Eitan Wagner, Renana Keydar, Omri Abend

Description:

Narratives can be modeled across many dimensions, such as time, space, and events. Focusing
on the spatial dimension of narratives, we present the task of event-location tracking in
narrative texts. This task seeks to extract a chain of locations or location categories for a given
sequence of text segments. We propose various neural architectures for the task, utilizing
state-of-the-art neural components. These architectures are capable of processing narratives as
whole units with different levels of context awareness. We also present methods for the
generation of location embeddings, which are crucial for the expansion of the work to other
categories and domains. We apply our methods to Holocaust testimonies, which provide a
unique case of a large set of narratives with a relatively similar pattern of location trajectories.
We compare our methods to several different baseline methods that use state-of-the-art
neural components in a local manner. We show that our methods significantly outperform the
baseline methods in terms of the list of locations that the model generates, thus providing
evidence of the importance of document-level processing.

Publications:

• Event-Location Tracking in Narratives: A Case Study on Holocaust Testimonies (submitted to
  EMNLP 2023 Narrative Understanding Workshop)