Semantic Patterns to Structure Timeframes for Event Ordering Enhancement
Abstract
Event ordering is a field in Event Extraction that
deals with the temporality aspect and order of occurrences of
events mentioned in a text. Event Ordering is essential because
any analysis of causalities and consequences of specific actions
or changes of state requires a time evaluation. Standard
approaches using machine learning models, with or without
inferences, start by identifying events in text and then generate
the temporal relationships between them individually. With no
consideration of flashbacks, flash-forward, and direct speech
temporal aspect, available models lack performance. In this
paper, we introduce a novel approach to group events in
temporal frames that we refer to as Timeframes. Three types
of timeframes will be presented: Publication, Narrative, and
Spoken. The purpose of this paper is to highlight the need of
this approach, define the different timeframes, introduce their
extraction process, evaluate the extraction and compare the
event ordering with and without the timeframe approach.