A Geometric Algebra Based Distributional Model to Encode Sentences Semantics
Word space models are used to encode the semantics of natural language elements by means of high dimensional vectors [23 ]. Latent Semantic Analysis (LSA) methodology [15 ] is well known and widely used for its generalization properties. Despite of its go
- PDF / 595,987 Bytes
- 14 Pages / 439.37 x 666.142 pts Page_size
- 114 Downloads / 212 Views
A Geometric Algebra Based Distributional Model to Encode Sentences Semantics Agnese Augello, Manuel Gentile, Giovanni Pilato and Giorgio Vassallo
Abstract Word space models are used to encode the semantics of natural language elements by means of high dimensional vectors [23]. Latent Semantic Analysis (LSA) methodology [15] is well known and widely used for its generalization properties. Despite of its good performance in several applications, the model induced by LSA ignores dynamic changes in sentences meaning that depend on the order of the words, because it is based on a bag of words analysis. In this chapter we present a technique that exploits LSA-based semantic spaces and geometric algebra in order to obtain a sub-symbolic encoding of sentences taking into account the words sequence in the sentence. Keywords Semantic spaces · Sentences encoding · Clifford algebra
1 Introduction Two rather orthogonal theories in Natural Language Processing are the symbolic [11] and distributional [25] paradigms: the former is compositional but only qualitative, the latter is non-compositional but quantitative [9].
A. Augello (B) · G. Pilato ICAR, CNR V.le delle Scienze - Ed.11, 90128 Palermo, Italy e-mail: [email protected] G. Pilato e-mail: [email protected] M. Gentile ITD, CNR Via Ugo La Malfa, 153, 90146 Palermo, Italy e-mail: [email protected] G. Vassallo DICGIM Università di Palermo, V.le delle Scienze, Ed. 6, 90128 Palermo, Italy e-mail: [email protected]
C. Lai et al. (eds.), Distributed Systems and Applications of Information Filtering and Retrieval, Studies in Computational Intelligence 515, DOI: 10.1007/978-3-642-40621-8_6, © Springer-Verlag Berlin Heidelberg 2014
101
102
A. Augello et al.
Distributional approaches try to quantify and categorize semantic correspondences between linguistic entities. The key idea is the distributional hypothesis, which states that words having similar meanings will occur in similar contexts [21]. This means that there is a correlation between distributional and meaning similarity, that makes it possible to estimate the latter starting from the former. Algorithms that try to acquire distributional meaning can be divided in two categories: the first one includes all approaches that try to build distributional profiles for words based on which other words surround them, while the other one embraces the techniques that build distributional profiles based on in which text regions word occur [23]. The core of the distributional approach is that linguistic meaning is essentially differential, i.e. differences of meaning are mediated by differences of distributions, therefore the distributional methodology deals only with meaning differences or semantic similarity. Usually the model that captures the pattern of distribution of single words across a set of contexts is a vector and the assessment of these models is often done by exploiting relations of semantic similarity between individual words. Saussure gave the foundation of what developed later as structuralism; in a
Data Loading...
