First, the innovations:
Given the many ways in relation extraction before the label only a one-hoe vector that relations between independent. This paper argues that relation is by the association. Thus, we propose a model RELE ( R & lt Elation E xtraction with Joint L Abel E mbedding), the label for embedding that, for the classification task. At the same time the use of structural information from KGs and textual information of entity descriptions to label study
Second, the related work:
This work related RE were good summary
1.Distant Supervision
(Mintz et al., 2009) proposed remote monitoring is an effective method, it is assumed that there is some relationship between two entities in a KG, all references to these entities are sentences expressing this relationship. This assumption does not hold true in all cases, resulting in the wrong label asked
question.
MultiR (Hoffmann et al., 2011 ) and MIMLRE (Surdeanu et Al., 2012) study the introduction of multi-instance, which refers to the same entity instances processed at the package level. However, these methods rely heavily on manual characteristics.
（Raphael Hoffmann, Congle Zhang, Xiao Ling, Luke Zettlemoyer, and Daniel S Weld. 2011. Knowledgebased weak supervision for information extraction of overlapping relations. In ACL, pages 541–550.）
（Mihai Surdeanu, Julie Tibshirani, Ramesh Nallapati, and Christopher D Manning. 2012. Multi-instance multi-label learning for relation extraction. In EMNLP-CoNLL, pages 455–465.）
2.Neural Relation Extraction
In recent years, with the development of deep learning, neural networks have been proven effective features can effectively automate the extraction of the sentence.
Multi-instance learning, the introduction of attention. . .
There are also some studies use other relevant information to improve relations extract
(Wenyuan Zeng, Yankai Lin, Zhiyuan Liu, and Maosong Sun. 2017. Incorporating relation paths in neural relation extraction. In EMNLP, pages 1768- 1777.)
(Shikhar Vashishth, Rishabh Joshi, Sai Suman Prayaga, Chiranjib Bhattacharyya, and Partha Talukdar. 2018. Reside: Improving distantly-supervised neural relation extraction using side information. In
EMNLP, pages 1257–1266.
RESIDE proposed model. Knowledge of the use-side available information, including an entity type information and the alias)
(Xu Han, Pengfei Yu, Zhiyuan Liu, Maosong Sun, and Peng Li. 2018b. Hierarchical relation extraction with coarse-to-fine grained attention. In EMNLP, pages 2236–2245
Proposed joint showing a learning framework and KGs example, embedded in the frame using the KG selected valid instance)
（Guoliang Ji, Kang Liu, Shizhu He, Jun Zhao, et al. 2017. Distant supervision for relation extraction with sentence-level attention and entity descriptions. In AAAI, pages 3060–3066.
Proposed APCNN + D model. Entity is described as an example of an effective choice of background knowledge, ignoring the imposed noise)
Comparison of existing work and the work of the author:
Third, some definitions:
1. We expressed by G = KG {(h, r, t )}, which contains a considerable number of groups of three (h, r, t) where h and t are the end entities and entity head, r is the relationship . Embedding thereof is represented as ( H , R & lt , T ).
2. In a given entity (h, t) and Examples bag (sentences) B = {s1, s2, ···, sm}, wherein si comprises each instance (h, t), the relationship between the extracted task is to train a classifier based on B from a predefined relationship between the concentration of prediction (h, t) relationship label y. If there is no relationship, we simply assign it to the NA.
3. 
, 
. D is a collection, each of the set of elements is described a enitty. Description is to find the first paragraph from Wikipedia. L is the length of this first segment. 
. V is the vocabulary.
Fourth, the model description:
1.Joint Label Embedding
KG Embedding：
Use TransE to KG embedding,
Entity Description Embedding:
Wherein, c is the convolution window length, 
corresponding to the output of the word Wi.
Gating Integration：
Wherein 
an integrated gated vector, e is the KG entity embedding (head, tail), de embedding are described.
Well, the last label is the type of embeddign.
Label Classifier：
2. Neural Relation Extraction
Instance Embedding：
Each word is embedded by the embedded pre-training and two positions.
其中，c是卷积的窗口大小。之后在对每个z的j的位置取最大，得到最后句子的嵌入。
把句子和两个entity的嵌入链接。
Attention over Instances.：
Relation Classifier.：
3. Model Training:
假设，在训练阶段，包一共有N个{B1, B2，··，BN}及其对应的标签{y1, y2，···，yN}，我们利用交叉熵对标签分类器L1的损失函数进行处理
D是描述，G是知识图谱。这个用来训练label embedding的分类器的
B是bag中的句子，这个用来训练关系的分类器
五。实验：
