Research On Document-Level Event Factuality Identification

Document-level event factuality identification aims to infer the degree of certainty of a specific event occurring in a document,i.e.,whether the event is a fact,a possibility,or a counterfact.This dissertation focuses on the limitations of the features and annotation information used in previous studies,as well as some problems in the existing corpus and related work,and conducts the following specific research:(1)Document-level Event Factuality Identification Using Negation and Speculation Scope FeatureTo address the problem of meaningless labels and semantic missing of syntactic path features for capturing negative and speculative information in previous work,this dissertation explores the effectiveness of negation and speculation scopes for document-level event factuality identification,and trains a BERT-CRF model from cross-domain corpora to capture event-related scopes in DLEF corpus with document-level factuality.After validating the effectiveness,this dissertation proposes two approaches for improving the performance of factuality identification using scope features.The experimental results show that the proposed models achieve significant improvements in performance compared to the state-ofthe-art baselines.(2)Document-level Event Factuality Identification Incorporating Factuality InferenceTo address the problem that previous studies on document-level event factuality fail to effectively exploit the annotated sentence-level factuality values in the same document,this dissertation innovatively proposes an event factuality inference task that semantically connects sentence-level and document-level factuality.In terms of model architecture design,this dissertation proposes a sentence-to-document inference network containing a multi-layer interaction module and a gated aggregation module to integrate the factuality inference and identification tasks,and a multi-task learning framework is used to improve the overall performance.Experimental results on English and Chinese DLEF corpora show that the proposed model significantly outperforms the state-of-the-art baselines.(3)Evidence-Based Document-Level Event Factuality IdentificationExisting work on document-level event factuality highly dependents on the syntactic and semantic features associated with annotated event triggers,which may lead to the omission of important information to identify factuality.To address the above issues,this dissertation proposes a hypothesis that the event factuality can be inferred from the complete set of evidential sentences.To this end,this dissertation constructs an evidence-based documentlevel event factuality corpus,EB-DLEF,and introduces a new evidential sentence selection task,which provides a new research direction for this field.In addition,this dissertation proposes a pipeline approach for the two-step work of evidential sentence selection and factuality identification,which outperforms various baselines.Following the original task definition of document-level event factuality identification,this dissertation solves some problems in previous work and improves the performance significantly from two aspects of obtaining better features and using annotation information more effectively.At the same time,this dissertation identifies the limitations of related studies caused by the existing corpus,and brings a new perspective to the research in this field,which provides a reference for the subsequent work.