Recognition and Generation of Sentences through Self-organizing Linguistic Hierarchy Using MTRNN

Abstract

Abstract. We show that a Multiple Timescale Recurrent Neural Network (MTRNN) can acquire the capabilities of recognizing and generating sentences by self-organizing a hierarchical linguistic structure. There have been many studies aimed at finding whether a neural system such as the brain can acquire languages without innate linguistic faculties. These studies have found that some kinds of recurrent neural networks could learn grammar. However, these models could not acquire the capability of deterministically generating various sentences, which is an essential part of language functions. In addition, the existing models require a word set in advance to learn the grammar. Learning languages without previous knowledge about words requires the capability of hierarchical composition such as characters to words and words to sentences, which is the essence of the rich expressiveness of languages. In our experiment, we trained our model to learn language using only a sentence set without any previous knowledge about words or grammar. Our experimental results demonstrated that the model could acquire the capabilities of recognizing and deterministically generating grammatical sentences even if they were not learned. The analysis of neural activations in our model revealed that the MTRNN had self-organized the linguistic structure hierarchically by taking advantage of differences in the time scale among its neurons, more concretely, neurons that change the fastest represented "characters," those that change more slowly represented "words," and those that change the slowest represented "sentences."

Communication with Speech and Gestures: Applications of Recurrent Neural Networks to Robot Language Learning

“…The first part of the project focuses on using the Long-Short Term Memory (LSTM) [6] recurrent neural network architecture for implementing, on an iCub humanoid robot [7] (depicted in Figure 1), the gesture-word combination stage in infants before and when transitioning to multi-word utterances. This method will be compared to that of using Multiple Timescale Recurrent Neural Networks (MTRNNs) [8], which will be developed during the second part of the project and focuses on the implementation and verification of constructivist theories for language learning.…”