Gemma Boleda

Roberto Dessì, Eleonora Gualdoni, Francesca Franzon et al.

We compare the 0-shot performance of a neural caption-based image retriever when given as input either human-produced captions or captions generated by a neural captioner. We conduct this comparison on the recently introduced ImageCoDe data-set (Krojer et al., 2022) which contains hard distractors nearly identical to the images to be retrieved. We find that the neural retriever has much higher performance when fed neural rather than human captions, despite the fact that the former, unlike the latter, were generated without awareness of the distractors that make the task hard. Even more remarkably, when the same neural captions are given to human subjects, their retrieval performance is almost at chance level. Our results thus add to the growing body of evidence that, even when the ``language'' of neural models resembles English, this superficial resemblance might be deeply misleading.

Corentin Kervadec, Iuliia Lysova, Iuri Macocco et al.

Transformer-based large language models (LLMs) are comprised of billions of parameters arranged in deep and wide computational graphs, but it is not clear that they exploit their full capacity for all inputs. We introduce the s-Trace method to efficiently estimate the subgraph of size s that best approximates a full model output. With this method, we find the computation in a variety of LLMs to be organized in two distinct phases. A small subgraph mostly composed of early-layer nodes can reconstruct the head of the full model output distribution. Adding further nodes, mostly located in later layers and increasingly consisting of attention heads, leads to incremental refinements in approximating the full output distribution. We find moreover that the amount of necessary computation per input correlates with model uncertainty, and that sparser subgraphs encode shallow statistics, such as unigram frequency. Overall, our results suggest a consistent modular organization in effective LLM computation, with a sparse early-layer core providing a rough prediction that is further refined through denser computations in later layers.

Corentin Kervadec, Iuliia Lysova, Marco Baroni et al.

Transformer-based large language models (LLMs) are comprised of billions of parameters arranged in deep and wide computational graphs. Several studies on LLM efficiency optimization argue that it is possible to prune a significant portion of the parameters, while only marginally impacting performance. This suggests that the computation is not uniformly distributed across the parameters. We introduce here a technique to systematically quantify computation density in LLMs. In particular, we design a density estimator drawing on mechanistic interpretability. We experimentally test our estimator and find that: (1) contrary to what has been often assumed, LLM processing generally involves dense computation; (2) computation density is dynamic, in the sense that models shift between sparse and dense processing regimes depending on the input; (3) per-input density is significantly correlated across LLMs, suggesting that the same inputs trigger either low or high density. Investigating the factors influencing density, we observe that predicting rarer tokens requires higher density, and increasing context length often decreases the density. We believe that our computation density estimator will contribute to a better understanding of the processing at work in LLMs, challenging their symbolic interpretation.

Yunke He, Xixian Liao, Jialing Liang et al.

Different speakers often produce different names for the same object or entity (e.g., "woman" vs. "tourist" for a female tourist). The reasons behind variation in naming are not well understood. We create a Language and Vision dataset for Mandarin Chinese that provides an average of 20 names for 1319 naturalistic images, and investigate how familiarity with a given kind of object relates to the degree of naming variation it triggers across subjects. We propose that familiarity influences naming variation in two competing ways: increasing familiarity can either expand vocabulary, leading to higher variation, or promote convergence on conventional names, thereby reducing variation. We find evidence for both factors being at play. Our study illustrates how computational resources can be used to address research questions in Cognitive Science.

Nora Graichen, Iria de-Dios-Flores, Gemma Boleda

We present a systematic review of 337 articles evaluating the syntactic abilities of Transformer-based language models, reporting on 1,015 model results from a range of syntactic phenomena and interpretability methods. Our analysis shows that the state of the art presents a healthy variety of methods and data, but an over-focus on a single language (English), a single model (BERT), and phenomena that are easy to get at (like part of speech and agreement). Results also suggest that TLMs capture these form-oriented phenomena well, but show more variable and weaker performance on phenomena at the syntax-semantics interface, like binding or filler-gap dependencies. We provide recommendations for future work, in particular reporting complete data, better aligning theoretical constructs and methods across studies, increasing the use of mechanistic methods, and broadening the empirical scope regarding languages and linguistic phenomena.

Yuqing Zhang, Ecesu Ürker, Tessa Verhoef et al.

Modeling the emergence of human-like lexicons in computational systems has advanced through the use of interacting neural agents, which simulate both learning and communicative pressures. The NeLLCom-Lex framework (Zhang et al., 2025) allows neural agents to develop pragmatic color naming behavior and human-like lexicons through supervised learning (SL) from human data and reinforcement learning (RL) in referential games. Despite these successes, the lexicons that emerge diverge systematically from human color categories, producing highly non-convex regions in color space, which contrast with the convexity typical of human categories. To address this, we introduce two factors, upsampling rare color terms during SL and multi-listener RL interactions, and adopt a convexity measure to quantify geometric coherence. We find that upsampling improves lexical diversity and system-level informativeness of the color lexicon, while many-listener setups promote more convex color categories. The combination of moderate upsampling and multiple listeners produces lexicons most similar to human systems.

Iuri Macocco, Nora Graichen, Gemma Boleda et al.

We study last-layer outlier dimensions, i.e. dimensions that display extreme activations for the majority of inputs. We show that outlier dimensions arise in many different modern language models, and trace their function back to the heuristic of constantly predicting frequent words. We further show how a model can block this heuristic when it is not contextually appropriate, by assigning a counterbalancing weight mass to the remaining dimensions, and we investigate which model parameters boost outlier dimensions and when they arise during training. We conclude that outlier dimensions are a specialized mechanism discovered by many distinct models to implement a useful token prediction heuristic.

Yuqing Zhang, Ecesu Ürker, Tessa Verhoef et al.

Lexical semantic change has primarily been investigated with observational and experimental methods; however, observational methods (corpus analysis, distributional semantic modeling) cannot get at causal mechanisms, and experimental paradigms with humans are hard to apply to semantic change due to the extended diachronic processes involved. This work introduces NeLLCom-Lex, a neural-agent framework designed to simulate semantic change by first grounding agents in a real lexical system (e.g. English) and then systematically manipulating their communicative needs. Using a well-established color naming task, we simulate the evolution of a lexical system within a single generation, and study which factors lead agents to: (i) develop human-like naming behavior and lexicons, and (ii) change their behavior and lexicons according to their communicative needs. Our experiments with different supervised and reinforcement learning pipelines show that neural agents trained to 'speak' an existing language can reproduce human-like patterns in color naming to a remarkable extent, supporting the further use of NeLLCom-Lex to elucidate the mechanisms of semantic change.

Gemma Boleda

Since the middle of the 20th century, a fierce battle is being fought between symbolic and distributed approaches to language and cognition. The success of deep learning models, and LLMs in particular, has been alternatively taken as showing that the distributed camp has won, or dismissed as an irrelevant engineering development. In this position paper, I argue that deep learning models for language actually represent a synthesis between the two traditions. This is because 1) deep learning architectures allow for both distributed/continuous/fuzzy and symbolic/discrete/categorical-like representations and processing; 2) models trained on language make use of this flexibility. In particular, I review recent research in interpretability that showcases how a substantial part of morphosyntactic knowledge is encoded in a near-discrete fashion in LLMs. This line of research suggests that different behaviors arise in an emergent fashion, and models flexibly alternate between the two modes (and everything in between) as needed. This is possibly one of the main reasons for their wild success; and it makes them particularly interesting for the study of language. Is it time for peace?

Sophia Harrison, Eleonora Gualdoni, Gemma Boleda

Gender bias in Language and Vision datasets and models has the potential to perpetuate harmful stereotypes and discrimination. We analyze gender bias in two Language and Vision datasets. Consistent with prior work, we find that both datasets underrepresent women, which promotes their invisibilization. Moreover, we hypothesize and find that a bias affects human naming choices for people playing sports: speakers produce names indicating the sport (e.g. 'tennis player' or 'surfer') more often when it is a man or a boy participating in the sport than when it is a woman or a girl, with an average of 46% vs. 35% of sports-related names for each gender. A computational model trained on these naming data reproduces the bias. We argue that both the data and the model result in representational harm against women.

Laura Aina, Xixian Liao, Gemma Boleda et al.

It is often posited that more predictable parts of a speaker's meaning tend to be made less explicit, for instance using shorter, less informative words. Studying these dynamics in the domain of referring expressions has proven difficult, with existing studies, both psycholinguistic and corpus-based, providing contradictory results. We test the hypothesis that speakers produce less informative referring expressions (e.g., pronouns vs. full noun phrases) when the context is more informative about the referent, using novel computational estimates of referent predictability. We obtain these estimates training an existing coreference resolution system for English on a new task, masked coreference resolution, giving us a probability distribution over referents that is conditioned on the context but not the referring expression. The resulting system retains standard coreference resolution performance while yielding a better estimate of human-derived referent predictability than previous attempts. A statistical analysis of the relationship between model output and mention form supports the hypothesis that predictability affects the form of a mention, both its morphosyntactic type and its length.

Kristina Gulordava, Thomas Brochhagen, Gemma Boleda

Children's tendency to associate novel words with novel referents has been taken to reflect a bias toward mutual exclusivity. This tendency may be advantageous both as (1) an ad-hoc referent selection heuristic to single out referents lacking a label and as (2) an organizing principle of lexical acquisition. This paper investigates under which circumstances cross-situational neural models can come to exhibit analogous behavior to children, focusing on these two possibilities and their interaction. To this end, we evaluate neural networks' on both symbolic data and, as a first, on large-scale image data. We find that constraints in both learning and selection can foster mutual exclusivity, as long as they put words in competition for lexical meaning. For computational models, these findings clarify the role of available options for better performance in tasks where mutual exclusivity is advantageous. For cognitive research, they highlight latent interactions between word learning, referent selection mechanisms, and the structure of stimuli of varying complexity: symbolic and visual.

Alba Herrera-Palacio, Carles Ventura, Carina Silberer et al.

The goal of this work is to segment the objects in an image that are referred to by a sequence of linguistic descriptions (referring expressions). We propose a deep neural network with recurrent layers that output a sequence of binary masks, one for each referring expression provided by the user. The recurrent layers in the architecture allow the model to condition each predicted mask on the previous ones, from a spatial perspective within the same image. Our multimodal approach uses off-the-shelf architectures to encode both the image and the referring expressions. The visual branch provides a tensor of pixel embeddings that are concatenated with the phrase embeddings produced by a language encoder. Our experiments on the RefCOCO dataset for still images indicate how the proposed architecture successfully exploits the sequences of referring expressions to solve a pixel-wise task of instance segmentation.

Laura Aina, Kristina Gulordava, Gemma Boleda

In neural network models of language, words are commonly represented using context-invariant representations (word embeddings) which are then put in context in the hidden layers. Since words are often ambiguous, representing the contextually relevant information is not trivial. We investigate how an LSTM language model deals with lexical ambiguity in English, designing a method to probe its hidden representations for lexical and contextual information about words. We find that both types of information are represented to a large extent, but also that there is room for improvement for contextual information.

Matthijs Westera, Gemma Boleda

Distributional semantics has had enormous empirical success in Computational Linguistics and Cognitive Science in modeling various semantic phenomena, such as semantic similarity, and distributional models are widely used in state-of-the-art Natural Language Processing systems. However, the theoretical status of distributional semantics within a broader theory of language and cognition is still unclear: What does distributional semantics model? Can it be, on its own, a fully adequate model of the meanings of linguistic expressions? The standard answer is that distributional semantics is not fully adequate in this regard, because it falls short on some of the central aspects of formal semantic approaches: truth conditions, entailment, reference, and certain aspects of compositionality. We argue that this standard answer rests on a misconception: These aspects do not belong in a theory of expression meaning, they are instead aspects of speaker meaning, i.e., communicative intentions in a particular context. In a slogan: words do not refer, speakers do. Clearing this up enables us to argue that distributional semantics on its own is an adequate model of expression meaning. Our proposal sheds light on the role of distributional semantics in a broader theory of language and cognition, its relationship to formal semantics, and its place in computational models.

Laura Aina, Carina Silberer, Matthijs Westera et al.

Humans use language to refer to entities in the external world. Motivated by this, in recent years several models that incorporate a bias towards learning entity representations have been proposed. Such entity-centric models have shown empirical success, but we still know little about why. In this paper we analyze the behavior of two recently proposed entity-centric models in a referential task, Entity Linking in Multi-party Dialogue (SemEval 2018 Task 4). We show that these models outperform the state of the art on this task, and that they do better on lower frequency entities than a counterpart model that is not entity-centric, with the same model size. We argue that making models entity-centric naturally fosters good architectural decisions. However, we also show that these models do not really build entity representations and that they make poor use of linguistic context. These negative results underscore the need for model analysis, to test whether the motivations for particular architectures are borne out in how models behave when deployed.

Gemma Boleda

Distributional semantics provides multi-dimensional, graded, empirically induced word representations that successfully capture many aspects of meaning in natural languages, as shown in a large body of work in computational linguistics; yet, its impact in theoretical linguistics has so far been limited. This review provides a critical discussion of the literature on distributional semantics, with an emphasis on methods and results that are of relevance for theoretical linguistics, in three areas: semantic change, polysemy and composition, and the grammar-semantics interface (specifically, the interface of semantics with syntax and with derivational morphology). The review aims at fostering greater cross-fertilization of theoretical and computational approaches to language, as a means to advance our collective knowledge of how it works.

Marco Del Tredici, Raquel Fernández, Gemma Boleda

We present the first exploration of meaning shift over short periods of time in online communities using distributional representations. We create a small annotated dataset and use it to assess the performance of a standard model for meaning shift detection on short-term meaning shift. We find that the model has problems distinguishing meaning shift from referential phenomena, and propose a measure of contextual variability to remedy this.

Abhijeet Gupta, Gemma Boleda, Sebastian Pado

In computational linguistics, a large body of work exists on distributed modeling of lexical relations, focussing largely on lexical relations such as hypernymy (scientist -- person) that hold between two categories, as expressed by common nouns. In contrast, computational linguistics has paid little attention to entities denoted by proper nouns (Marie Curie, Mumbai, ...). These have investigated in detail by the Knowledge Representation and Semantic Web communities, but generally not with regard to their linguistic properties. Our paper closes this gap by investigating and modeling the lexical relation of instantiation, which holds between an entity-denoting and a category-denoting expression (Marie Curie -- scientist or Mumbai -- city). We present a new, principled dataset for the task of instantiation detection as well as experiments and analyses on this dataset. We obtain the following results: (a), entities belonging to one category form a region in distributional space, but the embedding for the category word is typically located outside this subspace; (b) it is easy to learn to distinguish entities from categories from distributional evidence, but due to (a), instantiation proper is much harder to learn when using common nouns as representations of categories; (c) this problem can be alleviated by using category representations based on entity rather than category word embeddings.

Laura Aina, Carina Silberer, Ionut-Teodor Sorodoc et al.

This paper describes our winning contribution to SemEval 2018 Task 4: Character Identification on Multiparty Dialogues. It is a simple, standard model with one key innovation, an entity library. Our results show that this innovation greatly facilitates the identification of infrequent characters. Because of the generic nature of our model, this finding is potentially relevant to any task that requires effective learning from sparse or unbalanced data.

Gemma Boleda, Sebastian Padó, Nghia The Pham et al.

Reference is a crucial property of language that allows us to connect linguistic expressions to the world. Modeling it requires handling both continuous and discrete aspects of meaning. Data-driven models excel at the former, but struggle with the latter, and the reverse is true for symbolic models. This paper (a) introduces a concrete referential task to test both aspects, called cross-modal entity tracking; (b) proposes a neural network architecture that uses external memory to build an entity library inspired in the DRSs of DRT, with a mechanism to dynamically introduce new referents or add information to referents that are already in the library. Our model shows promise: it beats traditional neural network architectures on the task. However, it is still outperformed by Memory Networks, another model with external memory.

Gemma Boleda, Sebastian Padó, Marco Baroni

One of the most basic functions of language is to refer to objects in a shared scene. Modeling reference with continuous representations is challenging because it requires individuation, i.e., tracking and distinguishing an arbitrary number of referents. We introduce a neural network model that, given a definite description and a set of objects represented by natural images, points to the intended object if the expression has a unique referent, or indicates a failure, if it does not. The model, directly trained on reference acts, is competitive with a pipeline manually engineered to perform the same task, both when referents are purely visual, and when they are characterized by a combination of visual and linguistic properties.

Denis Paperno, Germán Kruszewski, Angeliki Lazaridou et al.

We introduce LAMBADA, a dataset to evaluate the capabilities of computational models for text understanding by means of a word prediction task. LAMBADA is a collection of narrative passages sharing the characteristic that human subjects are able to guess their last word if they are exposed to the whole passage, but not if they only see the last sentence preceding the target word. To succeed on LAMBADA, computational models cannot simply rely on local context, but must be able to keep track of information in the broader discourse. We show that LAMBADA exemplifies a wide range of linguistic phenomena, and that none of several state-of-the-art language models reaches accuracy above 1% on this novel benchmark. We thus propose LAMBADA as a challenging test set, meant to encourage the development of new models capable of genuine understanding of broad context in natural language text.

Alvaro Corral, Gemma Boleda, Ramon Ferrer-i-Cancho

Zipf's law is a fundamental paradigm in the statistics of written and spoken natural language as well as in other communication systems. We raise the question of the elementary units for which Zipf's law should hold in the most natural way, studying its validity for plain word forms and for the corresponding lemma forms. In order to have as homogeneous sources as possible, we analyze some of the longest literary texts ever written, comprising four different languages, with different levels of morphological complexity. In all cases Zipf's law is fulfilled, in the sense that a power-law distribution of word or lemma frequencies is valid for several orders of magnitude. We investigate the extent to which the word-lemma transformation preserves two parameters of Zipf's law: the exponent and the low-frequency cut-off. We are not able to demonstrate a strict invariance of the tail, as for a few texts both exponents deviate significantly, but we conclude that the exponents are very similar, despite the remarkable transformation that going from words to lemmas represents, considerably affecting all ranges of frequencies. In contrast, the low-frequency cut-offs are less stable.