Neural Activation Semantic Models: Computational lexical semantic models of localized neural activations

Proceedings of the 27th International Conference on Computational Linguistics, pages 2867–2878 Santa Fe, New Mexico, USA, August 20-26, 2018. 2867 Neural Activation Semantic Models: Computational lexical semantic models of localized neural activations Nikos Athanasiou School of ECE NTUA Athens, Greece el12074@central.ntua.gr Elias Iosif ∗ IFF, University of Nicosia Nicosia, Cyprus iosif.e@unic.ac.cy Alexandros Potamianos School of ECE NTUA Athens, Greece apotam@gmail.com Abstract Neural activation models that have been proposed in the literature use a set of example words for which fMRI measurements are available in order to find a mapping between word seman- tics and localized neural activations. Successful mappings let us expand to the full lexicon of concrete nouns using the assumption that similarity of meaning implies similar neural activation patterns. In this paper, we propose a computational model that estimates semantic similarity in the neural activation space and investigates the relative performance of this model for various natural language processing tasks. Despite the simplicity of the proposed model and the very small number of example words used to bootstrap it, the neural activation semantic model per- forms surprisingly well compared to state-of-the-art word embeddings. Specifically, the neural activation semantic model performs better than the state-of-the-art for the task of semantic simi- larity estimation between very similar or very dissimilar words, while performing well on other tasks such as entailment and word categorization. These are strong indications that neural acti- vation semantic models can not only shed some light into human cognition but also contribute to computation models for certain tasks. 1 Introduction The human ability of translating concepts into words and back depends on the ability of mind to decode and encode meaning. This mental process, which is not currently completely understood, has captivated the interest of both neuroscientists and computational linguists (Haxby et al., 2001; Ishai et al., 1999; Cree and McRae, 2003; G. Kanwisher et al., 1997; Just et al., 2010). Specifically, when a person experiences a visual stimulus of a concept, reads, speaks or writes a word, particular neuronal regions in the brain are activated (Pulvermller, 2001). Various studies have been carried out to explore brain encoding and decoding mechanisms when a stimulus is present, as detailed next. For visual stimuli, studies have shown that is feasible to discriminate and reconstruct images using patterns of neural activity, mainly found in the visual cortex (Kay et al., 2008; Thirion et al., 2006; Nishimoto et al., 2011; Naselaris et al., 2009; Miyawaki et al., 2008), the part of brain responsible for visual information processing. Other works have demonstrated the relationship between cognitive perception and speech (Benson et al., 2001; Formisano et al., 2008). Regarding textual stimuli, researchers have shown distributed semantic maps of words are present in our brains (Huth et al., 2012; Sudre et al., 2012). Lexical semantics are based on the assumption that similar words appear in similar contexts (Harris, 1954). Based on that assumption, two different approaches for building semantic models have been proposed. The first approach is to encode word semantics, by applying dimensionality …

Content truncated to 500 words