Deep Neural Architectures for End-to-End Relation Extraction

The rapid pace of scientific and technological advancements has led to a meteoric growth in knowledge, as evidenced by a sharp increase in the number of scholarly publications in recent years. PubMed, for example, archives more than 30 million biomedical articles across various domains and covers a wide range of topics including medicine, pharmacy, biology, and healthcare. Social media and digital journalism have similarly experienced their own accelerated growth in the age of big data. Hence, there is a compelling need for ways to organize and distill the vast, fragmented body of information (often unstructured in the form of natural human language) so that it can be assimilated, reasoned about, and ultimately harnessed. Relation extraction is an important natural language task toward that end. In relation extraction, semantic relationships are extracted from natural human language in the form of (subject, object, predicate) triples such that subject and object are mentions of discrete concepts and predicate indicates the type of relation between them. The difficulty of relation extraction becomes clear when we consider the myriad of ways the same relation can be expressed in natural language. Much of the current works in relation extraction assume that entities are known at extraction time, thus treating entity recognition as an entirely separate and independent task. However, recent studies have shown that entity recognition and relation extraction, when modeled together as interdependent tasks, can lead to overall improvements in extraction accuracy. When modeled in such a manner, the task is referred to as end-to-end relation extraction. In this work, we present four studies that introduce incrementally sophisticated architectures designed to tackle the task of end-to-end relation extraction. In the first study, we present a pipeline approach for extracting protein-protein interactions as affected by particular mutations. The pipeline system makes use of recurrent neural networks for protein detection, lexicons for gene normalization, and convolutional neural networks for relation extraction. In the second study, we show that a multi-task learning framework, with parameter sharing, can achieve state-of-the-art results for drug-drug interaction extraction. At its core, the model uses graph convolutions, with a novel attention-gating mechanism, over dependency parse trees. In the third study, we present a more efficient and general-purpose end-to-end neural architecture designed around the idea of the table-filling paradigm; for an input sentence of length n, all entities and relations are extracted in a single pass of the network in an indirect fashion by populating the cells of a corresponding n by n table using metric-based features. We show that this approach excels in both the general English and biomedical domains with extraction times that are up to an order of magnitude faster compared to the prior best. In the fourth and last study, we present an architecture for relation extraction that, in addition to being end-to-end, is able to handle cross-sentence and N-ary relations. Overall, our work contributes to the advancement of modern information extraction by exploring end-to-end solutions that are fast, accurate, and generalizable to many high-value domains