We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.CL

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo

Computer Science > Computation and Language

Title: Intrinsically Motivated Compositional Language Emergence

Abstract: Recently, there has been a great deal of research in emergent communication on artificial agents interacting in simulated environments. Recent studies have revealed that, in general, emergent languages do not follow the compositionality patterns of natural language. To deal with this, existing works have proposed a limited channel capacity as an important constraint for learning highly compositional languages. In this paper, we show that this is not a sufficient condition and propose an intrinsic reward framework for improving compositionality in emergent communication. We use a reinforcement learning setting with two agents -- a \textit{task-aware} Speaker and a \textit{state-aware} Listener that are required to communicate to perform a set of tasks. Through our experiments on three different referential game setups, including a novel environment gComm, we show intrinsic rewards improve compositionality scores by $\approx \mathbf{1.5-2}$ times that of existing frameworks that use limited channel capacity.
Comments: abridged version Accepted in NAACL 2021 workshop: Visually Grounded Interaction and Language (ViGIL) titled "gComm: An environment for investigating generalization in Grounded Language Acquisition"
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Cite as: arXiv:2012.05011 [cs.CL]
  (or arXiv:2012.05011v4 [cs.CL] for this version)

Submission history

From: Rishi Hazra [view email]
[v1] Wed, 9 Dec 2020 12:47:20 GMT (2106kb,D)
[v2] Sun, 13 Dec 2020 04:19:33 GMT (4215kb,D)
[v3] Wed, 19 May 2021 19:46:42 GMT (988kb,D)
[v4] Fri, 27 Jan 2023 11:04:58 GMT (3254kb,D)

Link back to: arXiv, form interface, contact.