Resolving sentence ambiguity with planning and working memory resources: Evidence from fMRI

Susana Novais-Santos, James C. Gee, Maliha Shah, Vanessa Troiani, Melissa Work, Murray GrossmanNeuroImage 2007 · DOI

TL;DR

This fMRI study reveals that our brains use distinct neural resources for resolving ambiguous sentences: the dorsolateral prefrontal cortex (dIPFC) for strategic planning when sentence structure and verb meaning clash, and the inferior parietal cortex (IPC) for managing working memory demands. These findings support a model where executive functions, rather than just core language areas, are crucial for complex language comprehension.

What Problem Does This Paper Address?

Understanding how the human brain processes language, especially complex sentences, is a long-standing challenge in neuroscience. A key area of investigation is how we handle sentences with temporary structural ambiguities – sentences that can be interpreted in more than one way for a short period before the correct meaning becomes clear. For example, in "The citizens heard the election result was fixed," we initially might think "the election result" is the direct object of "heard," but it turns out to be the subject of a subordinate clause.

Two main theories attempt to explain how we resolve these ambiguities:

  1. Parallel Processing: This theory suggests we build multiple possible interpretations simultaneously and hold them in working memory (WM) until the correct one is identified. This would require significant WM resources.
  2. Serial Processing: This theory proposes we use on-line planning and decision-making, relying on probabilistic information (like verb biases) to choose the most likely interpretation and pursue it. This would involve less WM but more executive control.

This paper aims to differentiate between these theories by investigating the neural correlates of resolving sentence ambiguity, specifically examining the roles of executive resources like planning and working memory.

Methods

The researchers used functional magnetic resonance imaging (fMRI) to observe brain activity in 19 healthy adults while they read sentences presented phrase-by-phrase. The study manipulated two key factors:

  • Verb-Sentence Consistency: Sentences were constructed with verbs that were either highly compatible (more consistent) or less compatible (less consistent) with the sentence's grammatical structure (direct object vs. sentential complement). For instance, a verb strongly associated with direct objects would be used in a sentence that ultimately requires a sentential complement structure.
  • Working Memory Load: Half of the sentences included an extra phrase inserted just before the final disambiguating phrase. This was designed to increase the demand on working memory without altering the core grammatical structure.

The fMRI data was analyzed by focusing on the brain activity during the final phrase of each sentence, where the ambiguity is resolved. They used a subtraction method, comparing activity during the disambiguating phrase to activity during an initial, unbiased phrase, to isolate the neural processes involved in resolving the ambiguity.

Datasets

The study used custom-designed sentences based on materials from Garnsey et al. (1997). The sentences were carefully constructed and categorized based on:

  • Sentence Structure: Direct Object (DO) or Sentential Complement (SC).
  • Verb Bias: Verbs were classified as DO-biased, SC-biased, or equally biased (EQ-bias).
  • Noun Bias: Nouns were also classified based on their plausibility as direct objects.

The researchers also introduced novel materials and carefully controlled for factors like verb and noun frequency, letter length, and polysemy to ensure the observed effects were due to the experimental manipulations.

Key Insights and Ideas

  • Distinct Neural Resources for Ambiguity Resolution: The study provides strong evidence that resolving sentence ambiguity isn't a one-size-fits-all process. Different brain regions are recruited depending on the nature of the ambiguity.
  • dIPFC for Strategic Planning: When a sentence presents a conflict between the verb's typical usage and its grammatical role (less consistent verb-sentence mapping), the dorsolateral prefrontal cortex (dIPFC) shows increased activation. This suggests the dIPFC is involved in strategic planning and decision-making to figure out the correct interpretation.
  • IPC for Working Memory: When the task demands more working memory (e.g., holding more information due to the extra phrase), the inferior parietal cortex (IPC), particularly the left IPC, shows increased activation. This points to the IPC's role as a phonological buffer or a general resource for managing working memory load during language processing.
  • Support for Serial Processing: The findings, especially the dIPFC activation for less consistent sentences, align better with serial processing models that emphasize on-line decision-making and planning, rather than purely parallel models that would predict similar WM demands across all ambiguous sentences.
  • Large-Scale Network: Sentence comprehension, especially when dealing with ambiguity, relies on a distributed neural network involving not just core language areas (like IFC) but also executive control regions (dIPFC and IPC).

Key Results

  • fMRI Findings:
    • Less Consistent Sentences: Showed greater activation in the dIPFC (bilaterally) compared to more consistent sentences.
  • Increased Working Memory Load: Led to increased activation in the left IPC.
  • Inferior Frontal Cortex (IFC): Showed bilateral activation in both "more consistent" and "less consistent" conditions, suggesting its general role in sentence processing.
  • Behavioral Findings:
    • Reading times for the final phrase of less consistent sentences were longer than for more consistent sentences.
  • Accuracy in responding to probes was high across all conditions, indicating participants were engaged with the task.

Limitations and Future Work

  • Controlled Environment: The study used written sentences presented phrase-by-phrase in an MRI scanner. While this minimizes confounds, it differs from naturalistic, spoken language processing. The authors note that the fixed 3-second presentation interval in the fMRI study might differ from the self-paced reading in the behavioral experiment.
  • Lexical Content: While efforts were made to control for lexical factors, the authors acknowledge that some differences between conditions might be influenced by factors related to lexical content beyond those controlled.
  • Future Directions: The study opens avenues for further research into the precise interplay between planning and working memory in language, and how these executive functions interact with core language processing networks. The authors suggest that the observed activations are consistent with a serial processing approach and highlight the plausible co-activation of prefrontal and parietal regions during complex language tasks.

Summary updated

BibTeX
@article{novaissantos2007resolving,
  author = {Susana Novais-Santos and James C. Gee and Maliha Shah and Vanessa Troiani and Melissa Work and Murray Grossman},
  title = {{Resolving sentence ambiguity with planning and working memory resources: Evidence from fMRI}},
  journal = {NeuroImage},
  year = {2007},
  doi = {10.1016/j.neuroimage.2007.03.077},
}
Figure