Relation extraction via one-shot dependency parsing on intersentential, higher-order, and nested relations

Abstract

Despite the emergence of digitalization, people still interact with institutions via traditional means such as submitting free formatted petitions, orders, or applications. These noisy documents generally consist of complex relations that are nested, higher-order, and intersentential. Most of the current approaches address extraction of only sentence-level and binary relations from grammatically correct text and generally require high-level linguistic features coming from preprocessors such as a parts-of-speech tagger, chunker, or syntactic parser. In this article, we focus on extracting complex relations in order to automate the task of understanding user intentions. We propose a novel language-agnostic and noise-immune approach that does not require preprocessing of input text. Unlike previous literature that uses dependency parsing outputs as input features, we formulate the relation extraction task directly as a one-shot dependency parsing problem. The presented method was evaluated using a representative dataset from the banking domain and obtained 91.84% labeled attachment score (LAS), which provides an improvement of 42.85 percentage points over a rule-based baseline.