Exploring Supervised LDA Models for Assigning Attributes to Adjective-Noun Phrases

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 540–551, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Exploring Supervised LDA Models for Assigning Attributes to Adjective-Noun Phrases Matthias Hartung and Anette Frank Computational Linguistics Department Heidelberg University {hartung,frank}@cl.uni-heidelberg.de Abstract This paper introduces an attribute selection task as a way to characterize the inherent mea- ning of property-denoting adjectives in adjec- tive-noun phrases, such as e.g. hot in hot sum- mer denoting the attribute TEMPERATURE, rather than TASTE. We formulate this task in a vector space model that represents adjec- tives and nouns as vectors in a semantic space defined over possible attributes. The vectors incorporate latent semantic information ob- tained from two variants of LDA topic mod- els. Our LDA models outperform previous ap- proaches on a small set of 10 attributes with considerable gains on sparse representations, which highlights the strong smoothing power of LDA models. For the first time, we extend the attribute selection task to a new data set with more than 200 classes. We observe that large-scale attribute selection is a hard prob- lem, but a subset of attributes performs ro- bustly on the large scale as well. Again, the LDA models outperform the VSM baseline. 1 Introduction Corpus-based statistical modeling of semantics is gaining increased attention in computational linguis- tics. This field of research includes distributional vector space models (VSMs), i.e., models that rep- resent the semantics of words or phrases as vectors over high-dimensional cooccurrence data (Turney and Pantel, 2010; Baroni and Lenci, 2010, i.a.), as well as latent variable models (LVMs) which aggre- gate distributional observations in ’hidden’, or latent variables, thereby reducing the dimensionality of the data. An example of the latter are topic models (Blei et al., 2003), which have recently been applied to modeling selectional preferences of verbs (Ritter et al., 2010; ´O S´eaghdha, 2010), or word sense disam- biguation (Li et al., 2010). A topic that is increasingly studied in distribu- tional semantics is the semantics of adjectives, both in isolation (Almuhareb, 2006) and in compositional adjective-noun phrases (Hartung and Frank, 2010; Guevara, 2010; Baroni and Zamparelli, 2010). In this paper, we propose a new approach to a problem we denote as attribute selection: The task is to predict the hidden attribute meaning expressed by a property-denoting adjective in composition with a noun. The adjective hot, e.g., may denote at- tributes such as TEMPERATURE, TASTE or EMO- TIONALITY. These adjective meanings can be com- bined with nouns such as tea, soup or debate, which can be characterized in terms of attributes as well. The goal of the task is to determine the hidden at- tribute meaning predicated over the noun in a given adjective-noun phrase, as illustrated in (1). (1) a. a hotvalue summerconcept b. TEMPERATURE(summer) = hot It is by way of the composition of adjective and noun that specific attributes are selected from the ad- jective’s space of possible attribute meanings, and typically lead to a disambiguation of the …

Content truncated to 500 words