Published as conference paper at ICLR 2015

E MBEDDING E NTITIES AND R ELATIONS FOR L EARN -

ING AND I NFERENCE IN K NOWLEDGE BASES

Bishan Yang1˚, Wen-tau Yih2 , Xiaodong He2 , Jianfeng Gao2 & Li Deng2
1
Department of Computer Science, Cornell University, Ithaca, NY, 14850, USA
bishan@cs.cornell.edu
2
Microsoft Research, Redmond, WA 98052, USA
{scottyih,xiaohe,jfgao,deng}@microsoft.com
arXiv:1412.6575v4 [cs.CL] 29 Aug 2015

A BSTRACT
We consider learning representations of entities and relations in KBs using the
neural-embedding approach. We show that most existing models, including
NTN (Socher et al., 2013) and TransE (Bordes et al., 2013b), can be generalized
under a unified learning framework, where entities are low-dimensional vectors
learned from a neural network and relations are bilinear and/or linear mapping
functions. Under this framework, we compare a variety of embedding models
on the link prediction task. We show that a simple bilinear formulation achieves
new state-of-the-art results for the task (achieving a top-10 accuracy of 73.2%
vs. 54.7% by TransE on Freebase). Furthermore, we introduce a novel ap-
proach that utilizes the learned relation embeddings to mine logical rules such
as BornInCitypa, bq ^ CityInCountrypb, cq ùñ N ationalitypa, cq. We
find that embeddings learned from the bilinear objective are particularly good
at capturing relational semantics, and that the composition of relations is char-
acterized by matrix multiplication. More interestingly, we demonstrate that our
embedding-based rule extraction approach successfully outperforms a state-of-
the-art confidence-based rule mining approach in mining Horn rules that involve
compositional reasoning.

1 I NTRODUCTION
Recent years have witnessed a rapid growth of knowledge bases (KBs) such as Freebase1 , DBPe-
dia (Auer et al., 2007), and YAGO (Suchanek et al., 2007). These KBs store facts about real-world
entities (e.g. people, places, and things) in the form of RDF triples2 (i.e. (subject, predicate, ob-
ject)). 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 large-scale multi-
relational 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 knowl-
edge bases. Tensor factorization (e.g. (Nickel et al., 2011; 2012)) and neural-embedding-based
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 low-dimensional representations of entities and rela-
tions. 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 neural-embedding models, where the representations
are learned using neural networks with energy-based objectives. Recent embedding models
TransE (Bordes et al., 2013b) and NTN (Socher et al., 2013) have shown state-of-the-art predic-
tion 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 rep-
resentations. Without careful comparison, it is not clear how different design choices affect the
˚
Work conducted while interning at Microsoft Research.
1
http://freebase.com
2
http://www.w3.org/TR/rdf11-concepts/

1
Published as conference paper at ICLR 2015

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 low-dimensional 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 multi-
relational learning that unifies most multi-relational embedding models developed in the past, in-
cluding 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 state-of-
the-art results for the task (a top-10 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 BornInCitypa, bq ^ CityOf Countrypb, cq ùñ N ationalitypa, cq.
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 embedding-based approach outperforms a state-of-the-art 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 multi-relational 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 R ELATED W ORK

Multi-relational learning has been an active research area for the past couple of years. Traditional
statistical learning approaches (Getoor & Taskar, 2007) such as Markov-logic networks (Richard-
son & Domingos, 2006) usually suffer from scalability issues. More recently, various types of
representation learning methods have been proposed to embed multi-relational knowledge into low-
dimensional 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 neural-embedding models as they have
shown good scalability and strong generalizability on large-scale KBs.
Existing neural embedding models (Bordes et al., 2013a;b; Socher et al., 2013) all represent entities
as low-dimensional 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 rela-
tion 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 rep-
resent 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 pre-trained vectors from external text corpora.
There has not been work that closely examines the effectiveness of these different design choices.
Our work on embedding-based rule extraction presented in part of this paper is related to the ear-
lier study on logical inference with learned continuous-space representations. Much existing work
along this line focuses on learning logic-based representations for natural language sentences. For
example, Socher et al. (2012) builds a neural network that recursively combines word representa-
tions 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 sim-
ulated 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 evalua-
tion 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

2
Published as conference paper at ICLR 2015

can capture the compositional semantics of relations. Moreover, we systematically evaluate our ap-
proach and compare it favorably with a state-of-the-art rule mining approach on the rule extraction
task on Freebase.

3 M ULTI -R ELATIONAL R EPRESENTATION L EARNING

In this section, we present a general neural network framework for multi-relational 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 pe1 , r, e2 q (denoting e1 (the subject) and
e2 (the object) that are in a certain relationship r), 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 relation-specific parameters.

3.1 E NTITY R EPRESENTATIONS

Each input entity corresponds to a high-dimensional vector, either a “one-hot” index vector or a
“n-hot” feature vector. Denote by xe1 and xe2 the input vectors for entity e1 and e2 , respectively.
Denote by W the first layer projection matrix. The learned entity representations, ye1 and ye2 can
be written as ` ˘ ` ˘
ye1 “ f Wxe1 , ye2 “ f Wxe2
where f can be a linear or non-linear function, and W is a parameter matrix, which can be randomly
initialized or initialized using pre-trained vectors.
Most existing embedding models adopt the “one-hot” 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
“bag-of-words” vectors as input and learning a projection matrix consisting of word vectors.

3.2 R ELATION R EPRESENTATIONS

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 gra ,
a bilinear transformation grb or their combination, where gra and grb are defined as
ˆ ˙
a T ye1
gr pye1 , ye2 q “ Ar and grb pye1 , ye2 q “ yeT1 Br ye2 , (1)
ye2
which Ar and Br are relation-specific parameters.
T
Models Br ` T Ar T ˘ Scoring Function
Distance (Bordes et al., 2011) - `QrT1 ´QTr2 ˘ ´||gra pye1 , ye2 q||1
Single Layer (Socher et al., 2013) - `Qr1 Qr2˘ uTr tanhpgra pye1 , ye2 qq
T T
TransE (Bordes et al., 2013b) I ` VTr ´VTr ˘ ´p2gra pye1 ,`ye2 q ´ 2grb pye1 , ye2 q ` ||Vr˘||22 q
NTN (Socher et al., 2013) Tr Qr1 Qr2 uTr tanh gra pye1 , ye2 q ` grb pye1 , ye2 q

Table 1: Comparisons among several multi-relational models in their scoring functions.

In Table 1, we summarize several popular scoring functions in the literature for a relation
triplet pe1 , r, e2 q, reformulated in terms of the above two functions. Denote by ye1 , ye2 P R n
two entity vectors. Denote by Qr1 , Qr2 P R nˆm and Vr P R n matrix or vector parameters
for linear transformation gra . Denote by Tr P R nˆnˆm tensor parameters for bilinear trans-
formation grb . I P R n is an identity matrix. ur P R m is an additional parameter for rela-
tion r. The scoring function for TransE (L2 formulation) is derived from ||ye1 ´ ye2 ` Vr ||22 “
2VrT pye1 ´ ye2 q ´ 2yeT1 ye2 ` ||Vr ||22 ` ||ye1 ||22 ` ||ye2 ||22 , where ye1 and ye2 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 one-dimensional vectors.

3
Published as conference paper at ICLR 2015

In this paper, we also consider the basic bilinear scoring function:

grb pye1 , ye2 q “ yeT1 Mr ye2 (2)
which is a special case of NTN without the non-linear layer and the linear operator, and uses a 2-d
matrix operator Mr P R nˆn 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; Garcı́a-Durán
et al., 2014) with different forms of regularization. Here, we consider a simple way to reduce the
number of relation parameters by restricting Mr 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 deep-structured 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 relation-independent scoring
function Gr ye1 , ye2 “ cosrye1 pWr q, ye2 pWr qs. The cosine scoring function is a special case of
grb with normalized ye1 , ye2 and with Br “ I.

3.3 PARAMETER L EARNING

The neural network parameters of all the models discussed above can be learned by minimizing
a margin-based 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 T , we can construct a set of “negative”
triplets T 1 by corrupting either one of the relation arguments, T 1 “ tpe11 , r, e2 q|e11 P E, pe11 , r, e2 q R
T u Y tpe1 , r, e12 q|e12 P E, pe1 , r, e12 q R T u. Denote the scoring function for triplet pe1 , r, e2 q as
Spe1 ,r,e2 q . The training objective is to minimize the margin-based ranking loss
ÿ ÿ
LpΩq “ maxtSpe11 ,r,e12 q ´ Spe1 ,r,e2 q ` 1, 0u (3)
pe1 ,r,e2 qPT pe11 ,r,e12 qPT 1

4 I NFERENCE TASK I: L INK P REDICTION

We first conduct a comparison study of different embedding models on the canonical link predic-
tion 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 Re-
ciprocal Rank (MRR) (an average of the reciprocal rank of an answered entity over all test triplets),
HITS@10 (top-10 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 4 tensor slices
as in (Socher et al., 2013); (2) Bilinear+Linear, NTN with 1 tensor slice and without the non-linear
layer; (3) TransE, a special case of Bilinear+Linear (see Table 1); (4) Bilinear: using scoring function
in Eq. (2); (5) Bilinear-diag: 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 151, 442 triplets with 40, 943 entities and 18 relations, and FB15k consists of
592, 213 triplets with 14, 951 entities and 1345 relations. We use the same training/validation/test
split as in (Bordes et al., 2013b). We also consider a subset of FB15k (FB15k-401) containing only
frequent relations (relations with at least 100 training examples). This results in 560, 209 triplets
with 14, 541 entities and 401 relations.

Implementation details All the models were implemented in C# and using GPU. Training was
implemented using mini-batch 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

4
Published as conference paper at ICLR 2015

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 mini-batches to 10, the dimensionality of the entity vector d “ 100, the regularization parameter
0.0001, and the number of training epochs T “ 100 on FB15k and FB15k-401 and T “ 300 on WN
(T was determined based on the learning curves where the performance of all models plateaued.)
The learning rate was initially set to 0.1 and then adapted during training by AdaGrad.

4.1 R ESULTS

FB15k FB15k-401 WN
MRR HITS@10 MRR HITS@10 MRR HITS@10
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 (D ISTADD) 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
Bilinear-diag (D IST M ULT) 0.35 57.7 0.36 58.5 0.83 94.2

Table 2: Performance comparisons among different embedding models

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 –
B ILINEAR - DIAG, clearly outperforms all baselines on FB and achieves comparable performance to
Bilinear on WN. Note that B ILINEAR - DIAG 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
B ILINEAR - DIAG is able to provide good prediction performance.
Multiplicative vs. Additive Interactions Note that B ILINEAR - DIAG and T RANS E have the same
number of model parameters and their difference can be viewed as the operational choices of the
composition of two entity vectors – B ILINEAR - DIAG uses weighted element-wise dot product (mul-
tiplicative operation) and T RANS E uses element-wise subtraction with a bias (additive operation).
To highlight the difference, here we use D IST M ULT and D ISTA DD to refer to B ILINEAR - DIAG and
T RANS E, 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 model-
ing entity relations. Overall, we observed superior performance of D IST M ULT 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 FB15k-401 when predicting the subject entity and the object entity respectively.
We can see that D IST M ULT significantly outperforms D ISTA DD in almost all the categories.

Predicting subject entities Predicting object entities

1-to-1 1-to-n n-to-1 n-to-n 1-to-1 1-to-n n-to-1 n-to-n
D ISTADD 70.0 76.7 21.1 53.9 68.7 17.4 83.2 57.5
D IST M ULT 75.5 85.1 42.9 55.2 73.7 46.7 81.0 58.8

Table 3: Results by relation categories: one-to-one, one-to-many, many-to-one and many-to-many

Initialization of Entity Vectors In the following, we examine the learning of entity representations
and introduce two further improvements: using non-linear projection and initializing entity vectors
with pre-trained vectors. We focus on D IST M ULT as our baseline and compare it with the two
modifications D IST M ULT-tanh (using f “ tanh for entity projection ) and D IST M ULT-tanh-EV-init

5
Published as conference paper at ICLR 2015

(initializing the entity parameters with the 1000-dimensional pre-trained entity vectors released by
word2vec (Mikolov et al., 2013)) on FB15k-401. 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 300-dimensional pre-trained word vectors released by
word2vec. We denote this method as D IST M ULT-tanh-WV-init. 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 HITS@10 MAP (w/ type checking)

D IST M ULT 0.36 58.5 64.5
D IST M ULT-tanh 0.39 63.3 76.0
D IST M ULT-tanh-WV-init 0.28 52.5 65.5
D IST M ULT-tanh-EV-init 0.42 73.2 88.2

Table 4: Evaluation with pre-trained vectors

In Table 4, we can see that D IST M ULT-tanh-EV-init provides the best performance on all the metrics.
Surprisingly, we observed performance drops by D IST M ULT-tanh-WV-init. We suspect that this
is because word vectors are not appropriate for modeling entities described by non-compositional
phrases (more than 73% of the entities in FB15k-401 are person names, locations, organizations and
films). The promising performance of D IST M ULT-tanh-EV-init suggests that the embedding model
can greatly benefit from pre-trained entity-level vectors using external textual resources.

5 I NFERENCE TASK II: RULE E XTRACTION

In this section, we focus on a complementary inference task, where we utilize the learned embed-
dings 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:
BornInCitypa, bq ^ CityOf Countrypb, cq ùñ N ationalitypa, cq
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. closed-paths 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 embedding-based 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 N OTATIONS

For a better illustration, we adopt the graph view of KB. Each binary relation rpa, bq is a directed
edge from node a to node b and with link type r. We are interested in extracting Horn rules that
consist of a head relation H and a sequence of body relations B1 , ..., Bn :
B1 pa1 , a2 q ^ B2 pa2 , a3 q ^ ... ^ Bn pan , an`1 q ùñ Hpa1 , an`1 q (4)
where ai are variables that can be substituted by entities. We constrain the body relations B1 , ..., Bn
to form a directed path in the graph and the head relation H 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
closed-path property. For consecutive relations that share one variable but do not form a path,

6
Published as conference paper at ICLR 2015

e,g, Bi´1 pa, bq ^ Bi pa, cq, we can replace one of the relations with its inverse relation, so that the
´1
relations are connected by an object and an subject, e.g. Bi´1 pb, aq ^ Bi pa, cq. We are interested
in mining rules that reflect relationships among different relation types, therefore we also constrain
B1 , ..., Bn , H to have distinct relation types. A rule is instantiated when all variables are substi-
tuted 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 r, we denote the domain of its subject argument (the set of entities that can appear
in the subject position) as Xr and similarly the domain of its object argument as Yr . Such domain
information can be extremely useful in restricting the search space of logical rules.

5.2 E MBEDDING - BASED RULE E XTRACTION

For simplicity, we consider Horn rules of length 2 (longer rules can be easily derived from this case):

B1 pa, bq ^ B2 pb, cq ùñ Hpa, cq (5)

Note that the body of the rule can be viewed as the composition of relations B1 and B2 , which is a
new relation that has the property that entities a and c are in a relation if and only if there is an entity
b which simultaneously satisfies two relations B1 pa, bq and B2 pb, cq.
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 T RANS E) and matrices (as in
B ILINEAR 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 V and assume ya `V´yb « 0 when Bpa, bq holds, then we have the property
that ya ` V1 « yb and yb ` V2 « yc implies ya ` pV1 ` V2 q « yc ; (2) if a relation corresponds
to a matrix M in the bilinear transformation and assume yaT Myb « 1 when Bpa, bq holds, also ya
and yb are unit vectors and yaT M is still a unit vector 3 , then we have the property that yaT M1 « ybT
and ybT M2 « ycT implies yaT pM1 M2 q « ycT .
To simulate the implication in 5, we want the composition result of relation B1 and B2 to demon-
strate similar behavior to the embedding of relation H. We assume the similarity between relation
embeddings is related to the Euclidean distance if the embeddings are vectors and to the Frobe-
nius 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 r, then we are only interested in relation pairs pp, qq that satisfy the type
constraints, namely: (1) Yp X Xq ‰ H; (2) Xp X Xr ‰ H; (3) Yq X Yr ‰ H. 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 closed-path Horn rules as in
Eq. (4). In Step 7, ˝ denotes vector addition or matrix multiplication. We apply a global threshold
value δ in our experiments to filter candidate sequences for each relation r, and then automatically
select the top remaining sequences by applying a heuristic thresholding strategy based on the differ-
ence of the distance scores: sort the sequences by increasing distance d1 , ..., dT and set the cut-off
point to be the j-th sequence where j “ arg maxi pdi`1 ´ di q.

5.3 E XPERIMENTS

We evaluate our rule extraction method (denoted as E MBED RULE) on the FB15k-401 dataset. In
our experiments, we remove the equivalence relations and relations whose domains have cardinality
3
These assumptions may not hold in our implementations. The intuition still leads to surprisingly good em-
pirical performance on Horn rule extraction. How to effectively enforce these constraints is worth investigating
in future work.

7
Published as conference paper at ICLR 2015

Algorithm 1 E MBED RULE

1: Input: KB “ tpe1 , r, e2 qu, relation set R
2: Output: Candidate rules Q
3: for each r in R do
4: Select the set of start relations S “ ts : Xs X Xr ‰ Hu
5: Select the set of end relations T “ tt : Yt X Yr ‰ Hu
6: Find all possible relation sequences
7: Select the K-NN sequences P 1 Ď P for r based on distpMr , Mp1 ˝ ¨ ¨ ¨ ˝ Mpn q
8: Form candidate rules using P 1 where r is the head relation and p P P 1 is the body in a rule
9: Add the candidate rules into Q
10: end for

1 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 E MBED RULE algorithm identifies 60,020
possible length-2 relation sequences and 2,156,391 possible length-3 relation sequences. We then
apply the thresholding method described in Section 5.2 to further select top „3.9K length-2 rules
and „2K length-3 rules 4 . 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 E MBED RULE using embeddings trained from T RANS E (D IS -
TA DD ), B ILINEAR , B ILINEAR - DIAG (D IST M ULT ) and D IST M ULT -tanh-EV-init with correspond-
ing composition functions. We also compare our approaches to AMIE (Galárraga et al., 2013), a
state-of-the-art rule mining system that can efficiently search for Horn rules in large-scale KBs by
using novel measurements of support and confidence. The system extracts close rules – a super-
set 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 length-1
rules and 46,975 possible length-2 rules, among which 3,952 rules have the close-path property. We
compare these length-2 rules with the similar number of length-2 rules extracted by E MBED RULE.
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, EmbedRule mines length-2 rules in 2 minutes and mines length-3 rules in
20 minutes (the computational time is similar when using different types of embeddings). AMIE
mines rules of length ď 2 in 9 minutes. All methods are evaluated on a machine with a 64-bit
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 R ESULTS

Figure 1 compares the predictions generated by the length-2 rules extracted by different meth-
ods. We plot the aggregated precision of the top rules that produce up to 10K predictions in to-
tal. From left to right, the n-th data point represents the total number of predictions of the top
n rules and the estimated precision of these predictions. We can see that E MBED RULE that uses
embeddings trained from the bilinear objective (B ILINEAR, D IST M ULT and D IST M ULT- TANH -
4
We consider K=100 nearest-neighbor sequences for each method, and set δ to 9.2, 36.3, 1.9 and 3.4 for
D IST M ULT- TANH -EV- INIT, D IST M ULT, B ILINEAR and D ISTA DD respectively for length-2 rules, and set it
to 9.1, 48.8, 2.9, and 1.1 for lengh-3 rules.

8
Published as conference paper at ICLR 2015

EV- INIT) consistently outperforms AMIE. This suggests that the bilinear embeddings contain
good amount of information about relations which allows for effective rule selection without look-
ing at the entities. For example, AMIE fails to extract T V P rogramCountryof Originpa, bq ^
CountryOf f icialLanguagepb, cq ùñ T V P rogramLanguagepa, cq by relying on the instan-
tiations of the rule occurred in the observed KB while all the bilinear variants of E MBED RULE
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 D IST-
M ULT and B ILINEAR) results in better performance compared to using additive composition of
vector embeddings (from D ISTA DD). We found many examples where D ISTA DD 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 D IST M ULT ranks the composi-
tion result much closer to the head relation. For example, D ISTA DD prunes the possible composition
F ilmDistributorInRegion^RegionGDP Currency for relation F ilmBudgetCurrency while
D IST M ULT ranks the composition as the nearest neighbor of F ilmBudgetCurrency.

Figure 1: Aggregated precision of top length-2 rules extracted by different methods

Figure 2: Aggregated precision of top length-3 rules extracted by different methods

We also look at the results for length-3 rules generated by different implementations of E MBED RULE
in Figure 2. We can see that the initial length-3 rules extracted by E MBED RULE can provide very
good precision in general. We can also see that B ILINEAR consistently outperforms D IST M ULT and
D ISTA DD on the top 1K predictions and D IST M ULT- TANH -EV- INIT tends to outperform the other
methods as more predictions are generated. We think that the fact that B ILINEAR starts to show more
advantage over D IST M ULT in extracting longer rules confirm the limitation of representing relations
by diagonal matrices, as longer rules requires the modeling of more complex relation semantics.

9
Published as conference paper at ICLR 2015

6 C ONCLUSION
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 state-of-
the-art 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 multi-relational 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 multi-relational
learning and inference.

A PPENDIX
A E XAMPLES OF THE EXTRACTED H ORN RULES

Examples of length-2 rules extracted by E MBED RULE with embeddings learned from D IST M ULT-
tanh-EV-init:
AwardInCeremanypa, bq ^ CeremanyEventT ypepb, cq ùñ AwardInEventT ypepa, cq
AtheleteP layInT eampa, bq ^ T eamP laySportpb, cq ùñ AtheleteP laySportpa, cq
T V P rogramInT V N etworkpa, bq^T V N etworkServiceLanguagepb, cq ùñ T V P rogramLanguagepa, cq
LocationInStatepa, bq ^ StateInCountrypb, cq ùñ LocationInCountrypa, cq
BornInLocationpa, bq ^ LocationInCountrypb, cq ùñ N ationalitypa, cq

Examples of length-3 rules extracted by E MBED RULE with embeddings learned from D IST M ULT-
tanh-EV-init:
SportP layByT eampa, bq^T eamInClubpb, cq^ClubHasP layerpc, dq ùñ SportP layByAtheletepa, dq
M usicT rackP erf ormerpa, bq^P eerInf luencepb, cq^P erf ormRolepc, dq ùñ M usicT rackRolepa, dq
F ilmHasActorpa, bq^CelebrityF riendshippb, cq^P ersonLanguagepc, dq ùñ F ilmLanguagepa, dq

B V ISUALIZATION OF THE RELATION EMBEDDINGS

Visualization of the relation embeddings learned by D IST M ULT and D ISTA DD using t-SNE (see
figure 3 and 4). We selected 189 relations in the FB15k-401 dataset. The embeddings learned by
D IST M ULT nicely reflect the clustering structures among these relations (e.g. /film/release region
is closed to /film/country); whereas the embeddings learned by D ISTA DD present structure that is
harder to interpret.

R EFERENCES
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, Antoine, Weston, Jason, Collobert, Ronan, and Bengio, Yoshua. Learning structured em-
beddings of knowledge bases. In AAAI, 2011.
Bordes, Antoine, Glorot, Xavier, Weston, Jason, and Bengio, Yoshua. A semantic matching energy
function for learning with multi-relational data. Machine Learning, pp. 1–27, 2013a.
Bordes, Antoine, Usunier, Nicolas, Garcia-Duran, Alberto, Weston, Jason, and Yakhnenko, Oksana.
Translating embeddings for modeling multi-relational data. In NIPS, 2013b.

10
Published as conference paper at ICLR 2015

Figure 3: Relation embeddings (D ISTA DD)

Figure 4: Relation embeddings (D IST M ULT)

Bowman, Samuel R. Can recursive neural tensor networks learn logical reasoning? In ICLR, 2014.

Chang, Kai-Wei, Yih, Wen-tau, Yang, Bishan, and Meek, Chris. Typed tensor decomposition of
knowledge bases for relation extraction. In EMNLP, 2014.

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, 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, Luis Antonio, Teflioudi, Christina, Hose, Katja, and Suchanek, Fabian. Amie: associa-
tion rule mining under incomplete evidence in ontological knowledge bases. In WWW, 2013.

Gao, Jianfeng, Pantel, Patrick, Gamon, Michael, He, Xiaodong, Deng, Li, and Shen, Yelong. Mod-
eling interestingness with deep neural networks. In EMNLP, 2014.

Garcı́a-Durán, Alberto, Bordes, Antoine, and Usunier, Nicolas. Effective blending of two and three-
way interactions for modeling multi-relational data. In Machine Learning and Knowledge Dis-
covery in Databases, pp. 434–449. Springer, 2014.

Getoor, Lise and Taskar, Ben (eds.). Introduction to Statistical Relational Learning. The MIT Press,
2007.

Grefenstette, Edward. Towards a formal distributional semantics: Simulating logical calculi with
tensors. In *SEM, 2013.

11
Published as conference paper at ICLR 2015

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, Po-Sen, 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, B, Deng, L., and Yu, D. Tensor deep stacking networks. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 35(8):1944–1957, 2013.
Jenatton, Rodolphe, Le Roux, Nicolas, Bordes, Antoine, and Obozinski, Guillaume. A latent factor
model for highly multi-relational data. In NIPS, 2012.
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, Tomas, Sutskever, Ilya, Chen, Kai, Corrado, Greg S, and Dean, Jeff. Distributed represen-
tations of words and phrases and their compositionality. In NIPS, pp. 3111–3119, 2013.
Nickel, Maximilian, Tresp, Volker, and Kriegel, Hans-Peter. A three-way model for collective learn-
ing on multi-relational data. In ICML, pp. 809–816, 2011.
Nickel, Maximilian, Tresp, Volker, and Kriegel, Hans-Peter. Factorizing YAGO: scalable machine
learning for linked data. In WWW, pp. 271–280, 2012.
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, Matthew and Domingos, Pedro. Markov logic networks. Machine learning, 62(1-2):
107–136, 2006.
Rocktäschel, Tim, Bošnjak, Matko, Singh, Sameer, and Riedel, Sebastian. Low-dimensional em-
beddings of logic. In ACL Workshop on Semantic Parsing, 2014.
Schoenmackers, Stefan, Etzioni, Oren, Weld, Daniel S, and Davis, Jesse. Learning first-order horn
clauses from web text. In EMNLP, 2010.
Shen, Yelong, He, Xiaodong, Gao, Jianfeng, Deng, Li, and Mesnil, Gregoire. A latent semantic
model with convolutional-pooling structure for information retrieval. In CIKM, 2014a.
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, Ajit P and Gordon, Geoffrey J. Relational learning via collective matrix factorization. In
KDD, pp. 650–658. ACM, 2008.
Socher, Richard, Huval, Brody, Manning, Christopher D., and Ng, Andrew Y. Semantic composi-
tionality through recursive matrix-vector spaces. In EMNLP-CoNLL, 2012.
Socher, Richard, Chen, Danqi, Manning, Christopher D., and Ng, Andrew Y. Reasoning with neural
tensor networks for knowledge base completion. In NIPS, 2013.
Suchanek, Fabian M, Kasneci, Gjergji, and Weikum, Gerhard. Yago: a core of semantic knowledge.
In WWW, 2007.
Sutskever, Ilya, Tenenbaum, Joshua B, and Salakhutdinov, Ruslan. Modelling relational data using
Bayesian clustered tensor factorization. In NIPS, pp. 1821–1828, 2009.
Vinyals, O., Jia, Y., Deng, L., and Darrell, T. Learning with recursive perceptual representations. In
NIPS, 2012.
Yih, Wen-tau, He, Xiaodong, and Meek, Christopher. Semantic parsing for single-relation question
answering. In ACL, 2014.
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.

12