Cross-Sentence N-ary Relation Extraction with Graph LSTMs

Abstract

This research was conducted when the authors were at Microsoft Research.

Text-mining clinically relevant cancer biomarkers for curation into the CIViC database

“…An alternative approach to this problem would have been to use a distant supervision method using the CIViC knowledgebase as seed data. This approach was taken by Peng et al, who also attempted to extract relations across sentence boundaries [44]. They chose to focus only on point mutations and extracted 530 within sentence biomarkers and 1,461 cross-sentence biomarkers.…”

DeepDrawing: A Deep Learning Approach to Graph Drawing

“…The key challenge of non-spectral approaches is how to define the neighborhood of a node as the receptive field and various methods have been proposed, including adaptive weight matrices [21], uniformly sampling [36], and transition matrices [1]. A closely related research direction explores using RNNs for graph-structured data [57,65,70,82]. For example, Li et al [57] modified the Gate Recurrent Units (GRU) and proposed a gated GNN to learn node representations.…”

“…Tai et al [70] proposed two types of tree-LSTM, generalizing the basic LSTM to tree-structure typologies, to predict the semantic relatedness of sentences. Peng et al [65] extended tree-LSTM by distinguishing different edge types in the graph and applied the model to the relation extraction problem in the Natural Language Processing (NLP) field. You et al [80] developed an RNN-based method for modeling complex distributions over multiple graphs and further generating graphs.…”

“…When using a general LSTM model, such kind of dependency information can still be weakened or lost, especially for the LSTM cells that are far from each other. Inspired by the recent work in natural language processing field [65,70], we propose adding direct connections between different LSTM cells to explicitly model the topological structure of input networks. Such direct connections are termed "skip connections" in the deep learning field [32].…”

“…Although our model architecture is similar to prior studies [65,70], it targets at different problems. Unlike natural language processing (NLP) problems, where the input text is already sequential data and the input feature vector can be directly gained through word embedding, it is necessary to carefully design the architecture and input feature vector to model the graph topology information when using deep learning for graph drawing.…”