Wikipedias, in CIDR, 2015.
 [25] H. Mousavi, M. Atzori, et al., Textmining, structured queries, and knowledge management on web document corpora, SIGMOD Record, 43 (2014).
 [26] H. Mousavi, S. Gao, , et al., Mining semantics structures from syntactic structures in web document corpora, International Journal of Semantic Computing, 8 (2014).
 [27] D. Needell, R. Ward, et al., Stochastic gradient descent, weighted sampling, and the randomized kaczmarz algorithm, in NIPS, 2014.
 [28] D. Q. Nguyen, K. Sirts, et al., Stranse: a novel embedding model of entities and relationships in knowledge bases, in NAACL HLT, 2016.
 [29] M. Nickel, L. Rosasco, et al., Holographic embeddings of knowledge graphs, in AAAI, 2016.
 [30] M. Nickel, V. Tresp, et al., A threeway model for collective learning on multirelational data, in ICML, 2011.
 [31] T. T. Quan, S. C. Hui, et al., Automatic generation of ontology for scholarly semantic web, in ISWC, 2004.
 [32] P. Ristoski and H. Paulheim, Rdf2vec: Rdf graph embeddings for data mining, in ISWC, 2016.
 [33] A. M. Saxe, J. L. McClelland, et al., Exact solutions to the nonlinear dynamics of learning in deep linear neural networks, ICLR, (2014).

[34]
R. Socher, D. Chen, et al.,
Reasoning with neural tensor networks for knowledge base completion
, in NIPS, 2013.  [35] R. Speer, J. Chin, et al., Conceptnet 5.5: An open multilingual graph of general knowledge., in AAAI, 2017.
 [36] T. Wang, Y. Li, K. Bontcheva, et al., Automatic extraction of hierarchical relations from text, in ESWC, 2006.
 [37] Z. Wang, J. Zhang, et al., Knowledge graph and text jointly embedding., in EMNLP, 2014.
 [38] Z. Wang, J. Zhang, et al., Knowledge graph embedding by translating on hyperplanes, in AAAI, 2014.
 [39] D. H. Widyantoro and J. Yen, A fuzzy ontologybased abstract search engine and its user studies, in FUZZIEEE, 2001.
 [40] C. Yang, Z. Liu, et al., Network representation learning with rich text information, in IJCAI, 2015.
 [41] C. Zaniolo, S. Gao, et al., Userfriendly temporal queries on historical knowledge bases, Information and Computation, (2017).

[42]
D. Zeng, K. Liu, et al.,
Distant supervision for relation extraction via piecewise convolutional neural networks.
, in EMNLP, 2015.  [43] H. Zhong, J. Zhang, et al., Aligning knowledge and text embeddings by entity descriptions, in EMNLP, 2015.
Comments
There are no comments yet.