Section: New Results

NLP and computational neurolinguistics

Participants : Éric Villemonte de La Clergerie, Murielle Fabre.

In the context of the CRCNS international network, the ANR-NSF NCM-ML project (dubbed “Petit Prince project”) aims to discover and explore correlations between features (or predictors) provided by NLP tools such as parsers, and brain imagery (fMRI) data resulting from listening of the novel Le Petit Prince. Following the availability of an increasing amount of fMRI datasets in French and English, the project has investigated the correlations between fMRI observations and an increasing number of parser-based features based on several parsers representing a number of architecture types (LSTM, RNN, Dyalog-SR [statistical], FRMG [hybrid symbolic/statistical]) [20].

While pursuing the purely computation goal of developing a method of variable beam size inference for Recurrent Neural Network Grammar (rnng ) the project investigated how different beam search methods can show different goodness of fit with fMRI signal recorded during naturalistic story listening [58]. This approach is part of a new trend that is now emerging under the name of cognitively inspired NLP, where the effort to leverage from what we know of human cognition to increase machine processing of language data. Drawing inspiration from sequential Monte-Carlo methods such as particle filtering, we illustrated the relevance of our new method for speeding up the computations of direct generative parsing for rnng , and revealing the potential cognitive interpretation of the underlying representations built by the search method and its beam activity through the analysis of neuro-imaging signal.

A second focus of the project is on compositionality, memory retrieval and syntactic composition during language comprehension. By using quantifications of these hypothesised processes as obtained from computational linguistics we seek to highlight their neural substrates and better understand or model human cognitition.

While linguistic expressions have been binarised as compositional and non-compositional given the lack of compositional linguistic analysis, the so-called Multi-word Expressions (MWEs) demonstrate finer-grained degrees of conventionalisation and predictability in psycho-linguisitcs, which can be quantified through computational Association Measures, like Point-wise Mutual Information and Dice's Coefficient [57]. An fMRI analysis was conducted to investigate to what extent these computational measures and the underlying cognitive processes they reflect are observable during on-line naturalistic sentence processing. Our results show that predictability, as quantified through Dice's Coefficent, is a better predictor of neural activation for processing MWEs and the more cognitively plausible computational metric. Computational results (1348) were obtained on MWE identification in French based on new method searching for frequent dependency-patterns [13]. These identifications in the Little Prince are contrasted with the ones published for English [69] and will yield an fMRI analysis comparing the two languages and the possible typological differences that the two languages may reflect in terms of morphological strategies to achieve lexical conventionalisation.