Nanyun Peng, et.al., “Cross-Sentence N-ary Relation Extraction with Graph LSTMs”, TACL 2017
目录
展开
1 Intro
- 姓名：Graph LSTM
- 机构：JHU, MSR, Google
- 任务：cross-sentence N-ary RE
- 流派：lstm
- 方法：Graph LSTM 就是把计算 LSTM 时所有前驱直接求和
- 性能：一顿操作猛如虎，最后和 Bi-LSTM 性能差不多
2 task definition
To extract the relation within a {tumor, mutation, drug} triplet. Furthermore, a sub-relation task is difined as extracting the relations between the two of a triplet.
3 methods
- Build a graph by connecting the root of dependency trees(but it’s not clear whether the sentence-roots are fully connected or just connect the adjacent roots)
- Bi-directional Graph LSTM: partition the document graph into two DAGs, one → and one ←. (still not clear how to partition in cross-sentence scenarios, I don’t think there is a unique solution for the document graph(more thoughts are needed))
- Graph LSTM Structure:
The left is vanilla LSTM, the middle is “Full Parameterization” Graph LSTM, which is simply sum all the predecessors when calculating 3 gates, the right is “Edge-Type embedding” Graph LSTM, which simplifies the “predecessor representation” into “predecessors’ edge-type representation”.
- Multi-task learning with sub-relations