Cross-document coreference resolution model based on neural entity embedding

Abstract

Natural Language Processing (NLP) is a way for computers to derive, analyze, and understand the meaning of human language in a smart and useful way. NLP considers the hierarchical structure of language that enables real-world applications such as automatic text summarization, event resolution, relationship extraction, and entity recognition to be presented in a proper human-computer interaction. One of the NLP components called Coreference Resolution (CR) is to determine whether the two noun phrases in natural language are referring to the same entity. In this context, an entity can be a real person, organization, place, or others, in which the referred term of such entity is called a mention. The task of CR when extended to resolve co-referent entities across multiple documents creates the Cross-Document Coreference Resolution (CDCR) task which requires special techniques to manage and address the mention chains within documents co-referring to the same entity across different documents. Currently, there are some limitations in the existing works in which the CDCR entities by variant referencing mentions are not well identified, and the grouping process to differentiate entities with lexical similarity is not well addressed. The main objective of this research is to propose a CDCR model using neural embedding of the entities and their mentions created by the representation of words using merely the input documents. This model created vectors of mentions and entities using neural embedding of mentions, regardless of the use of any external resources such as Knowledge Bases. For an advanced grouping of entities and their mentions, an improved density-based clustering technique containing DBSCAN and HDBSCAN clustering algorithms was employed. In addition, a prototype named CROCER was designed and developed as proof of concept to assess the model in an experimental environment. For evaluation, this model was applied to three publicly available datasets, called ‘John Smith Corpus’, ‘WePS-2 Collection’, and ‘Google Wikilinks’ from public open-source repositories. It measured the precision, recall, and F1 score of the model by three known scoring systems for Coreference Resolution, which are MUC, B3, and CEAF. Based on the findings, it can be concluded that the proposed model improved the F1 score of the datasets by almost 15.7%, 1.5%, and 9%, respectively.