【研究綜述】Relation extraction Summary（持續更新）

update time: 11-5

Relation definition:

A relation is defined in the form of a tuple $t = (e1,e2,...,en)$ where the $e_i$ are entities in a predefined relation $r$ within document $D$.

從這裏可以看出其實relation 的定義是十分廣泛的，只要在document中就可以了，對entity的數量，長度都沒有具體的要求，（不過這個觀點僅從一篇論文A Review of Relation Extraction 中得出，如果這樣的relation跨越了整個document，根據其他論文的觀點很可能會命名爲event extraction/event detection.

Relation extraction type:

number of entity:

• binary relation:two entities with one relation

• e.g.located-in(CMU, Pittsburgh)

• higher order relation: more than 2 entities and one relation.

• e.g.
  Sentence: At codons 12, the occurence of point mutations from G to T were observed
  Relation: point mutation(codon, 12, G, T)4 entities
  Paper 2005:Simple algorithms for complex relation extraction with applications to biomedical ie

range of extraction:

• sentence-level

• 基本上非常多的論文都是關於sentence-level的，因爲數據集最多，問題最簡單，方法可操作性多。主要集中在方法的改進上

• document-level (to be revised)

• 主要是基於數據集的改進, 對針對sentence-level的方法來說是很大的挑戰。

Sentence level relation extraction main stream methods

在深度學習盛行之前，根據其他論文的related work，可以將
總的分成feature-based和kernel based （這段時間mainly focus on deep learning method)
所以不提供paper notes和brief description了，僅提供link。

deep learning methods:

• 1. RNN based method

• 1. LSTM 1997
• 2. LSTM-based Context Aware Encoder 2017

• Paper:Context-Aware Representations for Knowledge Base Relation Extraction
• Paper Notes

• 2. CNN based methods

• 1. CNN with position features & lexical features & sentence features. 2014

Paper: Relation Classification via Convolutional Deep Neural Network
Paper Notes

• 2. Piece-wise CNN with Multi-instance Learning 2015

Paper: Distant Supervision for Relation Extraction via Piecewise Convolutional Neural Networks
Paper Notes

Feature based method:

Feature to extract

• (1) the entities themselves
• (2) the types of the two entities
• (3) word sequence between the entities
• (4) number of words between the entities
• (5) path in the parse tree containing the two entities.

typical methods

• log-linear model:for entity classification:

paper 2004 :Combining lexical, syntactic, and semantic features with maximum entropy models for extracting relations

• SVMs using polynomial and linear kernels

paper 2002:Exploring various knowledge in relation extraction
paper 2005:Extracting relations with integrated information using kernel methods

Kernel-based method:

• single kernel & string kernels

Paper 2002:Text classification using string kernels

• Bag of features Kernel ( 3 subkernels defined)

Paper 2005:Subsequence kernels for relation extraction

• Tree Kernels

Paper 2003: Kernel methods for relation extraction
Paper 2004: Dependency tree kernels for relation extraction.

Relation Extraction Learning Paradigm

distant supervision

不用標註的數據，只需要通過一些 general domain的relation數據集，通過數據集的supervise來進行訓練和測試。

Paper 2009: Distant supervision for relation extraction without labeled data
Paper Notes

Paper: Distant Supervision for Relation Extraction with Sentence-Level Attention and Entity Descriptions

semi-supervision

• Bootstrap method
• DIPRE(Dual Iterative Pattern Relation Expansion)

Paper 1998:Extracting patterns and relations from the world wide web.

• Snowball

Paper 2000:Snowball: Extracting relations from large plain-text collections

• KnowItAll

Paper 2005:Unsupervised Named-Entity Extraction from the Web: An Experimental Study

• TextRunner

Paper 2007:Open information extraction from the web.

supervision

上述的kernel-based & feature based

• without distant supervision entitled deep learning method papers.

unsupervision(to be revised)

這個不在研究範圍，暫時不寫了

References

• A Review of Relation Extraction