Multilingual processing of temporal expressions: detection and normalisation

Abstract

Temporal information processing allows us to identify when an event happened, to temporally relate different events and to create timelines with all the extracted information. To achieve these goals, detecting time expressions in a text (e.g. the week of March 6) and normalising them to a standardised value (e.g. a concrete week of a concrete year) is a fundamental step. In this work, we have developed a modular system for time expression processing in English and Spanish, which consists of a Flair sequence-labelling detector based on neural networks and a normaliser built on the TimeNorm synchronous context-free grammar system. Furthermore, we provide an exhaustive study on how we built our Spanish TimeNorm grammar, which is our main contribution and could be useful to adapt the normalisation system to other inflected languages. We evaluated this approach on the TempEval-3 challenge and obtained state-of-the-art results both in detection and normalisation for English and Spanish.