TL;DR
This paper introduces a modified computational model of sentence comprehension that accounts for slow, incorrect responses in individuals with aphasia (IWA). By allowing backtracking to fail, the model better explains IWA's comprehension difficulties, suggesting that aphasia involves more than just simple retrieval failures.
What Problem Does This Paper Address?
Understanding how the brain processes sentences is a complex challenge. One prominent theory, the direct-access model, suggests that we retrieve word meanings from memory using syntactic and semantic cues. While this model explains fast, correct responses, it struggles to account for slower, incorrect responses, particularly in individuals with aphasia (IWA). IWA often exhibit difficulties in sentence comprehension, and previous research has shown they sometimes produce slower incorrect responses, a phenomenon the standard direct-access model cannot explain. This paper aims to bridge this gap by proposing and evaluating a modified direct-access model that can account for these slow, erroneous responses in IWA.
Methods
The core of this research involves computational modeling. The authors implement two versions of the direct-access model:
- Base Direct-Access Model (DA): This is the standard model where retrieval cues allow parallel access to memory items. Backtracking is assumed to occur when initial retrieval fails, and this backtracking is always successful in finding the correct item, albeit with an added processing cost.
- Modified Direct-Access Model (MDA): This is the novel contribution. The MDA extends the base model by introducing a new parameter, $\theta_b$, which represents the probability of successful retrieval after backtracking. This means backtracking can now fail, leading to misretrievals and thus slow, incorrect responses.
Both models were implemented using probabilistic programming in Stan and fitted to experimental data from individuals with aphasia and control participants. The models were compared using posterior predictive checks and Bayes factors.
Datasets
The study uses data from a self-paced listening task in German, previously collected by Pregla et al. (2021). This dataset includes:
- Listening times (LT): Recorded at specific regions of interest within sentences.
- Accuracy: Performance on a picture-selection task following sentence presentation.
- Auditory lexical decision task (LDT) data: Used as a predictor for individual differences in lexical access speed.
The data was collected from 50 control participants and 21 individuals with aphasia (IWA).
Key Insights and Ideas
- Backtracking isn't always a perfect fix: The most significant insight is that the assumption of always-successful backtracking in the standard DA model is too simplistic. Allowing backtracking to fail is crucial for explaining the observed behavior of IWA.
- Aphasia is multifaceted: The findings suggest that IWA's comprehension deficits are not solely due to slow retrieval but likely involve a combination of factors, including slower processing speeds (higher $\mu$ and $\delta$), increased noise or intermittent failures (higher $\sigma$), and potentially reduced availability of target items (lower $\theta$).
- Interference effects are key: The study highlights how interference between similar items (e.g., "girl" and "boy" in sentence (1b)) impacts retrieval probability ($\theta$), and how this effect can be more pronounced in IWA.
- Model comparison is nuanced: While the MDA provides a better qualitative fit for slow incorrect responses, the quantitative comparison (Bayes factor) was inconclusive, suggesting that the MDA might be over-parameterized or that the data doesn't strongly differentiate between the models.
Key Results
- IWA exhibit slower processing: The MDA confirmed that IWA have higher mean listening times ($\mu$) and a greater cost of backtracking ($\delta$) compared to controls.
- IWA have lower initial retrieval probability: The probability of successfully retrieving the target at the first attempt ($\theta$) was estimated to be lower for IWA, especially in subject-control constructions, indicating a greater interference effect.
- Backtracking is less frequent in IWA: The probability of backtracking ($P_b$) was significantly lower for IWA than for controls.
- MDA better accounts for slow incorrect responses: Posterior predictive checks showed that the MDA provided a better fit for the distribution of listening times for incorrect responses compared to the base DA model, although some discrepancies remained.
- Bayes Factor is inconclusive: The Bayes factor of 2 for DA over MDA suggests that the models provide a similar quantitative fit to the data, meaning the added complexity of the MDA was not decisively justified by the current dataset.
Limitations and Future Work
- Inconclusive model comparison: The primary limitation is the inconclusive Bayes factor, which prevents a definitive choice between the DA and MDA models based on quantitative fit alone. The authors suggest the MDA might be over-parameterized, particularly with group adjustments for all parameters.
- Investigating parameter interactions: Future work could explore removing group adjustments for specific parameters (like $\mu, P_b, \delta$) one at a time to see if it improves the fit or clarifies parameter roles.
- Larger datasets and different populations: The authors suggest testing the MDA on larger datasets and potentially different populations with aphasia to better estimate the parameter $\theta_b$ and confirm its utility.
- Exploring other deficit theories: While the MDA aligns with slow syntax, resource reduction, and intermittent deficiencies, further work could explore its interaction with delayed lexical access more deeply.
Summary updated