Integrating Lexical Information into Entity Neighbourhood Representations for Relation Prediction

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3429–3436 June 6–11, 2021. ©2021 Association for Computational Linguistics 3429 Integrating Lexical Information into Entity Neighbourhood Representations for Relation Prediction Ian David Wood Macquarie University and CSIRO, Australia ian.wood@mq.edu.au Stephen Wan CSIRO, Australia Stephen.Wan@ data61.csiro.au Mark Johnson Oracle Digital Assistant mark.mj.johnson@ oracle.com Abstract Relation prediction informed from a combina- tion of text corpora and curated knowledge bases, combining knowledge graph comple- tion with relation extraction, is a relatively little studied task. A system that can per- form this task has the ability to extend an ar- bitrary set of relational database tables with information extracted from a document cor- pus. OpenKi (Zhang et al., 2019) addresses this task through extraction of named entities and predicates via OpenIE tools then learning relation embeddings from the resulting entity- relation graph for relation prediction, outper- forming previous approaches. We present an extension of OpenKi that incorporates embed- dings of text-based representations of the en- tities and the relations. We demonstrate that this results in a substantial performance in- crease over a system without this information. https://github.com/drevicko/OpenKI 1 Introduction Curated knowledge repositories such as knowledge bases and relational databases provide powerful tools for many practical knowledge related tasks. They require, however, substantial effort to cre- ate and maintain. Many applications deal with knowledge that is continuously changing, present- ing prohibitive maintenance costs and limiting the utility of explicit knowledge representation tech- nologies. The new knowledge is often available in text based formats such as reports, news items and memos. In this work, we use the term “proposition” to describe a triple (e1, r, e2) that indicates that a relation r holds between two entities e1 and e2. Work in the field has largely focussed on either extracting propositions directly from text or infer- ring missing propositions by examining knowledge graphs. What we are interested in here combines the two in a single model, utilising information from the knowledge base and collections of text together to infer relations, both mentioned in the text and implied by the text in combination with existing knowledge. Previous work following this approach draws on patterns in the curated knowledge graph in combi- nation with the graph of entity mentions in texts, allowing prediction of new knowledge base rela- tions (Riedel et al., 2013; Verga et al., 2015, 2017). Zhang et al. (2019) extend this work by incorpo- rating text predicates connecting entity mentions extracted using OpenIE tools (Fader et al., 2011; Lockard et al., 2019) and introducing the concept of “entity neighbourhoods” consisting of the binary OpenIE predicates and knowledge base relations1 that occur with a given entity as their subject or object. Drawing on the success of text based rep- resentations incorporated into entity recognition tasks (Gillick et al., 2019), we extend Zhang et.al.’s model by incorporating text based embeddings of entities and relations into the entity neighbourhood representations. Texts are drawn from knowledge base metadata and occurrences …

Content truncated to 500 words