论文: https://www.aclweb.org/anthology/C14-1220 或者 https://pan.baidu.com/s/1p8hNvfFh4cxEwp26BEsduw
机构: 中科院
数据: SemEval2010-Task8
结果: F1值,82.7%
备注: 利用了预训练好的词向量表示,以及WordNet
输入: 一个包含两个标注实体的句子,例如 [People] have been moving back into [downtown]
输出: 一个向量,各维度代表各个关系的概率
过程:
-
Word Representation :
直接利用已有公开的词向量资源,将词转化为词向量
原文:- each input word token is transformed into a vector by looking up word
embeddings. - directly utilize the trained embeddings
- each input word token is transformed into a vector by looking up word
-
Lexical Level Features:
以下这些词的词向量表示,拼接在一起组成词汇级别的特征
- Noun 1
- Noun 2
- Left and right tokens of noun 1
- Left and right tokens of noun 2
- WordNet hypernyms of nouns
-
Sentence Level Features:
句子中的每一个词,对应一个 Word Features(WF) 和一个Position Features(PF)
例如:句子为 [ P e o p l e ] 0 [People]_0 [People]0 h a v e 1 have_1 have1 b e e n 2 been_2 been2 m o v i n g 3 moving_3 moving3 b a c k 4 back_4 back4 i n t o 5 into_5 into5 [ d o w n t o w n ] 6 [downtown]_6 [downtown]6
- WF
moving 这个词,index为3,对应的词向量为 x 3 x_3 x3
若设置窗口大小(Window Size)为3
则WF为[ x 2 x_2 x2, x 3 x_3 x3, x 4 x_4 x4] - PF [论文创新点]
moving 这个词,距离[People] (实体1)的距离为3,距离[downtown] (实体2)的距离为-3
距离值通过查询distant embeddings得到距离值对应的向量
(distant embeddings随机初始化,在训练过程中更新)
d 1 d_1 d1为当前词距离实体1的距离值对应的向量, d 2 d_2 d2为当前词距离实体2的距离值对应的向量
则PF为 [ d 1 d_1 d1, d 2 d_2 d2]
然后利用卷积神经网络,提取特征,即为句子级别的特征
- WF
-
Output:
词汇级别的特征和句子级别的特征,拼接在一起,得到一个向量
通过softmax classifier得到最终结果
BibTeX:
@inproceedings{zeng-etal-2014-relation,
title = "Relation Classification via Convolutional Deep Neural Network",
author = "Zeng, Daojian and
Liu, Kang and
Lai, Siwei and
Zhou, Guangyou and
Zhao, Jun",
booktitle = "Proceedings of {COLING} 2014, the 25th International Conference on Computational Linguistics: Technical Papers",
month = aug,
year = "2014",
address = "Dublin, Ireland",
publisher = "Dublin City University and Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/C14-1220",
pages = "2335--2344",
}