【Paper Summary】 BRCNN for Relation Classification

BRCNN for Relation Classification is a paper published on the ACL Journal on 2016. It proposed a novel system Bidirectional Recurrent Convolutional Neural Network (BRCNN) to classify the relation of two entities in a sentence.

Read the original paper: https://www.aclweb.org/anthology/P16-1072.pdf

Key points:

1. overall INPUT: a sentence.
2. model INPUT: Shortest Dependency Path (SDP) including words and dependency relations.
3. Two-channel LSTM: one for words, another for dependency relations. They do not interact with each other during recurrent propagation. Used word embeddings, while the embedding of dependency relations are initialized randomly.
4. BRCNN: The recurrent neural network is a biased model, where later inputs are more dominant than earlier inputs. The issue is tackled by using a bidirectional model.
5. The two-channel LSTM is utilized to capture the global pattern in SDP; meanwhile, the convolution layer is utilized to capture local features.
6. Assumption: If we inverse the SDP, it corresponds the same relationship with an opposite direction. (SDP is a symmetrical structure.)

Thoughts:

1. How to determine and derive SDP?
2. How to embed dependency relations? lexical position of sentences?

Further Reading:

1. LSTM
2. NER
3. Word Embedding

Reference:

Cai, Rui, et al. “Bidirectional Recurrent Convolutional Neural Network for Relation Classification.” Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Association for Computational Linguistics, 2016, pp. 756–765. ACLWeb, doi:10.18653/v1/P16–1072.