Abstract
With the rapid development of deep learning, most of current state-of-the-arttechniques in natural langauge processing are based on deep learning modelstrained with argescaled static textual corpora. However, we human beings learnand understand in a different way. Thus, grounded language learning argues thatmodels need to learn and understand language by the experience and perceptionsobtained by interacting with enviroments, like how humans do. With the help ofdeep reinforcement learning techniques, there are already lots of worksfocusing on facilitating the emergence of communication protocols that havecompositionalities like natural languages among computational agentspopulation. Unlike these works, we, on the other hand, focus on the numericconcepts which correspond to abstractions in cognition and function words innatural language. Based on a specifically designed language game, we verifythat computational agents are capable of transmitting numeric concepts duringautonomous communication, and the emergent communication protocols can reflectthe underlying structure of meaning space. Although their encodeing method isnot compositional like natural languages from a perspective of human beings,the emergent languages can be generalised to unseen inputs and, moreimportantly, are easier for models to learn. Besides, iterated learning canhelp further improving the compositionality of the emergent languages, underthe measurement of topological similarity. Furthermore, we experiment anotherrepresentation method, i.e. directly encode numerals into concatenations ofone-hot vectors, and find that the emergent languages would becomecompositional like human natural languages. Thus, we argue that there are 2important factors for the emergence of compositional languages.