Extracting summaries of movement from text

Keywords: natural language processing, machine learning, migration, GeoAI, health informatics
Session Type: Virtual Paper Abstract
Day: Sunday
Session Start / End Time: 2/27/2022 05:20 PM (Eastern Time (US & Canada)) - 2/27/2022 06:40 PM (Eastern Time (US & Canada))
Room: Virtual 59

Abstract

Language remains central to how humans describe what they see (Jackendoff, 1987). In the evolving landscape of artificial intelligence and natural language understanding, the task of language models is to have machines properly replicate the human translation of information from one form to another, such as events that occur in the three dimensional world and the words used to communicate them. However, the extraction of spatial terms from unstructured data sources such as text is largely limited to geographic place names, and current methods fall short of being able to identify and extract more complicated spatial features such as dynamic paths of movement. While this represents a challenging machine learning task, this paper aims to address one way of bridging the gap between unstructured text and dynamic spatial features by annotating texts involving movement for not just their syntactical elements, but their conceptual elements, which can then be used in a machine learning task. This builds off of work first presented in Jackendoff (1987) by creating corresponding links between syntactical units of a sentence to conceptual functions of spatial events (movement), which can then inform broader spatial summaries of a text. This research contributes a methodological approach to the treatment of natural language expressions and their geographic and spatial components, with the intent to further the capacity of machine translations of text within diverse applications areas, such as detecting the spread of a disease reported in media or finding patterns in risk factors from transnational migration narratives.