Processing MWEs: Neurocognitive Bases of Verbal MWEs and Lexical Cohesiveness within MWEs

Shohini Bhattasali, Murielle Fabre, John HaleCOLING 2018

TL;DR

This study uses fMRI to investigate how the brain processes multiword expressions (MWEs), revealing distinct neural patterns for verbal vs. non-verbal MWEs and for MWEs with varying degrees of lexical cohesiveness. The findings suggest that grammatical category and lexical tightness play a crucial role in how MWEs are understood, offering neurocognitive evidence for their unique processing.

What Problem Does This Paper Address?

Multiword expressions (MWEs) – like "kick the bucket" or "boa constrictors" – are common in language but pose a challenge for computational linguistics due to their diverse forms and often non-literal meanings. While previous research has shown that MWEs are processed faster than comparable single words, the underlying neural mechanisms remain unclear. This paper aims to bridge this gap by using fMRI to explore whether different types of MWEs (verbal vs. non-verbal) and MWEs with varying degrees of lexical cohesiveness (how tightly bound the words are) activate distinct brain regions. This could shed light on whether MWEs are processed as single units or through more compositional means.

Methods

The researchers employed a functional Magnetic Resonance Imaging (fMRI) study. Participants listened to a spoken narrative of "The Little Prince" while their brain activity was recorded. The core of their analysis involved a regression approach where word-by-word predictors derived from the text were correlated with the fMRI timecourses.

Key methodological aspects include:

  • Stimulus: A naturalistic, spoken narrative to mimic real-world language comprehension.
  • MWE Identification: A transition-based MWE analyzer was used to identify 742 MWEs within the text.
  • Categorization: MWEs were categorized into verbal (headed by a verb) and non-verbal.
  • Cohesiveness Metric: Pointwise Mutual Information (PMI), a standard measure for identifying collocations, was repurposed as a gradient metric to quantify the lexical cohesiveness of MWEs. Higher PMI scores indicate greater cohesiveness.
  • Statistical Analysis: A General Linear Model (GLM) was used at both the subject (first-level) and group (second-level) levels to analyze the fMRI data. This allowed for the localization of brain activity associated with different MWE characteristics.

Datasets

  • Primary Stimulus: The audio stimulus was Antoine de Saint-Exupéry's "The Little Prince," translated by David Wilkinson and read by Nadine Eckert-Boulet. This comprised 15,388 words and lasted over an hour and a half.
  • MWE Identification Training Data: The MWE analyzer was trained on examples from the Children's Book Test (CBT) from the Facebook bAbI project.
  • External Lexicons for MWE Identification: Unitex lexicon, SAID corpus, Cambridge International Dictionary of Idioms, and Dictionary of American Idioms were used to supplement MWE identification.
  • PMI Calculation Corpus: Corpus of Contemporary English (COCA) was used for calculating PMI scores.

No new benchmark datasets were introduced.

Key Insights and Ideas

  • Grammatical Category Matters: The study provides strong evidence that the grammatical category of an MWE (specifically, whether it's verb-headed) influences its neural processing. Verbal MWEs and non-verbal MWEs activate distinct brain regions, suggesting that the inherent argument structure of verbs plays a role.
  • Cohesiveness as a Neural Gradient: PMI, a measure typically used for identifying MWEs, can effectively serve as a gradient predictor for their neural processing. This suggests that the degree of lexical tightness within an MWE is a cognitively relevant feature.
  • Repurposing PMI: The study successfully repurposes PMI, an association measure, as a cognitively plausible metric to model the cohesiveness and compositionality of MWEs, demonstrating its utility beyond simple identification.
  • Spatial Dissociation: The findings highlight a spatial dissociation in brain activation for different MWE types and cohesiveness levels, supporting the idea that the brain employs distinct mechanisms for processing these linguistic units.

Key Results

  • Verbal vs. Non-verbal MWEs: Analysis 1 revealed distinct activation patterns. Verbal MWEs primarily activated right-lateralized regions including the Inferior Frontal Gyrus (IFG) pars triangularis and Inferior Parietal Lobule. Non-verbal MWEs showed a broader, mostly right-lateralized pattern involving the Angular Gyrus, Supplementary Motor Area, Cerebellum, and Supramarginal Gyrus.
  • Lexical Cohesiveness: Analysis 2 showed that increasing cohesiveness (higher PMI) correlated with activation in the Precuneus and Supplementary Motor Area. Conversely, decreasing cohesiveness (lower PMI) was associated with activity in core language network nodes like Broca's area and the posterior Temporal Gyrus. This suggests that highly cohesive MWEs might be retrieved as units, while less cohesive ones are processed more compositionally.
  • Behavioral Results: Participants demonstrated high comprehension accuracy (90%) on the story, indicating attentive listening.

Limitations and Future Work

  • Limitations:
    • The study's dataset was skewed towards verbal MWEs, limiting strong conclusions about non-verbal MWEs, which were a "mixed bag" of types.
  • The study focused on native English speakers, and cross-linguistic comparisons are needed.
  • Future Work:
    • Replicate the experiment with non-native speakers to study the impact of acquisition timing on neural processing.
  • Compare compositional expressions (e.g., VPs) directly with non-compositional MWEs to further illustrate the spatial dissociation.
  • Explore alternative metrics for compositionality, such as word space models based on distributional semantics.
  • Investigate MWE processing cross-linguistically to compare processing similarities.

Summary updated

BibTeX
@inproceedings{bhattasali2018processing,
  author = {Shohini Bhattasali and Murielle Fabre and John Hale},
  title = {{Processing MWEs: Neurocognitive Bases of Verbal MWEs and Lexical Cohesiveness within MWEs}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2018},
}
Figure