Abstract
Behavioral studies show that humans process concrete words more similarly across languages than abstract words. This suggests that in bilingual brains, semantic representations may be more similar across languages for concrete concepts than for abstract concepts, but existing neuroimaging evidence is inconclusive. Here, we analyzed functional magnetic resonance imaging (fMRI) data from fluent Chinese-English bilinguals who read several hours of naturalistic narratives in both languages. We used encoding models to estimate voxelwise tuning towards concrete and abstract concepts in each language separately. We then quantified the similarity of cortical semantic representations across languages. First, we find that the cortical organization of concreteness tuning is consistent across languages. Second, semantic representations are similar across languages for both concrete-tuned voxels and abstract-tuned voxels. Third, we find that for abstract concepts, cross-language similarity of semantic representations may be driven by emotionality. Overall, these findings reveal how concreteness affects semantic representations in bilingual brains.
Introduction
Psycholinguistic models have long suggested that concrete words (like "apple") are easier to translate and process across languages than abstract words (like "justice"). The assumption was that concrete concepts have a shared physical reality, making their neural "blueprints" similar across languages, whereas abstract concepts might be more linguistically or culturally idiosyncratic.
However, previous neuroimaging studies provided conflicting results. Many used "controlled stimuli"—isolated words presented out of context—which don't capture how the brain processes language in the real world. Furthermore, most studies treated concreteness as a binary (either concrete or abstract), ignoring the nuanced spectrum of human thought. This paper seeks to map the continuous dimension of concreteness in the bilingual brain using naturalistic stimuli (stories) to see how language-invariant these representations truly are.
Methods and Experimental Design
The researchers employed voxelwise encoding models, a powerful computational framework used to map how specific features (like word meanings) relate to brain activity in individual "voxels" (3D pixels of brain tissue). Six fluent Chinese-English bilinguals read several hours of narratives from The Moth Radio Hour, manually translated and aligned between English and Simplified Chinese.
The team used fastText and mBERT (multilingual BERT) to convert words into high-dimensional vectors. They identified a "concreteness dimension" by creating a vector in the embedding space that spans from highly abstract to highly concrete words. For every voxel in the brain, they estimated "concreteness tuning" (how much a voxel prefers concrete vs. abstract concepts) and "cross-language similarity" (how well the neural weights for English predicted the weights for Chinese).
Key Results
Consistent Cortical Geography
The study revealed that the "map" of concreteness is remarkably stable across languages. Posterior areas, such as the lateral and medial parietal cortices, are tuned to concrete concepts, likely due to their proximity to visual and sensory systems. In contrast, abstract concepts consistently engage traditional language hubs like Broca’s area and the superior temporal sulcus. Concreteness tuning was strongly correlated between English and Chinese ($r = 0.83, p < .001$).
The "Heart-Shaped" Similarity Pattern
One of the most striking findings was the relationship between concreteness and cross-language similarity. Cross-language similarity is high for both ends of the spectrum. Both highly concrete and highly abstract concepts are represented similarly across languages, while "neutral" words in the middle show less consistency. This creates a distinctive "heart-shaped" distribution in the scatter plots comparing similarity to tuning.
Emotion as the Anchor for Abstraction
Why are abstract concepts so similar across languages if they lack a physical referent? The data suggests emotionality is the key. Abstract concepts that are highly emotional (e.g., "traumatized," "feelings") showed much higher cross-language similarity than "cold" abstract concepts. For abstract-tuned voxels, there was a moderate positive correlation ($r = 0.44$) between a voxel's emotionality tuning and its cross-language similarity, confirming that emotion is specifically important for stabilizing abstract representations.
Discussion and Future Work
While the location of the tuning was consistent, the magnitude was stronger in Chinese (slope = 1.22). The researchers suggest this might be due to the logographic nature of Chinese characters or the fact that Chinese was the participants' native language (L1).
Limitations include the small sample size (n=6), though this is common in "deep" fMRI studies where each subject provides hours of data. Future research could explore a broader range of language pairs to isolate the effects of writing systems from the effects of bilingual proficiency. Additionally, while naturalistic, the study focused on reading; future work could compare these representations during auditory speech or spontaneous production to further generalize these findings.