Paola Merlo

Paola Merlo

We motivate and formally define a new task for fine-tuning rule-like generalization in large language models. It is conjectured that the shortcomings of current LLMs are due to a lack of ability to generalize. It has been argued that, instead, humans are better at generalization because they have a tendency at extracting rules from complex data. We try to recreate this tendency to rule-based generalization. When exposed to tests of analytic intelligence, for example, the visual RAVEN IQ test, human problem-solvers identify the relevant objects in the picture and their relevant attributes and reason based on rules applied to these objects and attributes. Based on the induced rules, they are able to provide a solution to the test. We propose a task that translates this IQ task into language. In this paper, we provide the formal specification for the task and the generative process of its datasets.

Paola Merlo, Aixiu An, Maria A. Rodriguez

Current successes of machine learning architectures are based on computationally expensive algorithms and prohibitively large amounts of data. We need to develop tasks and data to train networks to reach more complex and more compositional skills. In this paper, we illustrate Blackbird's language matrices (BLMs), a novel grammatical dataset developed to test a linguistic variant of Raven's progressive matrices, an intelligence test usually based on visual stimuli. The dataset consists of 44800 sentences, generatively constructed to support investigations of current models' linguistic mastery of grammatical agreement rules and their ability to generalise them. We present the logic of the dataset, the method to automatically construct data on a large scale and the architecture to learn them. Through error analysis and several experiments on variations of the dataset, we demonstrate that this language task and the data that instantiate it provide a new challenging testbed to understand generalisation and abstraction.

Malvina Nissim, Danilo Croce, Viviana Patti et al.

The rapid progress of Large Language Models (LLMs) has transformed natural language processing and broadened its impact across research and society. Yet, systematic evaluation of these models, especially for languages beyond English, remains limited. "Challenging the Abilities of LAnguage Models in ITAlian" (CALAMITA) is a large-scale collaborative benchmarking initiative for Italian, coordinated under the Italian Association for Computational Linguistics. Unlike existing efforts that focus on leaderboards, CALAMITA foregrounds methodology: it federates more than 80 contributors from academia, industry, and the public sector to design, document, and evaluate a diverse collection of tasks, covering linguistic competence, commonsense reasoning, factual consistency, fairness, summarization, translation, and code generation. Through this process, we not only assembled a benchmark of over 20 tasks and almost 100 subtasks, but also established a centralized evaluation pipeline that supports heterogeneous datasets and metrics. We report results for four open-weight LLMs, highlighting systematic strengths and weaknesses across abilities, as well as challenges in task-specific evaluation. Beyond quantitative results, CALAMITA exposes methodological lessons: the necessity of fine-grained, task-representative metrics, the importance of harmonized pipelines, and the benefits and limitations of broad community engagement. CALAMITA is conceived as a rolling benchmark, enabling continuous integration of new tasks and models. This makes it both a resource -- the most comprehensive and diverse benchmark for Italian to date -- and a framework for sustainable, community-driven evaluation. We argue that this combination offers a blueprint for other languages and communities seeking inclusive and rigorous LLM evaluation practices.

Vivi Nastase, Chunyang Jiang, Giuseppe Samo et al.

In this paper, our goal is to investigate to what degree multilingual pretrained language models capture cross-linguistically valid abstract linguistic representations. We take the approach of developing curated synthetic data on a large scale, with specific properties, and using them to study sentence representations built using pretrained language models. We use a new multiple-choice task and datasets, Blackbird Language Matrices (BLMs), to focus on a specific grammatical structural phenomenon -- subject-verb agreement across a variety of sentence structures -- in several languages. Finding a solution to this task requires a system detecting complex linguistic patterns and paradigms in text representations. Using a two-level architecture that solves the problem in two steps -- detect syntactic objects and their properties in individual sentences, and find patterns across an input sequence of sentences -- we show that despite having been trained on multilingual texts in a consistent manner, multilingual pretrained language models have language-specific differences, and syntactic structure is not shared, even across closely related languages.

Vivi Nastase, Giuseppe Samo, Chunyang Jiang et al.

We investigate to what degree existing LLMs encode abstract linguistic information in Italian in a multi-task setting. We exploit curated synthetic data on a large scale -- several Blackbird Language Matrices (BLMs) problems in Italian -- and use them to study how sentence representations built using pre-trained language models encode specific syntactic and semantic information. We use a two-level architecture to model separately a compression of the sentence embeddings into a representation that contains relevant information for a task, and a BLM task. We then investigate whether we can obtain compressed sentence representations that encode syntactic and semantic information relevant to several BLM tasks. While we expected that the sentence structure -- in terms of sequence of phrases/chunks -- and chunk properties could be shared across tasks, performance and error analysis show that the clues for the different tasks are encoded in different manners in the sentence embeddings, suggesting that abstract linguistic notions such as constituents or thematic roles does not seem to be present in the pretrained sentence embeddings.

Vivi Nastase, Paola Merlo

Analyses of transformer-based models have shown that they encode a variety of linguistic information from their textual input. While these analyses have shed a light on the relation between linguistic information on one side, and internal architecture and parameters on the other, a question remains unanswered: how is this linguistic information reflected in sentence embeddings? Using datasets consisting of sentences with known structure, we test to what degree information about chunks (in particular noun, verb or prepositional phrases), such as grammatical number, or semantic role, can be localized in sentence embeddings. Our results show that such information is not distributed over the entire sentence embedding, but rather it is encoded in specific regions. Understanding how the information from an input text is compressed into sentence embeddings helps understand current transformer models and help build future explainable neural models.

Giuseppe Samo, Paola Merlo

We investigate how transformer models represent complex verb paradigms in Turkish and Modern Hebrew, concentrating on how tokenization strategies shape this ability. Using the Blackbird Language Matrices task on natural data, we show that for Turkish -- with its transparent morphological markers -- both monolingual and multilingual models succeed, either when tokenization is atomic or when it breaks words into small subword units. For Hebrew, instead, monolingual and multilingual models diverge. A multilingual model using character-level tokenization fails to capture the language non-concatenative morphology, but a monolingual model with morpheme-aware segmentation performs well. Performance improves on more synthetic datasets, in all models.

Giuseppe Samo, Paola Merlo

Large language models (LLMs) have shown remarkable performance across various sentence-based linguistic phenomena, yet their ability to capture cross-sentence paradigmatic patterns, such as verb alternations, remains underexplored. In this work, we present curated paradigm-based datasets for four languages, designed to probe systematic cross-sentence knowledge of verb alternations (change-of-state and object-drop constructions in English, German and Italian, and Hebrew binyanim). The datasets comprise thousands of the Blackbird Language Matrices (BLMs) problems. The BLM task -- an RPM/ARC-like task devised specifically for language -- is a controlled linguistic puzzle where models must select the sentence that completes a pattern according to syntactic and semantic rules. We introduce three types of templates varying in complexity and apply linguistically-informed data augmentation strategies across synthetic and natural data. We provide simple baseline performance results across English, Italian, German, and Hebrew, that demonstrate the diagnostic usefulness of the datasets.

Giuseppe Samo, Paola Merlo

This study compares the impact of natural and synthetic data on training and evaluating large language models (LLMs), using the case of passive verb alternation in French and Italian. We use Blackbird Language Matrices (BLMs), structured datasets designed to probe linguistic knowledge of underlying patterns across sentence sets. We compare structured templates instantiated with natural sentences extracted from Universal Dependencies to structured templates of synthetic sentences. Experiments show that while models achieve ceiling performance when trained and tested on synthetic datasets, they do not reliably generalize to natural sentences. In contrast, models trained on natural data exhibit robust performance across both natural and synthetic test suites, demonstrating their superior ability to capture abstract linguistic patterns. These results corroborate the value of natural data and of structured set ups in linguistic evaluation for probing LLMs' syntactic and semantic knowledge.

Paola Merlo, Chunyang Jiang, Giuseppe Samo et al.

This article describes a novel language task, the Blackbird Language Matrices (BLM) task, inspired by intelligence tests, and illustrates the BLM datasets, their construction and benchmarking, and targeted experiments on chunking and systematicity. BLMs are multiple-choice problems, structured at multiple levels: within each sentence, across the input sequence, within each candidate answer. Because of their rich structure, these curated, but naturalistic datasets are key to answer some core questions about current large language models abilities: do LLMs detect linguistic objects and their properties? Do they detect and use systematic patterns across sentences? Are they more prone to linguistic or reasoning errors, and how do these interact? We show that BLMs, while challenging, can be solved at good levels of performance, in more than one language, with simple baseline models or, at better performance levels, with more tailored models. We show that their representations contain the grammatical objects and attributes relevant to solve a linguistic task. We also show that these solutions are reached by detecting systematic patterns across sentences. The paper supports the point of view that curated, structured datasets support multi-faceted investigations of properties of language and large language models. Because they present a curated, articulated structure, because they comprise both learning contexts and expected answers, and because they are partly built by hand, BLMs fall in the category of datasets that can support explainability investigations, and be useful to ask why large language models behave the way they do.

Chunyang Jiang, Paola Merlo

Large language models achieve strong performance through training on vast datasets. Can analogical paradigm organization enable lightweight models to match this performance with minimal data? We develop a computational approach implementing three cognitive-inspired principles: analogical structure, contrastive learning, and minimal contextual cues. We test this approach with structured completion tasks where models identify correct sentence completions from analogical patterns with contrastive alternatives. Training lightweight models (BERT+CNN, $0.5M$ parameters) on only one hundred structured examples of English causative/inchoative alternations achieves $F1=0.95$, outperforming zero-shot \texttt{GPT-o3} ($F1=0.87$). Ablation studies confirm that analogical organization and contrastive structure improve performance, consistently surpassing randomly shuffled baselines across architectures. Cross-phenomenon validation using unspecified object alternations replicates these efficiency gains, confirming approach robustness. Our results show that analogical paradigm organization enables competitive linguistic rule learning with orders of magnitude less data than conventional approaches require.

Vivi Nastase, Paola Merlo

Sentence embeddings induced with various transformer architectures encode much semantic and syntactic information in a distributed manner in a one-dimensional array. We investigate whether specific grammatical information can be accessed in these distributed representations. Using data from a task developed to test rule-like generalizations, our experiments on detecting subject-verb agreement yield several promising results. First, we show that while the usual sentence representations encoded as one-dimensional arrays do not easily support extraction of rule-like regularities, a two-dimensional reshaping of these vectors allows various learning architectures to access such information. Next, we show that various architectures can detect patterns in these two-dimensional reshaped sentence embeddings and successfully learn a model based on smaller amounts of simpler training data, which performs well on more complex test data. This indicates that current sentence embeddings contain information that is regularly distributed, and which can be captured when the embeddings are reshaped into higher dimensional arrays. Our results cast light on representations produced by language models and help move towards developing few-shot learning approaches.

Vivi Nastase, Paola Merlo

Transformer models learn to encode and decode an input text, and produce contextual token embeddings as a side-effect. The mapping from language into the embedding space maps words expressing similar concepts onto points that are close in the space. In practice, the reverse implication is also assumed: words corresponding to close points in this space are similar or related, those that are further are not. Does closeness in the embedding space extend to shared properties for sentence embeddings? We present an investigation of sentence embeddings and show that the geometry of their embedding space is not predictive of their relative performances on a variety of tasks. We compute sentence embeddings in three ways: as averaged token embeddings, as the embedding of the special [CLS] token, and as the embedding of a random token from the sentence. We explore whether there is a correlation between the distance between sentence embedding variations and their performance on linguistic tasks, and whether despite their distances, they do encode the same information in the same manner. The results show that the cosine similarity -- which treats dimensions shallowly -- captures (shallow) commonalities or differences between sentence embeddings, which are not predictive of their performance on specific tasks. Linguistic information is rather encoded in weighted combinations of different dimensions, which are not reflected in the geometry of the sentence embedding space.

Vivi Nastase, Paola Merlo

Sentence embeddings from transformer models encode in a fixed length vector much linguistic information. We explore the hypothesis that these embeddings consist of overlapping layers of information that can be separated, and on which specific types of information -- such as information about chunks and their structural and semantic properties -- can be detected. We show that this is the case using a dataset consisting of sentences with known chunk structure, and two linguistic intelligence datasets, solving which relies on detecting chunks and their grammatical number, and respectively, their semantic roles, and through analyses of the performance on the tasks and of the internal representations built during learning.

Vivi Nastase, Paola Merlo

Sentence and word embeddings encode structural and semantic information in a distributed manner. Part of the information encoded -- particularly lexical information -- can be seen as continuous, whereas other -- like structural information -- is most often discrete. We explore whether we can compress transformer-based sentence embeddings into a representation that separates different linguistic signals -- in particular, information relevant to subject-verb agreement and verb alternations. We show that by compressing an input sequence that shares a targeted phenomenon into the latent layer of a variational autoencoder-like system, the targeted linguistic information becomes more explicit. A latent layer with both discrete and continuous components captures better the targeted phenomena than a latent layer with only discrete or only continuous components. These experiments are a step towards separating linguistic signals from distributed text embeddings and linking them to more symbolic representations.

Haozhou Wang, James Henderson, Paola Merlo

Generative adversarial networks (GANs) have succeeded in inducing cross-lingual word embeddings -- maps of matching words across languages -- without supervision. Despite these successes, GANs' performance for the difficult case of distant languages is still not satisfactory. These limitations have been explained by GANs' incorrect assumption that source and target embedding spaces are related by a single linear mapping and are approximately isomorphic. We assume instead that, especially across distant languages, the mapping is only piece-wise linear, and propose a multi-adversarial learning method. This novel method induces the seed cross-lingual dictionary through multiple mappings, each induced to fit the mapping for one subspace. Our experiments on unsupervised bilingual lexicon induction show that this method improves performance over previous single-mapping methods, especially for distant languages.

Haozhou Wang, James Henderson, Paola Merlo

Distributed representations of words which map each word to a continuous vector have proven useful in capturing important linguistic information not only in a single language but also across different languages. Current unsupervised adversarial approaches show that it is possible to build a mapping matrix that align two sets of monolingual word embeddings together without high quality parallel data such as a dictionary or a sentence-aligned corpus. However, without post refinement, the performance of these methods' preliminary mapping is not good, leading to poor performance for typologically distant languages. In this paper, we propose a weakly-supervised adversarial training method to overcome this limitation, based on the intuition that mapping across languages is better done at the concept level than at the word level. We propose a concept-based adversarial training method which for most languages improves the performance of previous unsupervised adversarial methods, especially for typologically distant language pairs.