Institutional Repository
Technical University of Crete
Technical University of Crete
EN
|
EL
| URI: | http://purl.tuc.gr/dl/dias/42E31263-711F-4297-90ED-9ED2BFA8F39B | ||
| Year | 2015 | ||
| Type of Item | Master Thesis | ||
| License |
|
||
| Bibliographic Citation | Spyridon Georgiladakis, "Semantic composition in DSMs: activational priming and transformational properties for similarity modeling", Master Thesis, School of Electronic and Computer Engineering, Technical University of Crete, Chania, Greece, 2015 https://doi.org/10.26233/heallink.tuc.64991 | ||
| Appears in Collections |
Distributional Semantic Models (DSMs) have been successful at modeling the meaning of words in isolation. Interest has recently shifted to compositional structures, i.e., lexical units that comprise of words that represent individual concepts, such as phrases and sentences. Network DSMs (NDSMs) represent and handle semantics via operations on word neighborhoods, i.e., semantic graphs comprising of a target lexical unit's semantically most similar words. Semantic networks are based on activational priming, a cognitively-based theorythat a specific area which shares common features can be activated upon the triggering of a related stimulus. In this thesis, a variety of activation composition and similarity modeling strategies is proposed that aims to address compositionality within the framework of the respective layers of NDSMs. In the activation layer, we propose several activation schemes, motivated by psycholinguistics, that utilize variable size activations in order to compose neighborhoods for complex structures. In the similarity layer, we model similarity metrics that operate on the derived neighborhoods to estimate similarity. The proposed schemes cover a range of approaches for modeling semantics in complex structures. We also investigate modifier properties and transformational models from the literature, and propose a fusion scheme that regulates the transformational properties of phrase modifiers in order to weight the contribution of its component models for handling semantics. To this end, the model utilizes network and transformational models under a fusion scheme that models similarity.It is shown that, by fusing strictly compositional with transformational models to realise a flexible model that adapts to phrase behavior by considering modifier roperties, performance gains can be achieved.
Copyright © DIAS 2013
