Automatic Open Knowledge Acquisition via Long Short Term Memory Networks with Feedback Negative Sampling

Abstract

There are two categories of approaches on Open IE. Studies in the first category heuristically defined extraction rule or pattern to find relationships that are not specified in advance. These approaches rely heavily on human knowledge and fail to identify relationships that are represented by more complex patterns. The second category includes studies that automatically learn extraction patterns from huge corpus. They gather dependency paths connecting relation word and arguments and make them as templates which are used to find relationships in extraction time. But the extraction process is just a pattern matching in if-else manner. They are not able to effectively make use of features in the path. In this paper, we solve these problems by applying LSTM network on the Open IE task. Bidirectional LSTM automatically extract features from paths connecting relation word to arguments and use them to decide whether there are relationships or not. To train the network, we automatically construct training set without any human labeling. Augmenting training set from a large number of seeds and feedback negative sampling enable to cope with various sentence structures. Test result on 100 randomly sampled sentences verifies that our approach extracts highly precise and abundant relationships without loss of precision.