Learning to interpret spatial natural language in terms of qualitative spatial relations*

Abstract

Computational approaches in spatial language understanding nowadays distinguish and use different aspects of spatial and contextual information. These aspects comprise linguistic grammatical features, qualitative formal representations, and situational context-aware data. In this chapter, we apply formal models and machine learning techniques to map spatial semantics in natural language to qualitative spatial representations. In particular, we investigate whether and how well linguistic features can be classified and automatically extracted and mapped to region-based qualitative relations using corpus-based learning. We separate the challenge of spatial language understanding into two tasks: (i) we identify and automatically extract those parts from linguistic utterances that provide specifically spatial information, and (ii) we map the extracted parts that result from the first task to qualitative spatial representations. In this chapter, we present both tasks and we particularly discuss experimental results of the second part of mapping linguistic features to qualitative spatial relations. Our results show that region-based spatial relations can indeed be learned to a high degree and that they are distinguishable on the basis of different linguistic features.2

Global machine learning for spatial ontology population

“…However, such a mapping is useful because formal spatial models enable automatic spatial 2 reasoning that is difficult to perform on natural language expressions. To overcome the complexity of this problem in a systematic way, our spatial ontology is divided into two abstraction layers [7,8,9]:…”

“…A layer of linguistic conceptual representation called spatial role labeling (SpRL), which predicts the existence of spatial information at the sentence level by identifying the words that play a particular spatial role as well as their spatial relationship [10]; 2. A layer of formal semantic representation called spatial qualitative labeling (SpQL), in which the spatial relation is described with semantic attributes based on qualitative spatial representation models (QSR) [11,12].…”

“…We discuss the two layers of semantics and the adequacy of mapping to qualitative spatial representation and reasoning models in [12,8,9]. We have previously developed machine learning models, but they were restricted to the annotation of text with the concepts of the SpRL layer [10,38].…”

“…The SpRL layer has been worked out by the participants of a semantic evaluation shared task that we have proposed [35,39,40,37]. Preliminary classification of the ground truth spatial relations to RCC classes is investigated by the authors in [12] using an SMO implementation of the SVM multi-class classifier.…”

Building a corpus of spatial relational expressions extracted from web documents

“…Understanding and formally modeling the semantics of SRE, like the ones listed above, is critical for a wide variety of applications, including information retrieval [23,6], human-robot interaction [20], and automatic text and speech processing and generation [17,1]. As a step in this direction, the work in this paper is directed toward the long-term goal of constructing and analyzing a large corpus of humanproduced expressions containing SREs alongside georeferenced locations of the involved objects, and the context in which these expressions have been produced.…”