Reconstructing Maps from Text

TL;DR

This paper investigates how language models can reconstruct spatial maps from text, revealing the importance of direct co-occurrence data and proposing an instance-based DSM that overcomes frequency limitations.

Contribution

It demonstrates the necessity of direct co-occurrence for traditional DSMs and introduces an instance-based DSM capable of map reconstruction independent of co-occurrence frequency.

Findings

Direct co-occurrence is necessary for traditional DSMs to reconstruct maps.

An instance-based DSM can reconstruct maps without relying on co-occurrence frequency.

The study constrains the understanding of semantic representations in language models.

Abstract

Previous research has demonstrated that Distributional Semantic Models (DSMs) are capable of reconstructing maps from news corpora (Louwerse & Zwaan, 2009) and novels (Louwerse & Benesh, 2012). The capacity for reproducing maps is surprising since DSMs notoriously lack perceptual grounding (De Vega et al., 2012). In this paper we investigate the statistical sources required in language to infer maps, and resulting constraints placed on mechanisms of semantic representation. Study 1 brings word co-occurrence under experimental control to demonstrate that direct co-occurrence in language is necessary for traditional DSMs to successfully reproduce maps. Study 2 presents an instance-based DSM that is capable of reconstructing maps independent of the frequency of co-occurrence of city names.