Marco Marelli
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Marco Marelli has a total of 500 co-authors. Here are the top co-authors by number of shared publications:
- Claudio Luzzatti
University of Milan Bicocca
11 shared publications - Simona Amenta
Department of Psychology, University of Milano-Bicocca, Milano, Italy
10 shared publications - Davide Crepaldi
Department of Cognitive Neuroscience, International School for Advanced Studies (SISSA), Trieste, Italy
6 shared publications - Fritz Günther
Department of Psychology, Humboldt-Universität zu Berlin
6 shared publications Marco Baroni
Center for Mind/Brain Sciences, University of Trento
5 shared publications- Francesca Foppolo
Department of Psychology, University of Milan-Bicocca
4 shared publications
Latest Publications
- Open Access
- 22 September 2026
- Published by Springer Nature in Behavior Research Methods
- Vol. 58 (11) , 297
The English Pseudoword Lexicon Project (EPLeP): A resource for investigating pseudoword processing
In recent years, psycholinguistics has seen an increase in research focused on the processing of written stimuli that are not attested in the lexicon of a given language. We introduce the English Pseudoword Lexicon Project (EPLeP), a new database of lexical decision data for English pseudowords, nonwords, and words. While large-scale behavioral data have been collected for lexical and non-lexical items in English and other languages, none of these datasets have specifically focused on pseudowords. The EPLeP database fills this gap by considering three types of English-like non-lexical stimuli: (1) morphologically simple pseudowords generated from existing English words; (2) morphologically complex pseudowords obtained by combining existing stems and existing suffixes; and (3) nonwords, not following English graphotactics. This registered report provides an extensive overview of the steps adopted in producing non-lexical stimuli and the criteria for their selection. We collected data for almost 13,500 non-lexical and 4,400 lexical stimuli from 1,416 English-speaking participants across four online studies. Analyses of behavioral responses show that morphologically complex pseudowords trigger the deepest level of processing, whereas nonwords elicit the shallowest processing, reflecting a crucial difference in how these stimuli affect responses to words. This database will be a valuable tool for researchers interested in both pseudoword and word processing, enabling them to evaluate hypotheses by analyzing the collected behavioral data or using them to select stimuli for future studies. The Stage 1 protocol was accepted in principle on July 4, 2025 ( https://osf.io/cz9rh/ )
- 22 September 2026
- Published by Springer Nature in Behavior Research Methods
- Vol. 58 (11) , 297
Open AccessThe English Pseudoword Lexicon Project (EPLeP): A resource for investigating pseudoword processing
In recent years, psycholinguistics has seen an increase in research focused on the processing of written stimuli that are not attested in the lexicon of a given language. We introduce the English Pseudoword Lexicon Project (EPLeP), a new database of lexical decision data for English pseudowords, nonwords, and words. While large-scale behavioral data have been collected for lexical and non-lexical items in English and other languages, none of these datasets have specifically focused on pseudowords. The EPLeP database fills this gap by considering three types of English-like non-lexical stimuli: (1) morphologically simple pseudowords generated from existing English words; (2) morphologically complex pseudowords obtained by combining existing stems and existing suffixes; and (3) nonwords, not following English graphotactics. This registered report provides an extensive overview of the steps adopted in producing non-lexical stimuli and the criteria for their selection. We collected data for almost 13,500 non-lexical and 4,400 lexical stimuli from 1,416 English-speaking participants across four online studies. Analyses of behavioral responses show that morphologically complex pseudowords trigger the deepest level of processing, whereas nonwords elicit the shallowest processing, reflecting a crucial difference in how these stimuli affect responses to words. This database will be a valuable tool for researchers interested in both pseudoword and word processing, enabling them to evaluate hypotheses by analyzing the collected behavioral data or using them to select stimuli for future studies. The Stage 1 protocol was accepted in principle on July 4, 2025 ( https://osf.io/cz9rh/ )
- Open Access
- 10 September 2026
- Published by Elsevier in SSRN Electronic Journal
Meaning beyond perceptual experience: sensory grounding of pseudowords through sub-lexical distributional cues
Embodied theories of cognition propose that conceptual representations are grounded in sensorimotor experience. Yet, such experience can also be activated for referents we have never directly experienced, raising the question of whether experiential features might be predicted from language alone. Pseudowords represent a critical test case for this question: given the absence of both perceptual and linguistic experience, they should, in principle, be impossible to ground in a word-centered approach. In this study, we hence explored whether pseudowords can evoke perceptual features through sub-lexical distributional cues. We first trained a model mapping distributed sub-lexical representations of existing words to their corresponding perceptual norms, then used it to predict perceptual estimates for 27,136 pseudowords across six modalities (vision, touch, hearing, taste, smell, and interoception). In Experiment 1, we tested whether composite measures of perceptual strength based on such perceptual estimates could predict reaction times in a lexical decision task. In Experiment 2, we collected perceptual ratings for a subset of 215 pseudowords using Best-Worst Scaling and compared these explicit intuitions to model predictions. Results showed that perceptual strength estimates influence linguistic processing, and demonstrated a modality-specific alignment between model predictions and human judgments, with the strongest correlations found for interoception, followed by taste and hearing. Together, these findings suggest that sub-lexical distributions can capture perceptual dimensions in pseudowords, supporting the idea that linguistic cues are crucial to infer meaning and for the subsequent activation of perceptual experiences.
- 10 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessMeaning beyond perceptual experience: sensory grounding of pseudowords through sub-lexical distributional cues
Embodied theories of cognition propose that conceptual representations are grounded in sensorimotor experience. Yet, such experience can also be activated for referents we have never directly experienced, raising the question of whether experiential features might be predicted from language alone. Pseudowords represent a critical test case for this question: given the absence of both perceptual and linguistic experience, they should, in principle, be impossible to ground in a word-centered approach. In this study, we hence explored whether pseudowords can evoke perceptual features through sub-lexical distributional cues. We first trained a model mapping distributed sub-lexical representations of existing words to their corresponding perceptual norms, then used it to predict perceptual estimates for 27,136 pseudowords across six modalities (vision, touch, hearing, taste, smell, and interoception). In Experiment 1, we tested whether composite measures of perceptual strength based on such perceptual estimates could predict reaction times in a lexical decision task. In Experiment 2, we collected perceptual ratings for a subset of 215 pseudowords using Best-Worst Scaling and compared these explicit intuitions to model predictions. Results showed that perceptual strength estimates influence linguistic processing, and demonstrated a modality-specific alignment between model predictions and human judgments, with the strongest correlations found for interoception, followed by taste and hearing. Together, these findings suggest that sub-lexical distributions can capture perceptual dimensions in pseudowords, supporting the idea that linguistic cues are crucial to infer meaning and for the subsequent activation of perceptual experiences.
- Open Access
- 8 June 2026
- Published by Springer Nature in Research Square
Emergent Sensorimotor Representation in Large Language Models: Revisiting Human-machine Alignment on Sensorimotor Intuitions of Concepts
Sensorimotor knowledge is vital for biological systems to build conceptual representations and interact with the physical world, yet whether and to what extent large language models (LLMs) can compensate for absent grounding remains an open question. Here we systematically assess ten LLMs on sensorimotor alignment with humans at both dimensional correlations and geometric relations. We collected approximately 1.17 million ratings for 4,442 English lexicalised concepts across 11 sensorimotor dimensions from six frontier LLMs, together with four models from prior work, and analysed them at both the population and individual levels. We find that frontier LLMs show substantial alignment with human sensorimotor intuitions, in sharp contrast to the negligible alignment of earlier models, with text-only DeepSeek-v3 matching or exceeding multimodal models. To control for data contamination, we further validated these findings on a novel Chinese sensorimotor dataset of 3,000 concepts rated by 291 native speakers, and found consistent alignment patterns across both languages. When considering individual-level ratings, frontier LLMs meet or exceed the human inter-rater baseline on 9 of 11 dimensions, except for chemosensory dimensions where alignment falls below this baseline. Together, our results suggest that sensorimotor representations can substantially emerge in frontier LLMs through the scaffold of language, providing empirical evidence for hybrid accounts of conceptual representation while highlighting persistent challenges in chemosensory dimensions and the attenuation of individual variability.
- 8 June 2026
- Published by Springer Nature in Research Square
Open AccessEmergent Sensorimotor Representation in Large Language Models: Revisiting Human-machine Alignment on Sensorimotor Intuitions of Concepts
Sensorimotor knowledge is vital for biological systems to build conceptual representations and interact with the physical world, yet whether and to what extent large language models (LLMs) can compensate for absent grounding remains an open question. Here we systematically assess ten LLMs on sensorimotor alignment with humans at both dimensional correlations and geometric relations. We collected approximately 1.17 million ratings for 4,442 English lexicalised concepts across 11 sensorimotor dimensions from six frontier LLMs, together with four models from prior work, and analysed them at both the population and individual levels. We find that frontier LLMs show substantial alignment with human sensorimotor intuitions, in sharp contrast to the negligible alignment of earlier models, with text-only DeepSeek-v3 matching or exceeding multimodal models. To control for data contamination, we further validated these findings on a novel Chinese sensorimotor dataset of 3,000 concepts rated by 291 native speakers, and found consistent alignment patterns across both languages. When considering individual-level ratings, frontier LLMs meet or exceed the human inter-rater baseline on 9 of 11 dimensions, except for chemosensory dimensions where alignment falls below this baseline. Together, our results suggest that sensorimotor representations can substantially emerge in frontier LLMs through the scaffold of language, providing empirical evidence for hybrid accounts of conceptual representation while highlighting persistent challenges in chemosensory dimensions and the attenuation of individual variability.
- Open Access
Brain signatures of semantic activation for words that do not exist
The experience of making sense of novel words is ubiquitous in human communication. Still, novel words have traditionally been considered as meaningless by cognitive research. Here we combined behavioral, univariate and multivariate fMRI techniques, and computational modelling to explore whether and how novel words activate the neurocognitive hallmarks of semantic processing triggered by existing words, and whether this process is influenced by the presence of familiar functional linguistic elements, i.e., morphemes. We observed that semantic activation for novel words is comparable to that of existing words, provided that novel words contain a concatenation of identifiable morphemes. In addition, representational similarity analysis highlighted that existing words and novel words containing morphemes (but not novel words not containing morphemes) can activate fine-grained semantic representations. These results suggest that the difference in the neurocognitive underpinnings of semantic processing for existing and novel words might be quantitative rather than qualitative and based on how reliably linguistic form points to meaning.
Open AccessBrain signatures of semantic activation for words that do not exist
The experience of making sense of novel words is ubiquitous in human communication. Still, novel words have traditionally been considered as meaningless by cognitive research. Here we combined behavioral, univariate and multivariate fMRI techniques, and computational modelling to explore whether and how novel words activate the neurocognitive hallmarks of semantic processing triggered by existing words, and whether this process is influenced by the presence of familiar functional linguistic elements, i.e., morphemes. We observed that semantic activation for novel words is comparable to that of existing words, provided that novel words contain a concatenation of identifiable morphemes. In addition, representational similarity analysis highlighted that existing words and novel words containing morphemes (but not novel words not containing morphemes) can activate fine-grained semantic representations. These results suggest that the difference in the neurocognitive underpinnings of semantic processing for existing and novel words might be quantitative rather than qualitative and based on how reliably linguistic form points to meaning.
- Open Access
- 17 January 2026
- Published by Elsevier in Journal of Memory and Language
- Vol. 148
Morphemes in the wild: Modelling affix learning from the noisy landscape of natural text
- 17 January 2026
- Published by Elsevier in Journal of Memory and Language
- Vol. 148
Open AccessMorphemes in the wild: Modelling affix learning from the noisy landscape of natural text
- 24 September 2025
- Published by Springer Nature in Nature Human Behaviour
- Vol. 10 (1) , 182-201
Measuring the semantic priming effect across many languages
Semantic priming has been studied for nearly 50 years across various experimental manipulations and theoretical frameworks. Although previous studies provide insight into the cognitive underpinnings of semantic representations, they have suffered from small sample sizes and a lack of linguistic and cultural diversity. In this Registered Report, we measured the size and the variability of the semantic priming effect across 19 languages (n = 25,163 participants analysed) by creating the largest available database of semantic priming values using an adaptive sampling procedure. We found evidence for semantic priming in terms of differences in response latencies between related word-pair conditions and unrelated word-pair conditions. Model comparisons showed that the inclusion of a random intercept for language improved model fit, providing support for variability in semantic priming across languages. This study highlights the robustness and variability of semantic priming across languages and provides a rich, linguistically diverse dataset for further analysis. The Stage 1 protocol for this Registered Report was accepted in principle on 15 July 2022. The protocol, as accepted by the journal, can be found at https://osf.io/u5bp6 (registration) or https://osf.io/q4fjy (preprint version 6, 31 May 2022).
- 24 September 2025
- Published by Springer Nature in Nature Human Behaviour
- Vol. 10 (1) , 182-201
Measuring the semantic priming effect across many languages
Semantic priming has been studied for nearly 50 years across various experimental manipulations and theoretical frameworks. Although previous studies provide insight into the cognitive underpinnings of semantic representations, they have suffered from small sample sizes and a lack of linguistic and cultural diversity. In this Registered Report, we measured the size and the variability of the semantic priming effect across 19 languages (n = 25,163 participants analysed) by creating the largest available database of semantic priming values using an adaptive sampling procedure. We found evidence for semantic priming in terms of differences in response latencies between related word-pair conditions and unrelated word-pair conditions. Model comparisons showed that the inclusion of a random intercept for language improved model fit, providing support for variability in semantic priming across languages. This study highlights the robustness and variability of semantic priming across languages and provides a rich, linguistically diverse dataset for further analysis. The Stage 1 protocol for this Registered Report was accepted in principle on 15 July 2022. The protocol, as accepted by the journal, can be found at https://osf.io/u5bp6 (registration) or https://osf.io/q4fjy (preprint version 6, 31 May 2022).
- Open Access
- 1 August 2025
- Published by Oxford University Press (OUP) in Cerebral Cortex
- Vol. 35 (8)
Compositionality in the semantic network: a model-driven representational similarity analysis
Semantic composition allows us to construct complex meanings (e.g., “dog house”, “house dog”) from simpler constituents (“dog”, “house”). Neuroimaging studies have often relied on high-level contrasts (e.g., meaningful > non-meaningful phrases) to identify brain regions sensitive to composition. However, such an approach is less apt at addressing how composition is carried out, namely what functions best characterize constituents integration. Here, we rely on simple computational models to explicitly characterize alternative compositional operations, and use representational similarity analysis to compare models to target regions of interest. We re-analyze fMRI data aggregated from four published studies (N = 85), all employing two-word combinations but differing in task requirements. Confirmatory and exploratory analyses reveal compositional representations in the left inferior frontal gyrus (BA45), even when the task did not require semantic access. These results suggest that BA45 represents combinatorial information automatically across task demands, and further characterize composition as the (symmetric) intersection of constituent features. Additionally, a cluster of compositional representations emerges in the left middle superior temporal sulcus, while semantic, but not compositional, representations are observed in the left angular gyrus. Overall, our work clarifies which brain regions represent semantic information compositionally across contexts and tasks, and qualifies which operations best describe composition.
- 1 August 2025
- Published by Oxford University Press (OUP) in Cerebral Cortex
- Vol. 35 (8)
Open AccessCompositionality in the semantic network: a model-driven representational similarity analysis
Semantic composition allows us to construct complex meanings (e.g., “dog house”, “house dog”) from simpler constituents (“dog”, “house”). Neuroimaging studies have often relied on high-level contrasts (e.g., meaningful > non-meaningful phrases) to identify brain regions sensitive to composition. However, such an approach is less apt at addressing how composition is carried out, namely what functions best characterize constituents integration. Here, we rely on simple computational models to explicitly characterize alternative compositional operations, and use representational similarity analysis to compare models to target regions of interest. We re-analyze fMRI data aggregated from four published studies (N = 85), all employing two-word combinations but differing in task requirements. Confirmatory and exploratory analyses reveal compositional representations in the left inferior frontal gyrus (BA45), even when the task did not require semantic access. These results suggest that BA45 represents combinatorial information automatically across task demands, and further characterize composition as the (symmetric) intersection of constituent features. Additionally, a cluster of compositional representations emerges in the left middle superior temporal sulcus, while semantic, but not compositional, representations are observed in the left angular gyrus. Overall, our work clarifies which brain regions represent semantic information compositionally across contexts and tasks, and qualifies which operations best describe composition.
- 4 June 2025
- Published by Springer Nature in Nature Human Behaviour
- Vol. 9 (7) , 1339-1341
High variability in LLMs’ analogical reasoning
- 4 June 2025
- Published by Springer Nature in Nature Human Behaviour
- Vol. 9 (7) , 1339-1341
High variability in LLMs’ analogical reasoning
- Open Access
- 3 June 2025
- Published by Elsevier in Neuropsychologia
- Vol. 217, 109190
A map of words: Retrieving the spatial layout of medium-scale geographical maps through distributional semantics
- 3 June 2025
- Published by Elsevier in Neuropsychologia
- Vol. 217, 109190
Open AccessA map of words: Retrieving the spatial layout of medium-scale geographical maps through distributional semantics