Acquisition of Common Sense Knowledge for Basic Level Concepts

International Conference RANLP 2009 - Borovets, Bulgaria, pages 23–27 Acquisition of common sense knowledge for basic level concepts Eduard Barbu Center for Mind/Brain Sciences Rovereto Trento, Italy eduard.barbu@unitn.it Abstract Feature norms can be regarded as repositories of common sense knowledge for basic level con- cepts. We acquire from very large corpora feature-norm-like concept descriptions using a combination of a weakly supervised method and an unsupervised method. The success in iden- tifying the specific properties listed in the fea- ture norms as well as the success in acquiring the classes of properties present in the norms are reported. Keywords basic level categories, common-sense knowledge, feature norms 1 Introduction The acquisition of common sense knowledge is the fo- cus of a series of projects originated in AI like CYC [5] or Open Mind [10]. The aim of this paper is the acqui- sition of every day knowledge for a restricted category of concepts: basic level concepts denoting concrete ob- jects. One of the main criteria for concept organization in initial studies carried both in psychology and AI [2] was thought to be the taxonomic criteria. Early work in psychology [9] showed that not all levels of taxon- omy are equal with respect to object categorization. There is a privileged level at which people consistently classify the objects in common speech called the basic level. For example, encountering an object (e.g. 19th century dinning table) in ordinary discussion we do not categorize it at its specific level (19th century dinning table) nor to its more general level (e.g. entity) but to its basic level (table). The basic level concept is the most inclusive level at which concepts share common features, it carves the world at its joints. Examples of basic level concepts are bird, dog, cat or car. To acquire common-sense knowledge for basic level concepts we rely on an ongoing effort in cognitive psy- chology: the feature norms. In a task called feature-generation subjects list what they believe the most important properties for a set of test concepts are. The experimenter processes the re- sulting conceptual descriptions and registers the final representation in the norm. Thus, a feature norm is a database containing a set of concepts and their most salient features (properties). The recorded properties are pieces of common sense knowledge. For example, in a norm one finds statements like: • An apple (concept) is a fruit (property)1. • An airplane (concept) is used for people trans- portation (property). In this paper we explore the possibility to acquire common-sense knowledge from very large corpora. The type of properties one finds in the norms guides the knowledge-extraction task. A double classification of the properties in the norms is used. At the morpho- logical level the properties are grouped according to the part of speech of the words used to express them (noun properties, adjective properties, verb proper- ties). At the semantic level we group the properties in semantic classes (taxonomic properties, part prop- erties, etc.). The properties in certain semantic classes are …

Content truncated to 500 words