1 Introduction
Relation extraction has been widely used for many applications, such as knowledge graph construction
Dong et al. (2014), information retrieval Liu et al. (2014), and question answering Ravichandran and Hovy (2002). Traditional supervised approaches require direct annotation on sentences with a relatively small number of relations Roth and Yih (2002); Kambhatla (2004).^{1}^{1}1We distinguish a relation (e.g., a predicate in a knowledge base) from the relation expression (e.g., the text surface between entities in a sentence) throughout the paper. With the development of largescale knowledge bases (KBs) such as Freebase Bollacker et al. (2008), relation extraction has been extended to larger scales comparable to KBs using the distant supervision Mintz et al. (2009). However, when the training corpus does not support the annotated relations showing in the KB, such approach could fail to find sufficient training examples. Distant supervision assumption can be violated by up to 31% for some relations when aligning to NYT corpus Riedel et al. (2010). More importantly, either traditional supervised learning or distantly supervised learning cannot discover new relations unseen in the training phase.
Unsupervised relation discovery tries to overcome the shortcomings of supervised or distantly supervised learning approaches. Existing approaches either extract surface or syntactic patterns from sentences and use relation expressions as predicates (which result in many noisy relations) Etzioni et al. (2004); Banko et al. (2007), or cluster the relation expressions based on the extracted triplets to form relation clusters Yao et al. (2011, 2012); Marcheggiani and Titov (2016). However, these approaches do not use existing highquality and largescale KBs when they are relevant to the relations to be discovered.
In this paper, we consider a new relation discovery problem where both the training corpus for relation clustering and a KB are available, but the relations in the training corpus and those in the KB are not overlapped. As shown in Figure 1, in the KB, we have entities Pink Floyd, Animals, etc., with some existing relations notable_work and has_member in the KB. However, when doing relation discovery, we can only get supporting sentences that suggest new relations based_on and influenced_by. This is a common and practical problem since predicates in KBs are limited to the annotator defined relations while the real relations in the world are always open and creative.
It is challenging when there is no overlapped relation between target relation clusters and the KB because in this case the KB is not a direct supervision. But if target relation clusters and the KB share some entities, we can use the shared entities as a bridge to introduce indirect supervision for the relation discovery problem. Specifically, we build constraints between pairs of tuples based on the KB. For example, in Figure 1, when we cluster the based_on relation, we can evaluate the similarity between the tuple (Animals, Animal Farm) and the tuple (Amused to Death, Amusing Ourselves to Death) based on the KB. If the KB tells us these two pairs of tuples are close to each other, then we put a constraint to force our relation clustering algorithm to group them together.
We use the discretestate variational autoencoder (DVAE) framework
Marcheggiani and Titov (2016) as our base relation discovery model since this framework is flexible to incorporate different features and currently the stateoftheart. We use KB embedding Bordes et al. (2013) to obtain entity embeddings in the KB and use entity embeddings to evaluate the similarity between a pair of tuples. Then constraints are constructed and incorporated into the DVAE framework in a way inspired by the mustlink and cannotlink based constrained clustering Basu et al. (2004). We show that with no overlapped relations between the KB and the training corpus, we can improve the relation discovery by a large margin.Our contributions are summarized as follows.

We study a new prevalent but challenging task of relation discovery where the training corpus and the KB have no overlapped relation.

We propose a new kind of indirect supervision to relation discovery which is built based on pairwise constraints between two tuples.

We show promising results using existing relation discovery datasets to demonstrate the effectiveness of our proposed learning algorithm for the new relation discovery task.
The code we used to train and evaluate our models is available at https://github.com/HKUSTKnowComp/RERegDVAE.
2 Problem Definition
We use to denote the set of all training sentences. is the set of named entities that are recognized by an NER system in , and is the pair of first and second entities in a given sentence . is the set of relation labels for . In addition, there exists an external knowledge base , consisting of a set of entities and relations and triplets where a triplet consists of two entities with their relation.
Our model is a relation extractor to predict the underlying semantic relation given sentences , with the help of . In particular, we focus on the challenging scenario where .
3 Model
In this section, we first review the discretestate variational autoencoder (DVAE) in §3.1. Then we introduce our new framework in §3.2.
3.1 DVAE for Relation Discovery
Assuming that we perform generative modeling, where each latent relation follows a uniform prior distribution , we follow Marcheggiani and Titov (2016) to optimize a pseudolikelihood:
(1)  
(2) 
where and are entities, and denotes the complement .
is the probability of one entity given another entity as well as the relation, where
denotes the set of parameters. Note that this probability is defined on the triplet which is universal across different sentences containing the two entities.The pseudolikelihood can be lowerbounded based on Jensen’s inequality through a variational posterior :
(3) 
where predicts the relation based on the whole sentence as an input and as the set of parameters.
is the entropy to regularize the probability distribution
, and is the hyperparameter to balance the regularization strength.This model consists of two components, an encoder which encodes sentence features into a relation distribution, and a decoder which predicts an entity given the relation cluster and another entity. Both are modeled by softmax functions:
(4)  
(5) 
where and
is a vector representation of sentence
, which can be highdimensional onehot feature encodings or lowdimensional sentence embeddings encoded by deep neural networks.
can be a general scoring function defined over triplets. We use the instantiation with the best performance shown by Marcheggiani and Titov (2016), which is a combination of bilinear model and selectional preference model:(6) 
where , is a matrix, is a vector for the relation , and are the vectors for head and tail entities respectively, and is the concatenation of the vector representations of the two entities.
The DVAE model directly optimizes the variational lower bound by doing gradient ascent for and jointly. Both encoder and decoder are implemented as neural networks. Standard training techniques and tricks can be applied.
3.2 Knowledge Base Constraint
Our KB constraint framework can be summarized as a twostep procedure: KB constraints construction and regularization for the learning model. In the constraints construction step, a set of sentences is formed as a query to KB and retrieves a set of constraints back. Then in the regularization step, we apply the constraint to regularize posterior distributions of the relation extractor.
Conceptually, given a set of sentences , we want to bias the learning result: After the entities are linked to the KB, if KB inference indicates that some pairs should be in a relation based on a set of rules , then the extractor should be constrained to output it. This constraint can be encoded into a feature function = “entity pairs in the same relation based on ” and put into the posterior regularization framework Gillenwater et al. (2011). However, the computational complexity of the feature function is exponential since we need to traverse the KB to find . We instead consider the mustlink and cannotlink constraints Basu et al. (2004), indicating respectively that a pair of sentences should be or should not be labeled as the same relation. For each pairwise constraint, the model assigns an associated cost of violating that constraint for the model regularization.
Euclidean distance  

KullbackLeibler (KL) divergence  
JensenShannon (JS) divergence 
3.2.1 KB Constraints Construction
From the perspective of KB, a mustlink constraint on sentences exists if two pairs of entities are similar given the KB, where is the entity pair belongs to sentence . This motivates us to define a similarity score for a pair of entity pairs. Instead of modeling the common relation paths or logic rules, which is computationally infeasible, we compare them in the latent embedding space. In particular, we model the KB using the TransE Bordes et al. (2013) model, where a relation is interpreted as a translation from the head entity to the tail entity, with a score function, for each gold triplet in the KB. This operation is fast and the latent embeddings are expressive in many cases. Then we can reason the latent relation representation of a particular pair in vector space by , without the need for extra parameters. Here is not necessarily a real relation between two entities in the KB but just reflects the geometric property. The penalty for violating a mustlink constraint between a pair of sentences with a high KB score should be higher than those with low KB scores. This further inspires us to define a soft constraint penalty based on the similarity of latent KB relations.
Here, we use the adjusted cosine similarity
Sarwar et al. (2001) between two latent relations as a mustlink confidence score(7) 
where if otherwise 0, is a threshold we defined to control the mustlink scope, is named entity in and is its embedding. The similarity between and evaluates whether two sentences indicate similar relations according to the KB embedding.
We also define the cannotlink in a similar way, where two sentences cannot be in the same cluster with a confidence
(8) 
where if otherwise 0, and is a threshold we defined to control the cannotlink scope. We simply set .
3.2.2 Clustering Regularization
For each pair of sentences , the relation extractor will predict a clustering posterior , which can be computed based on Eq. (4). We regularize the clustering result on the probability distance between sentence pairs, using either Euclidean distance, KullbackLeibler (KL) divergence, or JensenShannon (JS) divergence. The computation of the distance or divergences can be found in Table 1.
Then the soft constraints introduced in §3.2.1 are applied on the corresponding distance to calculate the regularization terms:
(9)  
(10) 
for must and cannot links respectively, where can be , , or . Taking mustlink constraint as an example, if the posterior distributions and are different from each other but KB suggests that these two sentences should be in the same cluster where is large, then being large means there is a large cost when and being different. Then in the training phase, we want to reduce this cost given the constraint.
The constraints above are defined in a
space. It is almost impossible to enumerate all of the constraints. To make it trainable, we instead gather the constraints within a minibatch. Since in different training epochs we randomly permute the training samples, it is possible to touch many pairs of sentences in practice.
3.3 Learning
The model parameters only exist in original autoencoder components (i.e., and ), which can be jointly optimized by maximizing the following objective function with regularization:
(11) 
where , , , and are hyperparameters to control the regularization strength. can be or depending on the cosine similarity between pairs. In practice, we apply annealing method over in an exponential way:
where is the initial value, and is the final value, and are the current and total training steps respectively. This method enables the extractor to explore more possibilities first and finally converge to a stable distribution.
It is difficult to directly compute the partition function in Eq. (5), as it requires to sum over . We use the same negative sampling method as Marcheggiani and Titov (2016) to substitute in Eq. (11) with:
where is the set of randomly sampled entities in and
is the sigmoid function.
4 Experiments
In this section, we show the experimental results.
4.1 Dataset and Preprocessing
We evaluate our model in the context of unsupervised relation discovery and compare to the baseline model, DVAE Marcheggiani and Titov (2016) which is the current stateoftheart of relation discovery. Distant supervision assumes that the relations should be aligned between the KB and the training text corpus, which is not available in our setting.
We tested our model on three different subsets of New York Times corpus (NYT) Sandhaus and Evan (2008).

The first one is widely used in unsupervised settings, which was developed by Yao et al. (2011) and has also been used by Marcheggiani and Titov (2016). This dataset contains articles 2000 to 2007, with named entities annotated and features processed (POS tagging, NER, and syntactic parsing). We use this dataset to compare with previous work directly Marcheggiani and Titov (2016).

The second and third ones are usually applied by supervised models. So when they generated the data, they tended to focus on relations with more supporting sentences. The second one was developed by Zeng et al. (2017). The data is built by aligning Wikidata Vrandečić (2012) relations with NYT corpus, as a result of 99 possible relations. It is built to contain more updated facts and richer structures of relations, e.g., a larger number of relation/relation paths. We use this dataset to amplify the effects coming from relation paths in KB, as the data was used to train a pathbased relation extraction model.

The third one was developed by Riedel et al. (2010) and has also been used by Lin et al. (2016). This dataset was generated by aligning Freebase Bollacker et al. (2008) relations with NYT in 20052007, and with 52 possible relations. We use this data to test the clustering result with a narrow relation domain.
Data  NYT122  NYT71  NYT27  

Text  # sentences  67,123  14,210  87,144 
# facts  9,207  2,274  8,559  
# entity pairs  20,939  3,539  36,714  
# entities  5,865  2,489  4,803  
# relations  122  71  27  
KB  # triplets  401,490  456,146  439,507 
# entity pairs  331,008  373,875  354,960  
# entities  14,907  14,933  14,911  
# relations  705  1,009  1,031 
We align these datasets against FB15K, which is a randomly sampled subset of Freebase developed by Bordes et al. (2013). For each of the datasets above, we hold out the triplets in FB15K that contains relations in corresponding text data, so that we ensure that KB cannot give any direct supervision on any relation labels. We then discard named entities in text corpus if they are not shown in KB, so that we can directly test the influence of our KB constraint model. Finally, we only keep a single label for each sentence, and , follow the occurrence order in the sentence. The resulting datasets contain 122, 71, and 27 relation labels respectively, so we name them as NYT122, NYT71, and NYT27. The statistics of the three datasets are shown in Table 2
. For NYT71 and NYT27, we perform the same feature extraction as NYT122 shown in
Marcheggiani and Titov (2016).4.2 Implementation Details
All the model parameters are initialized randomly. The number of negative samples is set to 5, minibatch size is set to 100 with 80 epochs. We optimize all the models using AdaGrad Duchi et al. (2011) with initial learning rate at 0.5. For NYT122, we induce 40 relations clusters, with , , , and . For NYT71, we induce 30 relations clusters, with , , , and . For NYT27, we induce 20 relations clusters, with , , , and . We train TransE as our KB embedding model with 50 dimensions and 1,000 epochs.
We report the average and standard deviation based on five different runs. We randomly split the data into validation:test=4:6. All the model selections were based on validation sets, and final evaluation results will be only based on test sets.
4.3 Evaluation and Discussion
As the scoring function, we use the Bagga and Baldwin (1998) which has also been used by our baseline Marcheggiani and Titov (2016), and Normalized Mutual Information (NMI) Strehl and Ghosh (2002) metrics. Both are standard measures for evaluating clustering tasks.
Model  Metrics  

Prediction based on encoder  Prediction based on decoder  
F1  NMI  F1  NMI  
Mean  Std  Mean  Std  Mean  Std  Mean  Std  
DVAE  0.417  0.011  0.339  0.009  0.419  0.011  0.337  0.014 
RegDVAE (Euclidean at encoder)  0.469  0.014  0.430  0.020  0.448  0.020  0.384  0.020 
RegDVAE (KL at encoder)  0.375  0.009  0.359  0.014  0.380  0.011  0.355  0.014 
RegDVAE (JS at encoder)  0.435  0.038  0.370  0.042  0.409  0.012  0.336  0.005 
RegDVAE (Euclidean at decoder)  0.416  0.019  0.329  0.017  0.350  0.012  0.201  0.054 
Regularization and Prediction Strategies.
We first report our results on NYT122 using different regularization and prediction settings, as this dataset was used by our baseline model DVAE.
Note that both encoder and decoder components can make relation predictions. In fact, the way of using encoder for each sentence is straightforward. Then based on the encoder, we predict relation on the basis of single occurrence of entity pair. When using the decoder, we need to renormalize as to make predictions. Based on the decoder, we make predictions for each unique entity pair. As a consequence, our constraints can be imposed on both encoder and decoder. The way of computing decoder probability distribution is the same as making predictions. So in this experiment, we report both results.
The results are shown in Table 3. From the table, we can see that regularization with Euclidean distance performs the best compared to KL and JS. Moreover, the regularization over encoder is better than the regularization over decoder. This may be because the way that we put constraints only over sampled sentences in a batch may hurt the regularization of decoder, since sampled unique pairs may be less than sample sentences. If we look at results comparing original DVAE prediction based on the encoder and the decoder, both result in similar F1 and NMI numbers. Thus, we can only conclude that currently in the way we do sampling, constraining over encoder is a better choice.
Comparison on Different Datasets.
We also compare our algorithm on the three datasets with different baseline settings. In order to evaluate our model rigorously, besides the original DVAE model, we compare two additional augmented baseline models with the same hyperparameter setting: DVAE with TransE embeddings appended to encoder input features (DVAE+E) and DVAE with decoder entity vectors replaced by pretrained KB embeddings (DVAE+D). For our method, we report RegDVAE with the best setting where we use Euclidean distance based constraints to regularize the encoder. Moreover, we report a setting with fixed embeddings in the decoder as the ones obtained from TransE (RegDVAE+D). This also makes sense since even though the TransE embeddings are not trained with the observation of the same relations as the text corpus, the embeddings already contain much semantic information about entities. Then by fixing the embeddings of entities in the decoder, we can significantly reduce the number of parameters that need to be trained. The results are shown in Table 4. As we can see that, RegDVAE+D can outperform the original DVAE by 823 points on F1. DVAE+D is also good but may fail when there are a lot of outofsample entities in the training corpus.
Model  NYT122  NYT71  NYT27  

F1  NMI  F1  NMI  F1  NMI  
Mean  Std  Mean  Std  Mean  Std  Mean  Std  Mean  Std  Mean  Std  
Majority  0.355    0    0.121    0    0.549    0   
DVAE  0.417  0.011  0.339  0.009  0.325  0.011  0.375  0.023  0.433  0.018  0.384  0.021 
DVAE+E  0.385  0.021  0.341  0.043  0.339  0.021  0.418  0.022  0.396  0.034  0.381  0.039 
DVAE+D  0.452  0.033  0.438  0.022  0.352  0.038  0.339  0.009  0.499  0.040  0.469  0.027 
RegDVAE  0.469  0.014  0.430  0.020  0.377  0.020  0.466  0.036  0.587  0.005  0.451  0.005 
RegDVAE+D  0.499  0.022  0.497  0.013  0.432  0.028  0.589  0.071  0.665  0.022  0.562  0.038 
Hyperparameter Sensitivity.
We have three hyperparameters in our algorithm: for the regularization of encoder entropy, for the regularization with our constraints, and for the threshold of KB based cosine similarities. Here, we test and , since the sensitivity result of is the same as the original DVAE work Marcheggiani and Titov (2016). The sensitivity of is shown in Figure 2(a). The results are good in a wide range from to . The sensitivity of is shown in Figure 2(b). It reveals some interesting patterns. At the beginning when is small, it hurts the performance. After getting greater than 0.7, it improves the performance, which means that only very similar relations indicated by KB embeddings are useful relations as constraints. In addition, (meaning only finding identical relations) is worse than , which means we indeed find some relations in our KB so that different triplets will be constrained.
KB Relation Overlap.
Although we assume that there is no overlapped relation between the KB and the training text corpus, in practice, we may find a lot of applications that the relations are partially observed in KB. Thus, we also test a setting when the KB has different proportions of overlapped relations with training text corpus. In this case, we train different KB embeddings for different percentages of overlapped relations, and then apply the embeddings into the constraints. The results are shown in Figure 2(c). As we can see, in general, more overlapped relations will result in better performance. The best number can be better than the number without overlapped relation by about two points. This again verifies that the KB embedding is very robust and represent the semantic meanings of entities even with part of the relations observed Bordes et al. (2013).
Contextual Sentence  Cluster  Similarity 

…Spain will become the third country in Europe…  12  0.926 
Portugal, with all that talent, goes home to Europe…  12  
Brazil, Latin America ’s largest economy …  12  0.916 
…Argentina was perhaps the most expensive country in Latin America for tourists.…  12 
Case Study.
We also show some examples of entity pair similarities in Table 5. From the Table we can see that our target relation cluster is /location/contained_by. In the first example, the similarity between entity pairs (Spain, Europe) and (Portugal, Europe) are high, which indicates the same cluster of pairs of sentences. The same constraint is applied in the second example, although there’s no direct connection between (Brazil, Latin America), (Argentina, Latin America).
5 Related Work
Supervised and Distantly Supervised Relation Extraction.
Traditional supervised relation extraction focuses on a limited number of relations Roth and Yih (2002); Kambhatla (2004); Chan and Roth (2010). Distant supervision uses KBs to obtain a lot of automatically annotated data Mintz et al. (2009); Riedel et al. (2010); Hoffmann et al. (2011); Surdeanu et al. (2012); Xu et al. (2013a); Zhang et al. ; Zeng et al. (2015); Lin et al. (2016); Zeng et al. (2017). There are two important assumptions behind these models, namely multiinstance learning Riedel et al. (2010) and multiinstance multilabel learning Hoffmann et al. (2011); Surdeanu et al. (2012). Our setting is similar to multiinstance learning but we assume there is no overlapped relation between KB and training text corpus. Universal schema Riedel et al. (2013); Verga et al. (2016); Toutanova et al. (2015); McCallum et al. (2017) can also exploit KB to help extract relations. It needs a lot of entity pairs in text to cooccur with KB triplets, which is under the same setting with distant supervision. Those surface patterns are preextracted and shown in the training phase, which makes it also a weakly supervised learning method.
Unsupervised Relation Extraction.
Open Domain Information Extraction (OpenIE) assumes that every relation expression can represent a unique relation Etzioni et al. (2004); Banko et al. (2007); Fader et al. (2011); Mausam et al. (2012); Xu et al. (2013b); Angeli et al. (2015). On the other hand, relation clustering approaches group all the related relation expressions to represent a relation Lin and Pantel (2001); Mohamed et al. (2011); Takamatsu et al. (2011); Yao et al. (2011, 2012); Nakashole et al. (2012a, b); Marcheggiani and Titov (2016). Our setting is based on Marcheggiani and Titov (2016) but we also introduce KB as a different kind of weak and indirect supervision.
Knowledge Base Representation.
Embedding based knowledge base representation learning methods Bordes et al. (2013); Wang et al. (2014); Lin et al. (2015); Trouillon et al. (2016) represent entities and relations as vectors, denoted as and respectively such that for a distances function , the value is maximized for all facts. Among all these methods, TransE model has a favorable property that the translation operation can be easily recovered by entity vectors . With its simplicity and high performance, TransE is enough for demonstration. Though our method is not restricted to the representation form of KB, we leave it for future evaluation.
Constraints can be made more explainable by paths finding. For instance, the Path Ranking Algorithm (PRA) Lao and Cohen (2010); Lao et al. (2011)
uses random walk to perform multihop reasoning based on logic rules. Later on, reinforcement Learning
Toutanova et al. (2015); Xiong et al. (2017); Das et al. (2017); Chen et al. (2018)is used to search for paths more effectively. Though heuristics are used to further reduce the number of mined relations, it is still very costly to find the paths for KB with hundreds of relations, if not impossible.
Constraint Modeling.
Originated from semisupervised learning
Chapelle et al. (2006), mustlink and cannotlink modeling has been well studied in machine learning community
Wagstaff et al. (2001); Basu et al. (2004, 2008). Such constraints were usually generated based on the ground truth labels of data. For document clustering, word constraints constructed based on WordNet similarities have been applied Song et al. (2013) and entity constraints based on entity types in an external KB have been used Wang et al. (a, 2016), both being considered as a kind of indirect supervision based on side information. For triplet relation clustering, relation surface similarity and entity type constraints have been explored Wang et al. (b). However the above constraints are applied to a particular form of models, coclustering models. Compared to existing approaches, our constraints are constructed based on more recently developed KB embeddings, which is more flexible and easy to incorporate into different models.In natural language processing community, constraints based on background knowledge are also well studied. For example, constrained conditional models (CCM)
Chang et al. (2012)provides a very flexible framework to decouple learning and inference, where in the inference step, background knowledge can be incorporated as an ILP (integer linear programming) problem. Posterior regularization (PR)
Ganchev et al. (2010) generalizes this idea so that it uses a joint learning and inference framework to incorporate the background knowledge. Both CCM and PR have many applications including the application to relation extraction Chan and Roth (2010); Chen et al. (2011). Compared to these existing approaches, our constraints are derived from the generalpurpose KB, which is quite different from their way of manually crafting some background knowledge as declarative rules.It is very interesting that we are similar to the PR framework. Since we use a DVAE framework as the base algorithm, there is no traditional Estep and Mstep in the variational inference. Instead, only and probabilities parameterized by neural networks are updated. In our framework, we can add constraints to either or probabilities (applying to
needs modification of normalization). It is the same that we draw a biased learning process when estimating the posteriors as PR does.
6 Conclusion
In this paper, we propose a new relation discovery setting where there is no overlapped relations between the training text corpus and the KB. We propose a new learning framework of KB regularization which uses mustlink and cannotlink constraints derived based on similarities in the KB embedding space. Our method improves the results over all baseline models without harming the scalability. We believe this framework is as flexible as other constraint models to be applied to many applications when we think the semantics of entities and relations provided by the KB is useful.
Acknowledgments
This paper was supported by the Early Career Scheme (ECS, No. 26206717) from Research Grants Council in Hong Kong. We thank Intel Corporation for supporting our deep learning related research. We also thank the anonymous reviewers for their valuable comments and suggestions that help improve the quality of this manuscript.
References
 Angeli et al. (2015) Gabor Angeli, Melvin Jose Johnson Premkumar, and Christopher D. Manning. 2015. Leveraging linguistic structure for open domain information extraction. In ACL, pages 344–354.
 Bagga and Baldwin (1998) Amit Bagga and Breck Baldwin. 1998. Algorithms for scoring coreference chains. In LREC, pages 563–566.
 Banko et al. (2007) Michele Banko, Michael J. Cafarella, Stephen Soderland, Matthew Broadhead, and Oren Etzioni. 2007. Open information extraction from the web. In IJCAI, pages 2670–2676.
 Basu et al. (2004) Sugato Basu, Mikhail Bilenko, and Raymond J. Mooney. 2004. A probabilistic framework for semisupervised clustering. In KDD, pages 59–68.
 Basu et al. (2008) Sugato Basu, Ian Davidson, and Kiri Wagstaff. 2008. Constrained Clustering: Advances in Algorithms, Theory, and Applications. Chapman & Hall/CRC.
 Bollacker et al. (2008) Kurt D. Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor. 2008. Freebase: a collaboratively created graph database for structuring human knowledge. In SIGMOD, pages 1247–1250.
 Bordes et al. (2013) Antoine Bordes, Nicolas Usunier, Alberto GarcíaDurán, Jason Weston, and Oksana Yakhnenko. 2013. Translating embeddings for modeling multirelational data. In NIPS, pages 2787–2795.
 Chan and Roth (2010) Yee Seng Chan and Dan Roth. 2010. Exploiting background knowledge for relation extraction. In COLING, pages 152–160.
 Chang et al. (2012) MingWei Chang, LevArie Ratinov, and Dan Roth. 2012. Structured learning with constrained conditional models. Machine Learning, 88(3):399–431.
 Chapelle et al. (2006) O. Chapelle, B. Schölkopf, and A. Zien, editors. 2006. SemiSupervised Learning. MIT Press.
 Chen et al. (2011) Harr Chen, Edward Benson, Tahira Naseem, and Regina Barzilay. 2011. Indomain relation discovery with metaconstraints via posterior regularization. In ACLHLT, pages 530–540.
 Chen et al. (2018) Wenhu Chen, Wenhan Xiong, Xifeng Yan, and William Yang Wang. 2018. Variational knowledge graph reasoning. In NAACLHLT, pages 1823–1832.
 Das et al. (2017) Rajarshi Das, Shehzaad Dhuliawala, Manzil Zaheer, Luke Vilnis, Ishan Durugkar, Akshay Krishnamurthy, Alexander J. Smola, and Andrew McCallum. 2017. Go for a walk and arrive at the answer: Reasoning over paths in knowledge bases using reinforcement learning. CoRR, abs/1711.05851.
 Dong et al. (2014) Xin Dong, Evgeniy Gabrilovich, Geremy Heitz, Wilko Horn, Ni Lao, Kevin Murphy, Thomas Strohmann, Shaohua Sun, and Wei Zhang. 2014. Knowledge vault: a webscale approach to probabilistic knowledge fusion. In KDD, pages 601–610.
 Duchi et al. (2011) John C. Duchi, Elad Hazan, and Yoram Singer. 2011. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12:2121–2159.
 Etzioni et al. (2004) Oren Etzioni, Michael Cafarella, and Doug Downey. 2004. Webscale information extraction in knowitall (preliminary results). In WWW, pages 100–110.
 Fader et al. (2011) Anthony Fader, Stephen Soderland, and Oren Etzioni. 2011. Identifying relations for open information extraction. In EMNLP, pages 1535–1545.
 Ganchev et al. (2010) Kuzman Ganchev, Joao Graça, Jennifer Gillenwater, and Ben Taskar. 2010. Posterior regularization for structured latent variable models. Journal of Machine Learning Research, 11:2001–2049.
 Gillenwater et al. (2011) Jennifer Gillenwater, Kuzman Ganchev, João Graça, Fernando Pereira, and Ben Taskar. 2011. Posterior sparsity in unsupervised dependency parsing. Journal of Machine Learning Research, 12:455–490.
 Hoffmann et al. (2011) Raphael Hoffmann, Congle Zhang, Xiao Ling, Luke S. Zettlemoyer, and Daniel S. Weld. 2011. Knowledgebased weak supervision for information extraction of overlapping relations. In ACL, pages 541–550.
 Kambhatla (2004) Nanda Kambhatla. 2004. Combining lexical, syntactic, and semantic features with maximum entropy models for information extraction. In ACL  Poster and Demonstration.
 Lao and Cohen (2010) Ni Lao and William W. Cohen. 2010. Relational retrieval using a combination of pathconstrained random walks. Machine Learning, 81(1):53–67.
 Lao et al. (2011) Ni Lao, Tom M. Mitchell, and William W. Cohen. 2011. Random walk inference and learning in A large scale knowledge base. In EMNLP, pages 529–539.
 Lin and Pantel (2001) Dekang Lin and Patrick Pantel. 2001. DIRT – discovery of inference rules from text. In KDD, pages 323–328.
 Lin et al. (2015) Yankai Lin, Zhiyuan Liu, HuanBo Luan, Maosong Sun, Siwei Rao, and Song Liu. 2015. Modeling relation paths for representation learning of knowledge bases. In EMNLP, pages 705–714.
 Lin et al. (2016) Yankai Lin, Shiqi Shen, Zhiyuan Liu, Huanbo Luan, and Maosong Sun. 2016. Neural relation extraction with selective attention over instances. In ACL, pages 2124–2133.
 Liu et al. (2014) Xitong Liu, Fei Chen, Hui Fang, and Min Wang. 2014. Exploiting entity relationship for query expansion in enterprise search. Inf. Retr., 17(3):265–294.
 Marcheggiani and Titov (2016) Diego Marcheggiani and Ivan Titov. 2016. Discretestate variational autoencoders for joint discovery and factorization of relations. Transactions of the Association for Computational Linguistic, 4:231–244.
 Mausam et al. (2012) Mausam, Michael Schmitz, Stephen Soderland, Robert Bart, and Oren Etzioni. 2012. Open language learning for information extraction. In EMNLPCoNLL, pages 523–534.
 McCallum et al. (2017) Andrew McCallum, Arvind Neelakantan, and Patrick Verga. 2017. Generalizing to unseen entities and entity pairs with rowless universal schema. In EACL, pages 613–622.
 Mintz et al. (2009) Mike Mintz, Steven Bills, Rion Snow, and Dan Jurafsky. 2009. Distant supervision for relation extraction without labeled data. In ACL/AFNLP, pages 1003–1011.
 Mohamed et al. (2011) Thahir Mohamed, Estevam R. Hruschka Jr., and Tom M. Mitchell. 2011. Discovering relations between noun categories. In EMNLP, pages 1447–1455.
 Nakashole et al. (2012a) Ndapandula Nakashole, Gerhard Weikum, and Fabian M. Suchanek. 2012a. Discovering and exploring relations on the web. PVLDB, 5(12):1982–1985.
 Nakashole et al. (2012b) Ndapandula Nakashole, Gerhard Weikum, and Fabian M. Suchanek. 2012b. PATTY: A taxonomy of relational patterns with semantic types. In EMNLP, pages 1135–1145.
 Ravichandran and Hovy (2002) Deepak Ravichandran and Eduard H. Hovy. 2002. Learning surface text patterns for a question answering system. In ACL, pages 41–47.
 Riedel et al. (2010) Sebastian Riedel, Limin Yao, and Andrew McCallum. 2010. Modeling relations and their mentions without labeled text. In ECML PKDD, pages 148–163.
 Riedel et al. (2013) Sebastian Riedel, Limin Yao, Andrew McCallum, and Benjamin M. Marlin. 2013. Relation extraction with matrix factorization and universal schemas. In NAACLHLT, pages 74–84.
 Roth and Yih (2002) Dan Roth and Wentau Yih. 2002. Probabilistic reasoning for entity & relation recognition. In COLING.
 Sandhaus and Evan (2008) Sandhaus and Evan. 2008. The new york times annotated corpus. Linguistic Data Consortium, Philadelphia.
 Sarwar et al. (2001) Badrul Sarwar, George Karypis, Joseph Konstan, and John Riedl. 2001. Itembased collaborative filtering recommendation algorithms. In WWW, pages 285–295.
 Song et al. (2013) Yangqiu Song, Shimei Pan, Shixia Liu, Furu Wei, Michelle X. Zhou, and Weihong Qian. 2013. Constrained text coclustering with supervised and unsupervised constraints. IEEE Trans. Knowl. Data Eng., 25(6):1227–1239.
 Strehl and Ghosh (2002) Alexander Strehl and Joydeep Ghosh. 2002. Cluster ensembles — A knowledge reuse framework for combining multiple partitions. Journal of Machine Learning Research, 3:583–617.
 Surdeanu et al. (2012) Mihai Surdeanu, Julie Tibshirani, Ramesh Nallapati, and Christopher D. Manning. 2012. Multiinstance multilabel learning for relation extraction. In EMNLPCoNLL, pages 455–465.
 Takamatsu et al. (2011) Shingo Takamatsu, Issei Sato, and Hiroshi Nakagawa. 2011. Probabilistic matrix factorization leveraging contexts for unsupervised relation extraction. In PAKDD, pages 87–99.
 Toutanova et al. (2015) Kristina Toutanova, Danqi Chen, Patrick Pantel, Hoifung Poon, Pallavi Choudhury, and Michael Gamon. 2015. Representing text for joint embedding of text and knowledge bases. In EMNLP, pages 1499–1509.
 Trouillon et al. (2016) Théo Trouillon, Johannes Welbl, Sebastian Riedel, Éric Gaussier, and Guillaume Bouchard. 2016. Complex embeddings for simple link prediction. In ICML, pages 2071–2080.
 Verga et al. (2016) Patrick Verga, David Belanger, Emma Strubell, Benjamin Roth, and Andrew McCallum. 2016. Multilingual relation extraction using compositional universal schema. In NAACLHLT, pages 886–896.
 Vrandečić (2012) Denny Vrandečić. 2012. Wikidata: A new platform for collaborative data collection. In WWW, pages 1063–1064.

Wagstaff et al. (2001)
Kiri Wagstaff, Claire Cardie, Seth Rogers, and Stefan Schrödl. 2001.
Constrained kmeans clustering with background knowledge.
In ICML, pages 577–584.  Wang et al. (a) Chenguang Wang, Yangqiu Song, Ahmed ElKishky, Dan Roth, Ming Zhang, and Jiawei Han. a. Incorporating world knowledge to document clustering via heterogeneous information networks. In KDD, pages 1215–1224.
 Wang et al. (b) Chenguang Wang, Yangqiu Song, Dan Roth, Chi Wang, Jiawei Han, Heng Ji, and Ming Zhang. b. Constrained informationtheoretic tripartite graph clustering to identify semantically similar relations. In IJCAI, pages 3882–3889.
 Wang et al. (2016) Chenguang Wang, Yangqiu Song, Dan Roth, Ming Zhang, and Jiawei Han. 2016. World knowledge as indirect supervision for document clustering. ACM Transactions on Knowledge Discovery from Data, 11(2):13:1–13:36.
 Wang et al. (2014) Zhen Wang, Jianwen Zhang, Jianlin Feng, and Zheng Chen. 2014. Knowledge graph embedding by translating on hyperplanes. In AAAI, pages 1112–1119.
 Xiong et al. (2017) Wenhan Xiong, Thien Hoang, and William Yang Wang. 2017. Deeppath: A reinforcement learning method for knowledge graph reasoning. In EMNLP, pages 564–573.
 Xu et al. (2013a) Wei Xu, Raphael Hoffmann, Le Zhao, and Ralph Grishman. 2013a. Filling knowledge base gaps for distant supervision of relation extraction. In ACL, pages 665–670.
 Xu et al. (2013b) Ying Xu, MiYoung Kim, Kevin Quinn, Randy Goebel, and Denilson Barbosa. 2013b. Open information extraction with tree kernels. In NAACLHLT, pages 868–877.
 Yao et al. (2011) Limin Yao, Aria Haghighi, Sebastian Riedel, and Andrew McCallum. 2011. Structured relation discovery using generative models. In EMNLP, pages 1456–1466.
 Yao et al. (2012) Limin Yao, Sebastian Riedel, and Andrew McCallum. 2012. Unsupervised relation discovery with sense disambiguation. In ACL, pages 712–720.

Zeng et al. (2015)
Daojian Zeng, Kang Liu, Yubo Chen, and Jun Zhao. 2015.
Distant supervision for relation extraction via piecewise convolutional neural networks.
In EMNLP, pages 1753–1762.  Zeng et al. (2017) Wenyuan Zeng, Yankai Lin, Zhiyuan Liu, and Maosong Sun. 2017. Incorporating relation paths in neural relation extraction. In EMNLP, pages 1768–1777.
 (61) Xingxing Zhang, Jianwen Zhang, Junyu Zeng, Jun Yan, Zheng Chen, and Zhifang Sui. Towards accurate distant supervision for relational facts extraction. In ACL (2), pages 810–815.
Comments
There are no comments yet.