Abstract
Philosopher and mathematician Gottlob Frege described word meaning in terms of sense and reference. One word refer to an almost unlimited number of things in the world (referents) while one referent can be referred to by a number of different expressions, with slightly different senses implying different perspectives. Morning star and evening star are names for the same entity, but have different senses Linguistic phenomena involved in this reference variation are synonymy, hyponymy, metonymy and metaphor. When one word has different senses, we speak of polysemy, in which case the word can refer to very different types of things. According to the Distributional Hypothesis, the meaning of words arises from their linguistic context. Recent methods in linguistics represent words in semantic spaces based purely on their surrounding words. In these representations, called word embeddings, similar words group together so that meaning relations can be inferred from the space. Individual word senses, let alone referents are no longer apparent from such representations as they conflate all senses and referents into one representation, leading to distortions in the semantic space. Nevertheless, they are widely used in Computational Linguistics and Digital Humanities. In this project, I will decompose word embeddings into the two components of meanings, sense and reference, and develop a model to predict word sense and reference from word embeddings alone. This will enable large-scale investigations of the fundamental semantic phenomena of sense and reference as well as figurative meanings, which currently rely on expensive manual and time consuming annotations.