Alessandro Lenci
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Alessandro Lenci has a total of 154 co-authors. Here are the top co-authors by number of shared publications:
- Emmanuele Chersoni
Department of Language Science and Technology The Hong Kong Polytechnic University
14 shared publications - Alessandro Bondielli
University of Pisa
7 shared publications - Chu-Ren Huang
The Hong Kong Polytechnic University
7 shared publications - Enrico Santus
DSIG -Bayer Pharmaceuticals New Jersey USA
7 shared publications - Lucia C. Passaro
CoLing Lab, Dipartimento di Filologia, Letteratura e Linguistica University of Pisa (Italy)
6 shared publications - Gianluca E. Lebani5 shared publications
Latest Publications
- Open Access
- 9 July 2026
- Published by Elsevier in SSRN Electronic Journal
Does Linguistic Structure Help with Grammar Learning? Evaluating Neurosymbolic Baby Language Modeling on Linguistic Minimal Pairs
Is linguistic structure necessary for mastering natural language? For a long time, this might have sounded like a rhetorical question in linguistics, but the high level of proficiency reached by the state-of-the-art Large Language Models (LLMs), which they achieve after training on huge amounts of raw text, is casting doubts on the necessity of structure-dependent biases forlanguage learning.In our work, we test whether providing structural guidance to LLMs during the training phase is really superfluous or not: we train different versions of the same LLMs on the small scale, cognitively-plausible corpus of the BabyLM challenge, and we compare the status of their grammatical knowledge with the one of models exposed to the same training materialsenriched with structural syntactic and semantic annotations.Our evaluation on the linguistic minimal pairs of the BLIMP dataset shows that exposure to structured representations yields a consistent advantage over a purely sequential baseline, with the largest gains concentrated on phenomena involving non-local syntactic dependencies, while morphological knowledge, already well captured by distributional regularities alone, benefits comparatively little from this additional structural signal. Crucially, this advantage emerges early in training and persists even when explicit graph annotations are removed at inference time, suggesting that structural bias is not merely exploited as an external signal but becomes internalized by the model itself
- 9 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessDoes Linguistic Structure Help with Grammar Learning? Evaluating Neurosymbolic Baby Language Modeling on Linguistic Minimal Pairs
Is linguistic structure necessary for mastering natural language? For a long time, this might have sounded like a rhetorical question in linguistics, but the high level of proficiency reached by the state-of-the-art Large Language Models (LLMs), which they achieve after training on huge amounts of raw text, is casting doubts on the necessity of structure-dependent biases forlanguage learning.In our work, we test whether providing structural guidance to LLMs during the training phase is really superfluous or not: we train different versions of the same LLMs on the small scale, cognitively-plausible corpus of the BabyLM challenge, and we compare the status of their grammatical knowledge with the one of models exposed to the same training materialsenriched with structural syntactic and semantic annotations.Our evaluation on the linguistic minimal pairs of the BLIMP dataset shows that exposure to structured representations yields a consistent advantage over a purely sequential baseline, with the largest gains concentrated on phenomena involving non-local syntactic dependencies, while morphological knowledge, already well captured by distributional regularities alone, benefits comparatively little from this additional structural signal. Crucially, this advantage emerges early in training and persists even when explicit graph annotations are removed at inference time, suggesting that structural bias is not merely exploited as an external signal but becomes internalized by the model itself
- 1 July 2026
- Published by Cambridge University Press (CUP) in Behavioral and Brain Sciences
- Vol. 49, e204
On the relationship between theoretical and empirical research in linguistics
Language models (LMs) call for a theoretical rethinking grounded within linguistics itself. Rather than signalling the “end of linguistics” or merely encouraging interdisciplinarity, LMs function as an empirical testing ground for formal linguistic concepts. They prompt a renewed examination of form–meaning mapping, theoretical autonomy, and the conditions under which computational systems can genuinely inform linguistic explanation.
- 1 July 2026
- Published by Cambridge University Press (CUP) in Behavioral and Brain Sciences
- Vol. 49, e204
On the relationship between theoretical and empirical research in linguistics
Language models (LMs) call for a theoretical rethinking grounded within linguistics itself. Rather than signalling the “end of linguistics” or merely encouraging interdisciplinarity, LMs function as an empirical testing ground for formal linguistic concepts. They prompt a renewed examination of form–meaning mapping, theoretical autonomy, and the conditions under which computational systems can genuinely inform linguistic explanation.
- Open Access
- 20 June 2026
- Published by Center for Open Science in PsyArXiv
Unraveling the interplay between sensory information and semantic architecture in brain concept representation: the role of abstraction
INTRODUCTION: Understanding how abstract and concrete words are represented in the brain requires disentangling the contribution of sensorimotor information from the organization of semantic knowledge. These dimensions, here referred to as abstractness and abstraction, respectively, are typically confounded, limiting our ability to identify their distinct neural correlates.OBJECTIVES: The present study aimed to dissociate the roles of abstractness and abstraction in word processing by introducing superordinate words as a critical comparison category.METHODS: Behavioral norming and feature production tasks were used to quantify sensory-related properties and semantic architecture, operationalized in terms of generalizability, feature relevance, and representational flexibility. These measures informed computational models capturing abstractness, abstraction, sensorimotor content, and distributional similarity. Neural data were acquired using electroencephalography and analyzed with representational similarity analysis to assess model–brain correspondence over time.RESULTS: Behaviorally, superordinate words patterned with concrete words on sensory properties but with abstract words on measures of semantic architecture. In the electrophysiological data, concrete words elicited stronger negativities than both abstract and superordinate words in the N400 and N700 time windows, whereas abstract and superordinate words did not differ. Representational similarity analysis revealed that neural response patterns were best explained by the abstraction model, outperforming models based on sensorimotor or distributional information.CONCLUSION: These results indicate that the neural dynamics of word processing are primarily shaped by semantic architecture rather than by sensorimotor grounding alone. Abstraction plays a central role in structuring conceptual representations, providing a unifying account that bridges embodied and formal theories of semantic representation.
- 20 June 2026
- Published by Center for Open Science in PsyArXiv
Open AccessUnraveling the interplay between sensory information and semantic architecture in brain concept representation: the role of abstraction
INTRODUCTION: Understanding how abstract and concrete words are represented in the brain requires disentangling the contribution of sensorimotor information from the organization of semantic knowledge. These dimensions, here referred to as abstractness and abstraction, respectively, are typically confounded, limiting our ability to identify their distinct neural correlates.OBJECTIVES: The present study aimed to dissociate the roles of abstractness and abstraction in word processing by introducing superordinate words as a critical comparison category.METHODS: Behavioral norming and feature production tasks were used to quantify sensory-related properties and semantic architecture, operationalized in terms of generalizability, feature relevance, and representational flexibility. These measures informed computational models capturing abstractness, abstraction, sensorimotor content, and distributional similarity. Neural data were acquired using electroencephalography and analyzed with representational similarity analysis to assess model–brain correspondence over time.RESULTS: Behaviorally, superordinate words patterned with concrete words on sensory properties but with abstract words on measures of semantic architecture. In the electrophysiological data, concrete words elicited stronger negativities than both abstract and superordinate words in the N400 and N700 time windows, whereas abstract and superordinate words did not differ. Representational similarity analysis revealed that neural response patterns were best explained by the abstraction model, outperforming models based on sensorimotor or distributional information.CONCLUSION: These results indicate that the neural dynamics of word processing are primarily shaped by semantic architecture rather than by sensorimotor grounding alone. Abstraction plays a central role in structuring conceptual representations, providing a unifying account that bridges embodied and formal theories of semantic representation.
- Open Access
CLASS-IT: Conversational and Lecture-Aligned Small-Scale Instruction Tuning for BabyLMs
This work investigates whether small-scale LMs can benefit from instruction tuning. We compare conversational and question-answering instruction tuning datasets, applied either in a merged or sequential curriculum, using decoder-only models with 100M and 140M parameters. Evaluation spans both fine-tuning (SuperGLUE) and zero-shot (BLiMP, EWoK, WUGs, entity tracking, and psycholinguistic correlation) settings. Results show that instruction tuning yields small but consistent gains in fine-tuning scenarios, with sequential curricula outperforming merged data; however, improvements do not consistently transfer to zero-shot tasks, suggesting a trade-off between interaction-focused adaptation and broad linguistic generalization. These results highlight both the potential and the constraints of adapting human-inspired learning strategies to low-resource LMs, and point toward hybrid, curriculum-based approaches for enhancing generalization under ecological training limits.
Open AccessCLASS-IT: Conversational and Lecture-Aligned Small-Scale Instruction Tuning for BabyLMs
This work investigates whether small-scale LMs can benefit from instruction tuning. We compare conversational and question-answering instruction tuning datasets, applied either in a merged or sequential curriculum, using decoder-only models with 100M and 140M parameters. Evaluation spans both fine-tuning (SuperGLUE) and zero-shot (BLiMP, EWoK, WUGs, entity tracking, and psycholinguistic correlation) settings. Results show that instruction tuning yields small but consistent gains in fine-tuning scenarios, with sequential curricula outperforming merged data; however, improvements do not consistently transfer to zero-shot tasks, suggesting a trade-off between interaction-focused adaptation and broad linguistic generalization. These results highlight both the potential and the constraints of adapting human-inspired learning strategies to low-resource LMs, and point toward hybrid, curriculum-based approaches for enhancing generalization under ecological training limits.
- Open Access
Doing Things with Words: Rethinking Theory of Mind Simulation in Large Language Models
Language is fundamental to human cooperation, facilitating not only the exchange of information but also the coordination of actions through shared interpretations of situational contexts. This study explores whether the Generative Agent-Based Model (GABM) Concordia can effectively model Theory of Mind (ToM) within simulated real-world environments. Specifically, we assess whether this framework successfully simulates ToM abilities and whether GPT-4 can perform tasks by making genuine inferences from social context, rather than relying on linguistic memorization. Our findings reveal a critical limitation: GPT-4 frequently fails to select actions based on belief attribution, suggesting that apparent ToM-like abilities observed in previous studies may stem from shallow statistical associations rather than true reasoning. Additionally, the model struggles to generate coherent causal effects from agent actions, exposing difficulties in processing complex social interactions. These results challenge current statements about emergent ToM-like capabilities in LLMs and highlight the need for more rigorous, action-based evaluation frameworks.
Open AccessDoing Things with Words: Rethinking Theory of Mind Simulation in Large Language Models
Language is fundamental to human cooperation, facilitating not only the exchange of information but also the coordination of actions through shared interpretations of situational contexts. This study explores whether the Generative Agent-Based Model (GABM) Concordia can effectively model Theory of Mind (ToM) within simulated real-world environments. Specifically, we assess whether this framework successfully simulates ToM abilities and whether GPT-4 can perform tasks by making genuine inferences from social context, rather than relying on linguistic memorization. Our findings reveal a critical limitation: GPT-4 frequently fails to select actions based on belief attribution, suggesting that apparent ToM-like abilities observed in previous studies may stem from shallow statistical associations rather than true reasoning. Additionally, the model struggles to generate coherent causal effects from agent actions, exposing difficulties in processing complex social interactions. These results challenge current statements about emergent ToM-like capabilities in LLMs and highlight the need for more rigorous, action-based evaluation frameworks.
- 23 September 2025
- Published by Springer Nature
- p. 188-200
Investigating Time-Scales in Deep Echo State Networks for Natural Language Processing
- 23 September 2025
- Published by Springer Nature
- p. 188-200
Investigating Time-Scales in Deep Echo State Networks for Natural Language Processing
- 31 July 2025
- Published by Springer Nature in Language Resources and Evaluation
- Vol. 59 (4) , 3659-3697
Prompting encoder models for zero-shot classification: a cross-domain study in Italian
- 31 July 2025
- Published by Springer Nature in Language Resources and Evaluation
- Vol. 59 (4) , 3659-3697
Prompting encoder models for zero-shot classification: a cross-domain study in Italian
- Open Access
BAMBI: Developing Baby Language Models for Italian
This paper presents BAMBI (BAby language Models Boostrapped for Italian), a series of Baby Language Models (BabyLMs) trained on data that mimic the linguistic input received by a five-year-old Italian-speaking child. The BAMBI models are tested using a benchmark specifically designed to evaluate language models, which takes into account the amount of training input the models received. The BAMBI models are compared against a large language model (LLM) and a multimodal language model (VLM) to study the contribution of extralinguistic information for language acquisition. The results of our evaluation align with the existing literature on English language models, confirming that while reduced training data support the development of relatively robust syntactic competence, they are insufficient for fostering semantic understanding. However, the gap between the training resources (data and computation) of the BAMBI models and the LLMs is not fully reflected in their performance: despite LLMs' massive training, their performance is not much better than that of BAMBI models. This suggests that strategies beyond scaling training resources, such as data curation, inclusion of multimodal input, and other training strategies such as curriculum learning, could play a crucial role in shaping model performance.
Open AccessBAMBI: Developing Baby Language Models for Italian
This paper presents BAMBI (BAby language Models Boostrapped for Italian), a series of Baby Language Models (BabyLMs) trained on data that mimic the linguistic input received by a five-year-old Italian-speaking child. The BAMBI models are tested using a benchmark specifically designed to evaluate language models, which takes into account the amount of training input the models received. The BAMBI models are compared against a large language model (LLM) and a multimodal language model (VLM) to study the contribution of extralinguistic information for language acquisition. The results of our evaluation align with the existing literature on English language models, confirming that while reduced training data support the development of relatively robust syntactic competence, they are insufficient for fostering semantic understanding. However, the gap between the training resources (data and computation) of the BAMBI models and the LLMs is not fully reflected in their performance: despite LLMs' massive training, their performance is not much better than that of BAMBI models. This suggests that strategies beyond scaling training resources, such as data curation, inclusion of multimodal input, and other training strategies such as curriculum learning, could play a crucial role in shaping model performance.
- Open Access
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 436-444
CLASS-IT: Conversational and Lecture-Aligned Small-Scale Instruction Tuning for BabyLMs
Luca Capone, Alessandro Bondielli, Alessandro Lenci. Proceedings of the First BabyLM Workshop. 2025.
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 436-444
Open AccessCLASS-IT: Conversational and Lecture-Aligned Small-Scale Instruction Tuning for BabyLMs
Luca Capone, Alessandro Bondielli, Alessandro Lenci. Proceedings of the First BabyLM Workshop. 2025.
- Open Access
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 469-480
Compositionality and Event Retrieval in Complement Coercion: A Study of Language Models in a Low-resource Setting
Matteo Radaelli, Emmanuele Chersoni, Alessandro Lenci, Giosuè Baggio. Proceedings of the 29th Conference on Computational Natural Language Learning. 2025.
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 469-480
Open AccessCompositionality and Event Retrieval in Complement Coercion: A Study of Language Models in a Low-resource Setting
Matteo Radaelli, Emmanuele Chersoni, Alessandro Lenci, Giosuè Baggio. Proceedings of the 29th Conference on Computational Natural Language Learning. 2025.