Categorial Grammar Induction as a Compositionality Measure for Emergent Languages in Signaling Games

Ryo Ueda, Taiga Ishii, Koki Washio, Yusuke Miyao

Keywords: Emergent Communication, Emergent Language, Categorial Grammar Induction, Syntax, Compositionality

TL;DR: This paper proposes a method for investigating the non-trivially compositional structure of emergent languages using Categorial Grammar Induction.

Abstract: This paper proposes a method to analyze the compositional structure of emergent languages using Categorial Grammar Induction (CGI). Emergent languages are communication protocols arising among agents in environments such as signaling games. Previous work has studied how similar or dissimilar emergent languages are to natural languages in compositionality. However, most of them focused on trivial compositionality, assuming flat structures in languages. We further focus on non-trivial compositionality, i.e., the relationship between hierarchical syntax and semantics. To this end, we apply CGI to emergent languages, inspired by previous NLP work. Given sentence-meaning pairs of a language, CGI induces 1) a categorial grammar that describes the syntax of the language and 2) a semantic parser that compositionally maps sentences to meanings. We also propose compositionality measures based on the grammar size and semantic parser performance. CGI and the proposed measures enable deeper insights into the non-trivial compositionality of emergent languages, while correlating well with existing measures like TopSim.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )