Simple Large-scale Relation Extraction from Unstructured Text

Simple Large-scale Relation Extraction from Unstructured Text Christos Christodoulopoulos, Arpit Mittal Amazon Research Cambridge {chrchrs, mitarpit}@amazon.co.uk Abstract Knowledge-based question answering relies on the availability of facts, the majority of which cannot be found in structured sources (e.g. Wikipedia info-boxes, Wikidata). One of the major components of extracting facts from unstructured text is Relation Extraction (RE). In this paper we propose a novel method for creating distant (weak) supervision labels for training a large-scale RE system. We also provide new evidence about the effectiveness of neural network approaches by decoupling the model architecture from the feature design of a state-of-the-art neural network system. Surprisingly, a much simpler classifier trained on similar features performs on par with the highly complex neural network system (at 75x reduction to the training time), suggesting that the features are a bigger contributor to the final performance. Keywords: relation extraction, distant supervision, unstructured text 1. Introduction Knowledge-based question answering relies on the avail- ability of facts – usually in the form of triples, stored in large-scale knowledge bases (KBs) e.g. Freebase (Bol- lacker et al., 2008), DBPedia (Auer et al., 2007). There are two main sources of facts for such a KB: structured data (e.g. Wikipedia info-boxes, Wikidata) or unstructured text. Undeniably, the former type of knowledge extraction is very accurate and has been the main source of knowledge behind the major industrial knowledge bases. However, the facts extracted from structured sources cover a limited set of high-importance relations, leaving a large number of them implicitly (or explicitly) mentioned in unstructured text (McCallum, 2005). In order to ground the following presentation, we will present a typical problem from the factual knowledge ex- traction domain with the following unstructured text from a Wikipedia page: “Carrie Fisher wrote several semi- autobiographical novels, including Postcards from the Edge.” The purpose of a fact extraction system is to extract the fol- lowing facts of the form of predicate (subject, object): instance of (postcards from the edge, novel), and au- thor of (postcards from the edge, carrie fisher), where the first part is a relation, and the other parts are the left and right entities participating in that relation. Typically three tasks are involved in generating facts: En- tity Recognition, Entity Resolution (or Entity Linking), and Relation Extraction (RE). Entity Recognition and Resolu- tion deal with the task of translating surface strings to KB entities. This includes nominal or pronominal coreference resolution: we should be able to extract the same entity even if the text stated that ‘Fisher wrote...’ (instead of resolving e.g. to Bobby Fisher) or ‘She wrote... ’ (pro- vided that Carrie Fisher’s name was mentioned in a previ- ous sentence). Relation Extraction extracts relation triples (or facts) involving those entities with appropriate relations (also part of the KB schema). Each of these components could be built and operated in isolation, but they affect the performance of each other. In this paper, we examine the task of RE focusing on ex- tracting knowledge to enrich a large-scale KB (∼billions of …

Content truncated to 500 words