A Structured Vector Space Model for Hidden Attribute Meaning in Adjective-Noun Phrases

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 430–438, Beijing, August 2010 A Structured Vector Space Model for Hidden Attribute Meaning in Adjective-Noun Phrases Matthias Hartung and Anette Frank Computational Linguistics Department Heidelberg University {hartung, frank}@cl.uni-heidelberg.de Abstract We present an approach to model hid- den attributes in the compositional se- mantics of adjective-noun phrases in a distributional model. For the represen- tation of adjective meanings, we refor- mulate the pattern-based approach for at- tribute learning of Almuhareb (2006) in a structured vector space model (VSM). This model is complemented by a struc- tured vector space representing attribute dimensions of noun meanings. The com- bination of these representations along the lines of compositional semantic principles exposes the underlying semantic relations in adjective-noun phrases. We show that our compositional VSM outperforms sim- ple pattern-based approaches by circum- venting their inherent sparsity problems. 1 Introduction In formal semantic theory, the compositional se- mantics of adjective-noun phrases can be modeled in terms of selective binding (Pustejovsky, 1995), i.e. the adjective selects one of possibly several roles or attributes1 from the semantics of the noun. (1) a. a blue car b. COLOR(car)=blue In this paper, we define a distributional frame- work that models the compositional process un- derlying the modification of nouns by adjectives. 1In the original statement of the theory, adjectives se- lect qualia roles that can be considered as collections of at- tributes. We focus on property-denoting adjectives as they are valuable for acquiring concept representations for, e.g., ontology learning. An approach for au- tomatic subclassification of property-denoting ad- jectives is presented in Hartung and Frank (2010). Our goal is to expose, for adjective-noun phrases as in (1a), the attribute in the semantics of the noun that is selected by the adjective, while not being overtly realized on the syntactic level. The semantic information we intend to capture for (1a) is formalized in (1b). Ideally, this kind of knowledge could be ex- tracted from corpora by searching for patterns that paraphrase (1a), e.g. the color of the car is blue. However, linguistic patterns that explicitly relate nouns, adjectives and attributes are very rare. We avoid these sparsity issues by reducing the triple r=⟨noun, attribute, adjective⟩ that encodes the relation illustrated in (1b) to tu- ples r′=⟨noun, attribute⟩ and r′′=⟨attribute, adjective⟩, as suggested by Turney and Pantel (2010) for similar tasks. Both r′ and r′′ can be observed much more frequently in text corpora than r. Moreover, this enables us to model adjec- tive and noun meanings as distinct semantic vec- tors that are built over attributes as dimensions. Based on these semantic representations, we make use of vector composition operations in order to reconstruct r from r′ and r′′. This, in turn, al- lows us to infer complete noun-attribute-adjective triples from individually acquired noun-attribute and adjective-attribute representations. The contributions of our work are as follows: (i) We propose a framework for attribute selection based on structured vector space models (VSM), using as meaning dimensions attributes elicited 430 by adjectives; …

Content truncated to 500 words