Temporal Information Extraction by Predicting Relative Time-lines

Abstract

The current leading paradigm for temporal information extraction from text consists of three phases: (1) recognition of events and temporal expressions, (2) recognition of temporal relations among them, and (3) time-line construction from the temporal relations. In contrast to the first two phases, the last phase, time-line construction, received little attention and is the focus of this work. In this paper, we propose a new method to construct a linear time-line from a set of (extracted) temporal relations. But more importantly, we propose a novel paradigm in which we directly predict start and end-points for events from the text, constituting a time-line without going through the intermediate step of prediction of temporal relations as in earlier work. Within this paradigm, we propose two models that predict in linear complexity, and a new training loss using TimeML-style annotations, yielding promising results.

Temporal Common Sense Acquisition with Minimal Supervision

“…On the other hand, time has long been an important research area in NLP. Prior works have focused on the extraction and normalization of temporal expressions (Strötgen and Gertz, 2010;Angeli et al, 2012;Lee et al, 2014;Vashishtha et al, 2019), temporal relation extraction (Ning et al, 2017(Ning et al, , 2018cVashishtha et al, 2019), and timeline construction (Leeuwenberg and Moens, 2018). Recently, MCTACO (Zhou et al, 2019) summarizes five types of TCS and the three temporal dimensions studied here are all in their proposal.…”

Machine Reading of Historical Events

“…McClosky and Manning (2012) address the problem of ensuring semantically consistent timelines by finding patterns in the ordering of endpoints of different event types, which adds a common sense reasoning component to the system. Leeuwenberg and Moens (2018) construct a relative timeline of events directly, which allows them to circumvent typical pitfalls of pair-wise classifiers, such as computationally intractable inference and constructing globally inconsistent orderings (with cycles). Our work takes a similar approach but instead is able to construct an absolute timeline for the restricted domain of historical events.…”

Towards Extracting Absolute Event Timelines From English Clinical Reports

“…Current state-of-the-art systems are mostly neural-networkbased models [27], [28], [29], [30], [31]. [32] construct relative timelines from TimeML-style predictions, where each event is modeled as a timeline interval. We adopt this method to construct absolute interval-based timelines from TimeML as a baseline.…”

“…To be able to better compare our timeline annotations with the existing TimeML annotations, we follow the strategy of [32] to evaluate relative timelines using TimeML. Based on the timeline, We assign each TimeML-annotated event pair a temporal link (TLink), and calculate accuracy with the originally annotated TLinks.…”

“…First, to extract TLinks, we retrain a neural state-of-the-art clinical TLink extraction model [51] that is publicly available on our data split, using the existing TimeML annotations for training. From the extracted TLinks (of types before, after, and overlap), we position the events on the timeline following the TLinks-to-Timeline method by [32]: (1) Each event is assigned an interval with start variable x µ s and end variable x µ e . (2) The variables are set such that the predicted TLinks between the events are satisfied on the timeline 14 .…”