1 Introduction
Recent years have witnessed a rapid growth of knowledge bases (KBs) such as Freebase^{1}^{1}1http://freebase.com, DBPedia (Auer et al., 2007), and YAGO (Suchanek et al., 2007). These KBs store facts about realworld entities (e.g. people, places, and things) in the form of RDF triples^{2}^{2}2http://www.w3.org/TR/rdf11concepts/ (i.e. (subject, predicate, object)). Today’s KBs are large in size. For instance, Freebase contains millions of entities and billions of facts (triples) involving a large variety of predicates (relation types). Such largescale multirelational data provide an excellent potential for improving a wide range of tasks, from information retrieval, question answering to biological data mining.
Recently, much effort has been invested in relational learning methods that can scale to large knowledge bases. Tensor factorization (e.g.
(Nickel et al., 2011, 2012)) and neuralembeddingbased models (e.g. (Bordes et al., 2013a, b; Socher et al., 2013)) are two popular kinds of approaches that learn to encode relational information using lowdimensional representations of entities and relations. These representation learning methods have shown good scalability and reasoning ability in terms of validating unseen facts given the existing KB.In this work, we focus on the study of neuralembedding models, where the representations are learned using neural networks with energybased objectives. Recent embedding models TransE (Bordes et al., 2013b) and NTN (Socher et al., 2013) have shown stateoftheart prediction performance compared to tensor factorization methods such as RESCAL (Nickel et al., 2012). They are similar in model forms with slight differences on the choices of entity and relation representations. Without careful comparison, it is not clear how different design choices affect the learning results. In addition, the performance of the embedding models are evaluated on the link prediction task (i.e. predicting the correctness of unseen triples). This only indirectly shows the meaningfulness of lowdimensional embeddings. It is hard to explain what relational properties are being captured and to what extent they are captured during the embedding process.
We make three main contributions in this paper. (1) We present a general framework for multirelational learning that unifies most multirelational embedding models developed in the past, including NTN (Socher et al., 2013) and TransE (Bordes et al., 2013b). (2) We empirically evaluate different choices of entity representations and relation representations under this framework on the canonical link prediction task and show that a simple bilinear formulation achieves new stateoftheart results for the task (a top10 accuracy of 73.2% vs. 54.7% by TransE when evaluated on Freebase). (3) We propose and evaluate a novel approach that utilizes the learned embeddings to mine logical rules such as . We show that such rules can be effectively extracted by modeling the composition of relation embeddings, and that the embeddings learned from the bilinear objective are particularly good at capturing the compositional semantics of relations via matrix multiplication. Furthermore, we demonstrate that our embeddingbased approach outperforms a stateoftheart rule mining system AMIE (Galárraga et al., 2013) on mining rules that involve compositional reasoning.
The rest of this paper is structured as follows. Section 2 discusses related work. Section 3 presents the general framework for learning multirelational representations. Sections 4 and 5 present two inference tasks: a canonical link prediction task and a novel rule extraction task where the learned embeddings are empirically evaluated. Section 6 concludes the paper.
2 Related Work
Multirelational learning has been an active research area for the past couple of years. Traditional statistical learning approaches (Getoor & Taskar, 2007) such as Markovlogic networks (Richardson & Domingos, 2006) usually suffer from scalability issues. More recently, various types of representation learning methods have been proposed to embed multirelational knowledge into lowdimensional representations of entities and relations, including tensor/matrix factorization (Singh & Gordon, 2008; Nickel et al., 2011, 2012), Bayesian clustering framework (Kemp et al., 2006; Sutskever et al., 2009), and neural networks (Paccanaro & Hinton, 2001; Bordes et al., 2013a, b; Socher et al., 2013). Our work focuses on the study of neuralembedding models as they have shown good scalability and strong generalizability on largescale KBs.
Existing neural embedding models (Bordes et al., 2013a, b; Socher et al., 2013) all represent entities as lowdimensional vectors and represent relations as operators that combine the representations of two entities. They differ in different parametrization of relation operators. For instance, given two entity vectors, the model of Neural Tensor Network (NTN) (Socher et al., 2013) represents each relation as a bilinear tensor operator followed by a linear matrix operator. The model of TransE (Bordes et al., 2013b), on the other hand, represents each relation as a single vector that linearly interacts with the entity vectors. Likewise, variations on entity representations also exist. Most methods represent each entity as a unit vector while NTN (Socher et al., 2013) represent entities as an average of word vectors and initializing word vectors with pretrained vectors from external text corpora. There has not been work that closely examines the effectiveness of these different design choices.
Our work on embeddingbased rule extraction presented in part of this paper is related to the earlier study on logical inference with learned continuousspace representations. Much existing work along this line focuses on learning logicbased representations for natural language sentences. For example, Socher et al. (2012) builds a neural network that recursively combines word representations based on parse tree structures and shows that such neural network can simulate the behavior of conjunction and negation. Bowman (2014) further demonstrates that recursive neural network can capture certain aspects of natural logical reasoning on examples involving quantifiers like some and all. Recently, Grefenstette (2013) shows that in theory most aspects of predicate logic can be simulated using tensor calculus. Rocktäschel et al. (2014) further implements the idea by introducing a supervised objective that trains embeddings to be consistent with given logical rules. The evaluation was conducted on toy data and uses limited logical forms. Different from these earlier studies, we propose a novel approach to utilizing embeddings learned without explicit logical constraints to directly mine logical rules from KBs. We demonstrate that the learned embeddings of relations can capture the compositional semantics of relations. Moreover, we systematically evaluate our approach and compare it favorably with a stateoftheart rule mining approach on the rule extraction task on Freebase.
3 MultiRelational Representation Learning
In this section, we present a general neural network framework for multirelational representation learning. We discuss different design choices for the representations of entities and relations which will be empirically compared in Section 4.
Given a KB that is represented as a list of relation triplets (denoting (the subject) and (the object) that are in a certain relationship ), we want to learn representations for entities and relations such that valid triplets receive high scores (or low energies). The embeddings can be learned via a neural network. The first layer projects a pair of input entities to low dimensional vectors, and the second layer combines these two vectors to a scalar for comparison via a scoring function with relationspecific parameters.
3.1 Entity Representations
Each input entity corresponds to a highdimensional vector, either a “onehot” index vector or a “nhot” feature vector. Denote by and the input vectors for entity and , respectively. Denote by the first layer projection matrix. The learned entity representations, and can be written as
where can be a linear or nonlinear function, and is a parameter matrix, which can be randomly initialized or initialized using pretrained vectors.
Most existing embedding models adopt the “onehot” input vectors except for NTN (Socher et al., 2013) which represents each entity as an average of its word vectors. This can be viewed as adopting “bagofwords” vectors as input and learning a projection matrix consisting of word vectors.
3.2 Relation Representations
The choice of relation representations reflects in the form of the scoring function. Most of the existing scoring functions in the literature can be unified based on a basic linear transformation
, a bilinear transformation or their combination, where and are defined as(1) 
which and are relationspecific parameters.
Models  Scoring Function  

Distance (Bordes et al., 2011)    
Single Layer (Socher et al., 2013)    
TransE (Bordes et al., 2013b)  
NTN (Socher et al., 2013) 
In Table 1, we summarize several popular scoring functions in the literature for a relation triplet , reformulated in terms of the above two functions. Denote by two entity vectors. Denote by and matrix or vector parameters for linear transformation . Denote by tensor parameters for bilinear transformation .
is an identity matrix.
is an additional parameter for relation . The scoring function for TransE (L2 formulation) is derived from , where and are unit vectors.Note that NTN is the most expressive model as it contains both linear and bilinear relation operators as special cases. In terms of the number of parameters, TransE is the simplest model which only parametrizes the linear relation operators with onedimensional vectors.
In this paper, we also consider the basic bilinear scoring function:
(2) 
which is a special case of NTN without the nonlinear layer and the linear operator, and uses a 2d matrix operator instead of a tensor operator. Such bilinear formulation has been used in other matrix factorization models such as in (Nickel et al., 2011; Jenatton et al., 2012; GarcíaDurán et al., 2014) with different forms of regularization. Here, we consider a simple way to reduce the number of relation parameters by restricting to be a diagonal matrix. This results in the same number of relation parameters as TransE. Our experiments in Section 4 demonstrate that this simple formulation enjoys the same scalable property as TransE and it achieves superior performance over TransE and other more expressive models on the task of link prediction.
This general framework for relationship modeling also applies to the recent deepstructured semantic model (Huang et al., 2013; Shen et al., 2014a, b; Gao et al., 2014; Yih et al., 2014), which learns the relevance or a single relation between a pair of word sequences. The framework above applies when using multiple neural network layers to project entities and using a relationindependent scoring function . The cosine scoring function is a special case of with normalized and with .
3.3 Parameter Learning
The neural network parameters of all the models discussed above can be learned by minimizing a marginbased ranking objective , which encourages the scores of positive relationships (triplets) to be higher than the scores of any negative relationships (triplets). Usually only positive triplets are observed in the data. Given a set of positive triplets , we can construct a set of “negative” triplets by corrupting either one of the relation arguments, . Denote the scoring function for triplet as . The training objective is to minimize the marginbased ranking loss
(3) 
4 Inference Task I: Link Prediction
We first conduct a comparison study of different embedding models on the canonical link prediction task, which is to predict the correctness of unseen triplets. As in (Bordes et al., 2013b), we formulate link prediction as an entity ranking task. For each triplet in the test data, we treat each entity as the target entity to be predicted in turn. Scores are computed for the correct entity and all the corrupted entities in the dictionary and are ranked in descending order. We consider Mean Reciprocal Rank (MRR) (an average of the reciprocal rank of an answered entity over all test triplets), HITS@10 (top10 accuracy), and Mean Average Precision (MAP) (as used in (Chang et al., 2014)
) as the evaluation metrics.
We examine five embedding models in decreasing order of complexity: (1) NTN with tensor slices as in (Socher et al., 2013); (2) Bilinear+Linear, NTN with tensor slice and without the nonlinear layer; (3) TransE, a special case of Bilinear+Linear (see Table 1); (4) Bilinear: using scoring function in Eq. (2); (5) Bilineardiag: a special case of Bilinear where the relation matrix is a diagonal matrix.
Datasets
We used the WordNet (WN) and Freebase (FB15k) datasets introduced in (Bordes et al., 2013b). WN contains triplets with entities and relations, and FB15k consists of triplets with entities and relations. We use the same training/validation/test split as in (Bordes et al., 2013b). We also consider a subset of FB15k (FB15k401) containing only frequent relations (relations with at least training examples). This results in triplets with entities and relations.
Implementation details
All the models were implemented in C# and using GPU. Training was implemented using minibatch stochastic gradient descent with AdaGrad
(Duchi et al., 2011). At each gradient step, we sampled for each positive triplet two negative triplets, one with a corrupted subject entity and one with a corrupted object entity. The entity vectors are renormalized to have unit length after each gradient step (it is an effective technique that empirically improved all the models). For the relation parameters, we used standard L2 regularization. For all models, we set the number of minibatches to , the dimensionality of the entity vector , the regularization parameter, and the number of training epochs
on FB15k and FB15k401 and on WN ( was determined based on the learning curves where the performance of all models plateaued.) The learning rate was initially set to and then adapted during training by AdaGrad.4.1 Results
FB15k  FB15k401  WN  
MRR  HITS10  MRR  HITS10  MRR  HITS10  
NTN  0.25  41.4  0.24  40.5  0.53  66.1 
Blinear+Linear  0.30  49.0  0.30  49.4  0.87  91.6 
TransE (DistADD)  0.32  53.9  0.32  54.7  0.38  90.9 
Bilinear  0.31  51.9  0.32  52.2  0.89  92.8 
Bilineardiag (DistMult)  0.35  57.7  0.36  58.5  0.83  94.2 
Table 2 shows the results of all compared methods on all the datasets. In general, we observe that the performance increases as the complexity of the model decreases on FB. NTN, the most complex model, provides the worst performance on both FB and WN, which suggests overfitting. Compared to the previously published results of TransE (Bordes et al., 2013b), our implementation achieves much better results (53.9% vs. 47.1% on FB15k and 90.9% vs. 89.2% on WN) using the same evaluation metric (HITS@10). We attribute such discrepancy mainly to the different choice of SGD optimization: AdaGrad vs. constant learning rate. We also found that Bilinear consistently provides comparable or better performance than TransE, especially on WN. Note that WN contains much more entities than FB, it may require the parametrization of relations to be more expressive to better handle the richness of entities. Interestingly, we found that a simple variant of Bilinear – Bilineardiag, clearly outperforms all baselines on FB and achieves comparable performance to Bilinear on WN. Note that Bilineardiag has the limitation of encoding the difference between a relation and its inverse. Still, as there is a large variety of relations in FB and the average number of training examples seen by each relation is relatively small (compared to WN), the simple form of Bilineardiag is able to provide good prediction performance.
Multiplicative vs. Additive Interactions Note that Bilineardiag and TransE have the same number of model parameters and their difference can be viewed as the operational choices of the composition of two entity vectors – Bilineardiag uses weighted elementwise dot product (multiplicative operation) and TransE uses elementwise subtraction with a bias (additive operation). To highlight the difference, here we use DistMult and DistAdd to refer to Bilineardiag and TransE, respectively. Comparisons between these two models can provide us more insights on the effect of two common choices of compositional operations – multiplication and addition for modeling entity relations. Overall, we observed superior performance of DistMult on all the datasets in Table 2. Table 3 shows the HITS@10 score on four types of relation categories (as defined in (Bordes et al., 2013b)) on FB15k401 when predicting the subject entity and the object entity respectively. We can see that DistMult significantly outperforms DistAdd in almost all the categories.
Predicting subject entities  Predicting object entities  
1to1  1ton  nto1  nton  1to1  1ton  nto1  nton  
DistADD  70.0  76.7  21.1  53.9  68.7  17.4  83.2  57.5 
DistMult  75.5  85.1  42.9  55.2  73.7  46.7  81.0  58.8 
Initialization of Entity Vectors In the following, we examine the learning of entity representations and introduce two further improvements: using nonlinear projection and initializing entity vectors with pretrained vectors. We focus on DistMult as our baseline and compare it with the two modifications DistMulttanh (using for entity projection ) and DistMulttanhEVinit (initializing the entity parameters with the dimensional pretrained entity vectors released by word2vec (Mikolov et al., 2013)) on FB15k401. We also reimplemented the initialization technique introduced in (Socher et al., 2013) – each entity is represented as an average of its word vectors and the word vectors are initialized using the dimensional pretrained word vectors released by word2vec. We denote this method as DistMulttanhWVinit. Inspired by (Chang et al., 2014), we design a new evaluation setting where the predicted entities are automatically filtered according to “entity types” (entities that appear as the subjects/objects of a relation have the same type defined by that relation). This provides us with better understanding of the model performance when some entity type information is provided.
MRR  HITS10  MAP (w/ type checking)  

DistMult  0.36  58.5  64.5 
DistMulttanh  0.39  63.3  76.0 
DistMulttanhWVinit  0.28  52.5  65.5 
DistMulttanhEVinit  0.42  73.2  88.2 
In Table 4, we can see that DistMulttanhEVinit provides the best performance on all the metrics. Surprisingly, we observed performance drops by DistMulttanhWVinit. We suspect that this is because word vectors are not appropriate for modeling entities described by noncompositional phrases (more than 73% of the entities in FB15k401 are person names, locations, organizations and films). The promising performance of DistMulttanhEVinit suggests that the embedding model can greatly benefit from pretrained entitylevel vectors using external textual resources.
5 Inference Task II: Rule Extraction
In this section, we focus on a complementary inference task, where we utilize the learned embeddings to extract logical rules from the KB. For example, given the fact that a person was born in New York and New York is a city of the United States, then the person’s nationality is the United States:
Such logical rules can serve four important purposes. First, they can help deduce new facts and complete the existing KBs. Second, they can help optimize data storage by storing only rules instead of large amounts of extensional data, and generate facts only at inference time. Third, they can support complex reasoning. Lastly, they can provide explanations for inference results, e.g. we may infer that people’s professions usually involve the specialization of the field they study, etc.
The key problem of extracting Horn rules like the aforementioned example is how to effectively explore the search space. Traditional rule mining approaches directly operate on the KB graph – they search for possible rules (i.e. closedpaths in the graph) by pruning rules with low statistical significance and relevance (Schoenmackers et al., 2010). These approaches often fail on large KB graphs due to scalability issues. In the following, we introduce a novel embeddingbased rule mining approach whose efficiency is not affected by the size of the KB graph but rather by the number of distinct types of relations in the KB (which is usually relatively small). It can also mine better rules due to its strong generalizability.
5.1 Background and Notations
For a better illustration, we adopt the graph view of KB. Each binary relation is a directed edge from node to node and with link type . We are interested in extracting Horn rules that consist of a head relation and a sequence of body relations :
(4) 
where are variables that can be substituted by entities. We constrain the body relations to form a directed path in the graph and the head relation to from a directed edge that close the path (from the start of the path to the end of the path). We denote such property as the closedpath property. For consecutive relations that share one variable but do not form a path, e,g, , we can replace one of the relations with its inverse relation, so that the relations are connected by an object and an subject, e.g. . We are interested in mining rules that reflect relationships among different relation types, therefore we also constrain to have distinct relation types. A rule is instantiated when all variables are substituted by entities. We denote the length of the rule as the number of body relations. In general longer rules are harder to extract due to the exponential search space. In our experiments, we focus on extracting rules of length 2 and 3.
In KBs, entities usually have types and relations often can only take arguments of certain types. For example, BornInCity relation can only take a person as the subject and a location as the object. For each relation , we denote the domain of its subject argument (the set of entities that can appear in the subject position) as and similarly the domain of its object argument as . Such domain information can be extremely useful in restricting the search space of logical rules.
5.2 Embeddingbased Rule Extraction
For simplicity, we consider Horn rules of length 2 (longer rules can be easily derived from this case):
(5) 
Note that the body of the rule can be viewed as the composition of relations and , which is a new relation that has the property that entities and are in a relation if and only if there is an entity which simultaneously satisfies two relations and .
We model relation composition as multiplication or addition of two relation embeddings. Here we focus on relation embeddings that are in the form of vectors (as in TransE) and matrices (as in Bilinear and its variants). The composition results in a new embedding that lies in the same relation space. Specifically, we use addition for relation vector embeddings and multiplication for relation matrix embeddings. This is inspired by two different properties: (1) if a relation corresponds to a translation vector and assume when holds, then we have the property that and implies ; (2) if a relation corresponds to a matrix in the bilinear transformation and assume when holds, also and are unit vectors and is still a unit vector ^{3}^{3}3These assumptions may not hold in our implementations. The intuition still leads to surprisingly good empirical performance on Horn rule extraction. How to effectively enforce these constraints is worth investigating in future work., then we have the property that and implies .
To simulate the implication in 5, we want the composition result of relation and to demonstrate similar behavior to the embedding of relation . We assume the similarity between relation embeddings is related to the Euclidean distance if the embeddings are vectors and to the Frobenius norm if the embeddings are matrices. This distance metric allows us to rank possible pairs of relations with respect to the relevance of their composition to the target relation.
Note that we do not need to enumerate all possible pairs of relations in the KB. For example, if the relation in the head is , then we are only interested in relation pairs that satisfy the type constraints, namely: (1) ; (2) ; (3) . As mentioned before, the arguments (entities) of relations are usually strongly typed in KBs. Applying such domain constraints can effectively reduce the search space.
In Algorithm 1, we describe our rule extraction algorithm for general closedpath Horn rules as in Eq. (4). In Step 7, denotes vector addition or matrix multiplication. We apply a global threshold value in our experiments to filter candidate sequences for each relation
, and then automatically select the top remaining sequences by applying a heuristic thresholding strategy based on the difference of the distance scores: sort the sequences by increasing distance
and set the cutoff point to be the th sequence where .5.3 Experiments
We evaluate our rule extraction method (denoted as EmbedRule) on the FB15k401 dataset. In our experiments, we remove the equivalence relations and relations whose domains have cardinality since rules involving these relations are not interesting. This results in training data that contains 485,741 facts, 14,417 entities, and 373 relations. Our EmbedRule algorithm identifies 60,020 possible length2 relation sequences and 2,156,391 possible length3 relation sequences. We then apply the thresholding method described in Section 5.2 to further select top 3.9K length2 rules and 2K length3 rules ^{4}^{4}4We consider =100 nearestneighbor sequences for each method, and set to 9.2, 36.3, 1.9 and 3.4 for DistMulttanhEVinit, DistMult, Bilinear and DistAdd respectively for length2 rules, and set it to 9.1, 48.8, 2.9, and 1.1 for lengh3 rules.. By default all the extracted rules are ranked by decreasing confidence, which is computed as the ratio of the correct predictions to the total number of predictions, where predictions are triplets that are derived from the instantiated rules where the body relations are observed.
We implemented four versions of EmbedRule using embeddings trained from TransE (DistAdd), Bilinear, Bilineardiag (DistMult) and DistMulttanhEVinit with corresponding composition functions. We also compare our approaches to AMIE (Galárraga et al., 2013), a stateoftheart rule mining system that can efficiently search for Horn rules in largescale KBs by using novel measurements of support and confidence. The system extracts close rules – a superset of the rules we consider in this paper: every relation in the body is connected to the following relation by sharing an entity variable, and every entity variable in the rule appears at least twice. We run AMIE with the default setting on the same training set. It extracts 2,201 possible length1 rules and 46,975 possible length2 rules, among which 3,952 rules have the closepath property. We compare these length2 rules with the similar number of length2 rules extracted by EmbedRule. By default AMIE ranks rules by PCA confidence (a normalized confidence that takes into account the incompleteness of KBs). However we found that ranking by the standard confidence gives better performance than the PCA confidence on the Freebase dataset we use.
For computational cost, mines length2 rules in minutes and mines length3 rules in minutes (the computational time is similar when using different types of embeddings). AMIE mines rules of length in minutes. All methods are evaluated on a machine with a 64bit processor, 2 CPUs and 8GB memory.
We consider precision as the evaluation metric, which is the ratio of predictions that are in the test (unseen) data to all the generated unseen predictions. Note that this is an estimation, since a prediction is not necessarily “incorrect” if it is not seen.
Galárraga et al. (2013) suggested to identify incorrect predictions based on the functional property of relations. However, we find that most relations in our datasets are not functional. For a better estimation, we manually labeled the top 30 unseen facts predicted by each method by checking Wikipedia. We also remove rules where the head relations are hard to justified due to dynamic factors (i.e. involving the word “current”).5.4 Results
Figure 1 compares the predictions generated by the length2 rules extracted by different methods. We plot the aggregated precision of the top rules that produce up to predictions in total. From left to right, the th data point represents the total number of predictions of the top rules and the estimated precision of these predictions. We can see that EmbedRule that uses embeddings trained from the bilinear objective (Bilinear, DistMult and DistMulttanhEVinit) consistently outperforms AMIE. This suggests that the bilinear embeddings contain good amount of information about relations which allows for effective rule selection without looking at the entities. For example, AMIE fails to extract by relying on the instantiations of the rule occurred in the observed KB while all the bilinear variants of EmbedRule successfully extract the rule purely based on the embeddings of the three involved relations.
We can also see that in general, using multiplicative composition of matrix embeddings (from DistMult and Bilinear) results in better performance compared to using additive composition of vector embeddings (from DistAdd). We found many examples where DistAdd fails to retrieve rules because it assigns large distance between the composition of the body relations and the head relation in the embedding space while its multiplicative counterpart DistMult ranks the composition result much closer to the head relation. For example, DistAdd prunes the possible composition for relation while DistMult ranks the composition as the nearest neighbor of .
We also look at the results for length3 rules generated by different implementations of EmbedRule in Figure 2. We can see that the initial length3 rules extracted by EmbedRule can provide very good precision in general. We can also see that Bilinear consistently outperforms DistMult and DistAdd on the top 1K predictions and DistMulttanhEVinit tends to outperform the other methods as more predictions are generated. We think that the fact that Bilinear starts to show more advantage over DistMult in extracting longer rules confirm the limitation of representing relations by diagonal matrices, as longer rules requires the modeling of more complex relation semantics.
6 Conclusion
In this paper, we present a general framework for learning representations of entities and relations in KBs. Under the framework, we empirically evaluate different embedding models on knowledge inference tasks. We show that a simple formulation of bilinear model can outperform the stateoftheart embedding models for link prediction on Freebase. Furthermore, we examine the learned embeddings by utilizing them to extract logical rules from KBs. We show that embeddings learned from the bilinear objective can capture compositional semantics of relations and be successfully used to extract Horn rules that involve compositional reasoning. For future work, we aim to exploit deep structure in the neural network framework. As learning representations using deep networks has shown great success in various applications (Hinton et al., 2012; Vinyals et al., 2012; Deng et al., 2013)
, it may also help capturing hierarchical structure hidden in the multirelational data. Further, tensor constructs have been usefully applied to some deep learning architectures
(Yu et al., 2013; Hutchinson et al., 2013). Related constructs and architectures may help improve multirelational learning and inference.Appendix
A Examples of the extracted Horn rules
Examples of length2 rules extracted by EmbedRule with embeddings learned from DistMulttanhEVinit:
Examples of length3 rules extracted by EmbedRule with embeddings learned from DistMulttanhEVinit:
B Visualization of the relation embeddings
Visualization of the relation embeddings learned by DistMult and DistAdd using tSNE (see figure 3 and 4). We selected relations in the FB15k401 dataset. The embeddings learned by DistMult nicely reflect the clustering structures among these relations (e.g. /film/release_region is closed to /film/country); whereas the embeddings learned by DistAdd present structure that is harder to interpret.
References
 Auer et al. (2007) Auer, Sören, Bizer, Christian, Kobilarov, Georgi, Lehmann, Jens, Cyganiak, Richard, and Ives, Zachary. Dbpedia: A nucleus for a web of open data. In The semantic web, pp. 722–735. Springer, 2007.
 Bordes et al. (2011) Bordes, Antoine, Weston, Jason, Collobert, Ronan, and Bengio, Yoshua. Learning structured embeddings of knowledge bases. In AAAI, 2011.
 Bordes et al. (2013a) Bordes, Antoine, Glorot, Xavier, Weston, Jason, and Bengio, Yoshua. A semantic matching energy function for learning with multirelational data. Machine Learning, pp. 1–27, 2013a.
 Bordes et al. (2013b) Bordes, Antoine, Usunier, Nicolas, GarciaDuran, Alberto, Weston, Jason, and Yakhnenko, Oksana. Translating embeddings for modeling multirelational data. In NIPS, 2013b.
 Bowman (2014) Bowman, Samuel R. Can recursive neural tensor networks learn logical reasoning? In ICLR, 2014.
 Chang et al. (2014) Chang, KaiWei, Yih, Wentau, Yang, Bishan, and Meek, Chris. Typed tensor decomposition of knowledge bases for relation extraction. In EMNLP, 2014.
 Deng et al. (2013) Deng, Li, Hinton, G., and Kingsbury, B. New types of deep neural network learning for speech recognition and related applications: An overview. In in ICASSP, 2013.
 Duchi et al. (2011) Duchi, John, Hazan, Elad, and Singer, Yoram. Adaptive subgradient methods for online learning and stochastic optimization. The Journal of Machine Learning Research, 12:2121–2159, 2011.
 Galárraga et al. (2013) Galárraga, Luis Antonio, Teflioudi, Christina, Hose, Katja, and Suchanek, Fabian. Amie: association rule mining under incomplete evidence in ontological knowledge bases. In WWW, 2013.
 Gao et al. (2014) Gao, Jianfeng, Pantel, Patrick, Gamon, Michael, He, Xiaodong, Deng, Li, and Shen, Yelong. Modeling interestingness with deep neural networks. In EMNLP, 2014.
 GarcíaDurán et al. (2014) GarcíaDurán, Alberto, Bordes, Antoine, and Usunier, Nicolas. Effective blending of two and threeway interactions for modeling multirelational data. In Machine Learning and Knowledge Discovery in Databases, pp. 434–449. Springer, 2014.
 Getoor & Taskar (2007) Getoor, Lise and Taskar, Ben (eds.). Introduction to Statistical Relational Learning. The MIT Press, 2007.
 Grefenstette (2013) Grefenstette, Edward. Towards a formal distributional semantics: Simulating logical calculi with tensors. In *SEM, 2013.
 Hinton et al. (2012) Hinton, Geoff, Deng, L., Yu, D., Dahl, G., Mohamed, A., Jaitly, N., Senior, A., Vanhoucke, V., Nguyen, P., Sainath, T., and Kingsbury, B. Deep neural networks for acoustic modeling in speech recognition. IEEE Sig. Proc. Mag., 29:82–97, 2012.
 Huang et al. (2013) Huang, PoSen, He, Xiaodong, Gao, Jianfeng, Deng, Li, Acero, Alex, and Heck, Larry. Learning deep structured semantic models for Web search using clickthrough data. In CIKM, 2013.
 Hutchinson et al. (2013) Hutchinson, B, Deng, L., and Yu, D. Tensor deep stacking networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):1944–1957, 2013.
 Jenatton et al. (2012) Jenatton, Rodolphe, Le Roux, Nicolas, Bordes, Antoine, and Obozinski, Guillaume. A latent factor model for highly multirelational data. In NIPS, 2012.
 Kemp et al. (2006) Kemp, Charles, Tenenbaum, Joshua B, Griffiths, Thomas L, Yamada, Takeshi, and Ueda, Naonori. Learning systems of concepts with an infinite relational model. In AAAI, volume 3, pp. 5, 2006.
 Mikolov et al. (2013) Mikolov, Tomas, Sutskever, Ilya, Chen, Kai, Corrado, Greg S, and Dean, Jeff. Distributed representations of words and phrases and their compositionality. In NIPS, pp. 3111–3119, 2013.
 Nickel et al. (2011) Nickel, Maximilian, Tresp, Volker, and Kriegel, HansPeter. A threeway model for collective learning on multirelational data. In ICML, pp. 809–816, 2011.
 Nickel et al. (2012) Nickel, Maximilian, Tresp, Volker, and Kriegel, HansPeter. Factorizing YAGO: scalable machine learning for linked data. In WWW, pp. 271–280, 2012.
 Paccanaro & Hinton (2001) Paccanaro, Alberto and Hinton, Geoffrey E. Learning distributed representations of concepts using linear relational embedding. IEEE Transactions on Knowledge and Data Engineering, 13(2):232–244, 2001.
 Richardson & Domingos (2006) Richardson, Matthew and Domingos, Pedro. Markov logic networks. Machine learning, 62(12):107–136, 2006.
 Rocktäschel et al. (2014) Rocktäschel, Tim, Bošnjak, Matko, Singh, Sameer, and Riedel, Sebastian. Lowdimensional embeddings of logic. In ACL Workshop on Semantic Parsing, 2014.
 Schoenmackers et al. (2010) Schoenmackers, Stefan, Etzioni, Oren, Weld, Daniel S, and Davis, Jesse. Learning firstorder horn clauses from web text. In EMNLP, 2010.
 Shen et al. (2014a) Shen, Yelong, He, Xiaodong, Gao, Jianfeng, Deng, Li, and Mesnil, Gregoire. A latent semantic model with convolutionalpooling structure for information retrieval. In CIKM, 2014a.

Shen et al. (2014b)
Shen, Yelong, He, Xiaodong, Gao, Jianfeng, Deng, Li, and Mesnil, Grégoire.
Learning semantic representations using convolutional neural networks for Web search.
In WWW, pp. 373–374, 2014b.  Singh & Gordon (2008) Singh, Ajit P and Gordon, Geoffrey J. Relational learning via collective matrix factorization. In KDD, pp. 650–658. ACM, 2008.
 Socher et al. (2012) Socher, Richard, Huval, Brody, Manning, Christopher D., and Ng, Andrew Y. Semantic compositionality through recursive matrixvector spaces. In EMNLPCoNLL, 2012.
 Socher et al. (2013) Socher, Richard, Chen, Danqi, Manning, Christopher D., and Ng, Andrew Y. Reasoning with neural tensor networks for knowledge base completion. In NIPS, 2013.
 Suchanek et al. (2007) Suchanek, Fabian M, Kasneci, Gjergji, and Weikum, Gerhard. Yago: a core of semantic knowledge. In WWW, 2007.
 Sutskever et al. (2009) Sutskever, Ilya, Tenenbaum, Joshua B, and Salakhutdinov, Ruslan. Modelling relational data using Bayesian clustered tensor factorization. In NIPS, pp. 1821–1828, 2009.
 Vinyals et al. (2012) Vinyals, O., Jia, Y., Deng, L., and Darrell, T. Learning with recursive perceptual representations. In NIPS, 2012.
 Yih et al. (2014) Yih, Wentau, He, Xiaodong, and Meek, Christopher. Semantic parsing for singlerelation question answering. In ACL, 2014.
 Yu et al. (2013) Yu, D., Deng, L., and Seide, F. The deep tensor neural network with applications to large vocabulary speech recognition. IEEE Trans. Audio, Speech and Language Proc., 21(2):388 –396, 2013.
Comments
There are no comments yet.