Title: Entities in Distributional Semantics

Abstract:
The distributional semantics community has extensively addressed various 
phenomena arising from concepts (like, common nouns: country, president) 
such as modelling similarity, hypernymy, etc. However, entities still 
remain an under-scrutinized phenomenon. Although distributional 
representations easily yield that an entity like Italy is more similar 
to Spain than Germany but fine-grained facts about Italy (like area, 
population, official language, GDP) are still difficult to express in 
vector algebra and geometry. Consequently, there exists a widely 
accepted notion that distributional embeddings contain only coarse-
grained information and current approaches treat the embeddings of 
entities and concepts at par.
However, ontologically and according to model-theoretic semantics, 
entities <e> are distinct from concepts <e,t> and we hypothesize that 
this distinction can be recovered from the standard distributional 
models. In this presentation, I will present empirical evidence from our 
experiments which supports our primary hypothesis. The experiments deal 
with three research questions:
• Do the word embeddings of entities capture fine-grained information?
Eg.: Germany has a population of 80 million and Germany’s capital is 
Berlin.
• Are entities (Mozart) ontologically distinct from their related 
concepts (composer)?
• What are the aspects, w.r.t. distributional geometry and behaviour, 
which differentiate entities and concepts within the same distributional 
space?