Recurrent Neural Network for Syntax Learning with Flexible Representations - Inria - Institut national de recherche en sciences et technologies du numérique Access content directly
Conference Papers Year : 2016

Recurrent Neural Network for Syntax Learning with Flexible Representations

Xavier Hinaut


We present a Recurrent Neural Network (RNN), namely an Echo State Network (ESN), that performs sentence comprehension and can be used for Human-Robot Interaction (HRI). The RNN is trained to map sentence structures to meanings (e.g. predicates). We have previously shown that this ESN is able to generalize to unknown sentence structures in English and French. The meaning representations it can learn to produce are flexible: it enables one to use any kind of " series of slots " (or more generally a vector representation) and are not limited to predicates. Moreover, preliminary work has shown that the model could be trained fully incrementally. Thus, it enables the exploration of language acquisition in a developmental approach. Furthermore, an " inverse " version of the model has been also studied, which enables to produce sentence structure from meaning representations. Therefore, if these two models are combined in a same agent, one can investigate language (and in particular syntax) emergence through agent-based simulations. This model has been encapsulated in a ROS module which enables one to use it in a cognitive robotic architecture, or in a distributed agent simulation.
Fichier principal
Vignette du fichier
ICDL-Epirob_Workshop-on-language-learning_2016__paper-folder.pdf (217.93 Ko) Télécharger le fichier
Origin : Files produced by the author(s)

Dates and versions

hal-01417060 , version 1 (15-12-2016)


  • HAL Id : hal-01417060 , version 1


Xavier Hinaut. Recurrent Neural Network for Syntax Learning with Flexible Representations. IEEE ICDL-EPIROB Workshop on Language Learning, Dec 2016, Cergy, France. ⟨hal-01417060⟩
504 View
197 Download


Gmail Facebook Twitter LinkedIn More