Preview

PDF, English - main document Download (2MB) | Terms of use

Abstract

Attributes such as SIZE, WEIGHT or COLOR are at the core of conceptualization, i.e., the formal representation of entities or events in the real world. In natural language, formal attributes find their counterpart in attribute nouns which can be used in order to generalize over individual properties (e.g., 'big' or 'small' in case of SIZE, 'blue' or 'red' in case of COLOR). In order to ascribe such properties to entities or events, adjective-noun phrases are a very frequent linguistic pattern (e.g., 'a blue shirt', 'a big lion'). In these constructions, attribute meaning is conveyed only implicitly, i.e., without being overtly realized at the phrasal surface. This thesis is about modeling attribute meaning in adjectives and nouns in a distributional semantics framework. This implies the acquisition of meaning representations for adjectives, nouns and their phrasal combination from corpora of natural language text in an unsupervised manner, without tedious handcrafting or manual annotation efforts. These phrase representations can be used to predict implicit attribute meaning from adjective-noun phrases -- a problem which will be referred to as attribute selection throughout this thesis. The approach to attribute selection proposed in this thesis is framed in structured distributional models. We model adjective and noun meanings as distinct semantic vectors in the same semantic space spanned by attributes as dimensions of meaning. Based on these word representations, we make use of vector composition operations in order to construct a phrase representation from which the most prominent attribute(s) being expressed in the compositional semantics of the adjective-noun phrase can be selected by means of an unsupervised selection function. This approach not only accounts for the linguistic principle of compositionality that underlies adjective-noun phrases, but also avoids inherent sparsity issues that result from the fact that the relationship between an adjective, a noun and a particular attribute is rarely explicitly observed in corpora. The attribute models developed in this thesis aim at a reconciliation of the conflict between specificity and sparsity in distributional semantic models. For this purpose, we compare various instantiations of attribute models capitalizing on pattern-based and dependency-based distributional information as well as attribute-specific latent topics induced from a weakly supervised adaptation of Latent Dirichlet Allocation. Moreover, we propose a novel framework of distributional enrichment in order to enhance structured vector representations by incorporating additional lexical information from complementary distributional sources. In applying distributional enrichment to distributional attribute models, we follow the idea to augment structured representations of adjectives and nouns to centroids of their nearest neighbours in semantic space, while keeping the principle of meaning representation along structured, interpretable dimensions intact. We evaluate our attribute models in several experiments on the attribute selection task framed for various attribute inventories, ranging from a thoroughly confined set of ten core attributes up to a large-scale set of 260 attributes. Our results show that large-scale attribute selection from distributional vector representations that have been acquired in an unsupervised setting is a challenging endeavor that can be rendered more feasible by restricting the semantic space to confined subsets of attributes. Beyond quantitative evaluation, we also provide a thorough analysis of performance factors (based on linear regression) that influence the effectiveness of a distributional attribute model for attribute selection. This investigation reflects strengths and weaknesses of the model and sheds light on the impact of a variety of linguistic factors involved in attribute selection, e.g., the relative contribution of adjective and noun meaning. In conclusion, we consider our work on attribute selection as an instructive showcase for applying methods from distributional semantics in the broader context of knowledge acquisition from text in order to alleviate issues that are related to implicitness and sparsity.