Towards Extracting Absolute Event Timelines From English Clinical Reports

Artuur Leeuwenberg,Marie-Francine Moens

doi:10.1109/taslp.2020.3027201

Abstract

Temporal information extraction is a challenging but important area of automatic natural language understanding. Existing approaches annotate and extract various parts of the temporal information conveyed in language like relative event order, temporal expressions, or event durations. Most schemes focus primarily on annotation of temporally certain (often explicit) information, resulting in partial annotation, and under-representation of implicit information. In this article, we propose an approach towards extraction of more complete (implicit and explicit) temporal information for all events, and obtain probabilistic absolute event timelines by modeling temporal uncertainty with information bounds. As a case study, we use our scheme to annotate a set of English clinical reports, and propose and evaluate a multi-regression model for predicting probabilistic absolute timelines, obtaining promising results.

Highlights

I N THIS article, we address the new task of bounded absolute timeline construction from text
Many temporal annotation schemes have been developed, all focusing on different aspects of temporality: relative event order [6]–[9], event durations [10], [11], and explicit temporal cues like temporal expressions [7], [12]–[14]
We believe this is the primary reason that our models perform better than the state-of-the-art D-Long Short-Term Memory (LSTM) baseline for durations, as this model uses GloVe embeddings. Another observation is that for most metrics, in general most models perform slightly better on the TimeML-bounded subset. We believe this is due to the slightly higher inter-annotator agreement (IAA) on these events, which can in turn be the result of the fact that TimeML focuses on explicit temporal information, whereas we focus on both explicit and implicit information

Summary

Introduction

I N THIS article, we address the new task of bounded absolute timeline construction from text. High quality temporal extraction from text could be an important enrichment of the structured electronic health record, with much potential for applications [4], [5]. Many temporal annotation schemes have been developed, all focusing on different aspects of temporality: relative event order [6]–[9], event durations [10], [11], and explicit temporal cues like temporal expressions [7], [12]–[14]. For a majority of events, existing schemes provide only partial event time information, leaving many event times

Objectives

Methods

Results

Conclusion