e-Repositori UPF :: Congressos (Departament de Traducció i Ciències del Llenguatge)

Enviaments recents

Contexts of language learning: predicting child language by interactive speech in 9 languages
(Cascadilla Proceedings Project, 2025) Rüst, Olivier; Baroni, Marco; Stoll, Sabine
A central challenge for the language learning child is to extract the components of their communicative system from the input. While children are known to be excellent statistical learners (e.g., Saffran et al. 1999), less is understood about the structure of the input that supports this learning. Recent research suggests that, beyond input quantity, input quality (Anderson et al. 2021) and especially the linguistic interactions children experience play a crucial (Feurstein et al. 2022; Donnelly & Kidd 2021; Newman et al. 2016; Weisleder & Fernald 2013; Rowe 2012; Shneidman & Goldin-Meadow 2012; Meredith L. Rowe & Ayoub 2005). This study investigates whether and how caregiver-child interactions facilitate language learning. Specifically, we hypothesize that interactive moments characterized by increased turn-taking between child and caregiver create optimal conditions for learning. This hypothesis is tested across nine typologically diverse languages: Chintang (Sino - Tibetan), Cree (Algic), Indonesian (Austronesian), Japanese (Japonic), Ku Waru (Trans - New Guinea), Russian (Indo - European), Sesotho (Niger - Congo), Turkish (Turkic), and Yucatec (Mayan).
Evil twins are not that evil: qualitative insights into machine-generated prompts
(ACL (Association for Computational Linguistics), 2025) Carraz Rakotonirina, Nathanaël; Kervadec, Corentin; Franzon, Francesca; Baroni, Marco
It has been widely observed that language models (LMs) respond in predictable ways to algorithmically generated prompts that are seemingly unintelligible. This is both a sign that we lack a full understanding of how LMs work, and a practical challenge, because opaqueness can be exploited for harmful uses of LMs, such as jailbreaking. We present the first thorough analysis of opaque machine-generated prompts, or autoprompts, pertaining to 6 LMs of different sizes and families. We find that machinegenerated prompts are characterized by a last token that is often intelligible and strongly affects the generation. A small but consistent proportion of the previous tokens are prunable, probably appearing in the prompt as a by-product of the fact that the optimization process fixes the number of tokens. The remaining tokens fall into two categories: filler tokens, which can be replaced with semantically unrelated substitutes, and keywords, that tend to have at least a loose semantic relation with the generation, although they do not engage in wellformed syntactic relations with it. Additionally, human experts can reliably identify the most influential tokens in an autoprompt a posteriori, suggesting these prompts are not entirely opaque. Finally, some of the ablations we applied to autoprompts yield similar effects in natural language inputs, suggesting that autoprompts emerge naturally from the way LMs process linguistic inputs in general.
Not a nuisance but a useful heuristic: outlier dimensions favor frequent tokens in language models
(ACL (Association for Computational Linguistics), 2025) Macocco, Iuri; Graichen, Nora; Boleda, Gemma; Baroni, Marco
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.
Documenting the final days of monolingual English learners' dictionaries using the archived web
(Lexical Computing CZ s.r.o., 2025) Rees, Geraint Paul
Online dictionaries have many advantages over their physical counterparts. However, the ephemeral nature of web content means that they are often changed without notice and no ostensible record of what came before remains. This makes research on historical online dictionaries difficult and perhaps explains why, while the history of printed monolingual English learners’ dictionaries (MELDs) has been comprehensively explored, studies of online dictionaries have tended to take a cross-sectional rather than longitudinal view. This is not ideal since it means that a large period of MELD history is yet to be explored. Moreover, given recent predictions of the decline of MELDs, as we know them, in light of developments with AI chatbots and other digital tools, this gap is all the more significant. In an attempt to remedy this situation, this study applies Brügger’s (2018) framework for archived web research to explore the feasibility of using the web archive, the Wayback Machine, to trace the development of websites that give, or have given, access to ‘the big five’ MELDs. Some key challenges of using archived web material to conduct lexicographic research are discussed along with suggestions for potential solutions.
Prediction hubs are context-informed frequent tokens in LLMs
(ACL (Association for Computational Linguistics), 2025) Nielsen, Beatrix MG; Macocco, Iuri; Baroni, Marco
Hubness, the tendency for a few points to be among the nearest neighbours of a disproportionate number of other points, commonly arises when applying standard distance measures to high-dimensional data, often negatively impacting distance-based analysis. As autoregressive large language models (LLMs) operate on high-dimensional representations, we ask whether they are also affected by hubness. We first prove that the only large-scale representation comparison operation performed by LLMs, namely that between context and unembedding vectors to determine continuation probabilities, is not characterized by the concentration of distances phenomenon that typically causes the appearance of nuisance hubness. We then empirically show that this comparison still leads to a high degree of hubness, but the hubs in this case do not constitute a disturbance. They are rather the result of context-modulated frequent tokens often appearing in the pool of likely candidates for next token prediction. However, when other distances are used to compare LLM representations, we do not have the same theoretical guarantees, and, indeed, we see nuisance hubs appear. There are two main takeaways. First, hubness, while omnipresent in high-dimensional spaces, is not a negative property that needs to be mitigated when LLMs are being used for next token prediction. Second, when comparing representations from LLMs using Euclidean or cosine distance, there is a high risk of nuisance hubs and practitioners should use mitigation techniques if relevant.
Emergence of a high-dimensional abstraction phase in language transformers
(International Conference on Learning Representations, 2025) Cheng, Emily; Doimo, Diego; Kervadec, Corentin; Macocco, Iuri; Yu, Lei; Laio, Alessandro; Baroni, Marco
A language model (LM) is a mapping from a linguistic context to an output token. However, much remains to be known about this mapping, including how its geometric properties relate to its function. We take a high-level geometric approach to its analysis, observing, across five pre-trained transformer-based LMs and three input datasets, a distinct phase characterized by high intrinsic dimensionality. During this phase, representations (1) correspond to the first full linguistic abstraction of the input; (2) are the first to viably transfer to downstream tasks; (3) predict each other across different LMs. Moreover, we find that an earlier onset of the phase strongly predicts better language modelling performance. In short, our results suggest that a central high-dimensionality phase underlies core linguistic processing in many common LM architectures.
Unagreement and how morphology sees syntax
(Lexical-Functional Grammar, 2024) Alsina i Keith, Àlex
The phenomenon of unagreement, found in Spanish, Catalan, and Greek, among other languages, poses four theoretical problems: 1) how to account for an apparent mismatch between trigger and target in an agreement relation; 2) how to account for the fact that not all languages have this phenomenon; 3) how to account for variation in the NPs that trigger unagreement within a given language and across languages; 4) how to account for the correlation between the presence or absence of unagreement and the type of adnominal pronoun construction (APC) allowed in the language. The analysis assumes a lexicalist unencapsulated view of the relationship between syntax and inflectional morphology, which implies that agreement is a strictly morphological phenomenon. The fundamental idea is that some determiners in some languages do not specify person information. This implies that a phrase headed by such a determiner is compatible with any person feature.
A multilingual annotated corpus for the study of information structure
(Narr Francke Attempto Verlag, 2009) Brunetti, Lisa; Bott, Stefan Markus; Costa, Joan; Vallduví, Enric
This paper presents a corpus of spoken narrative texts in Catalan, Italian, Spanish, English, and German. The aim of this corpus compilation is to create an empirical resource for a comparative study of Information Structure. A total of 68 speakers were asked to tell a story in an acoustically isolated room by looking at the pictures of three textless books. A total of 222 narrations resulted in about 16 hours of speech. The recordings have been transcribed and an original annotation of non-canonical constructions for the Romance subgroup has been proposed, namely of morphosyntactically/prosodically marked constructions that relate informational categories such as topic, focus, and contrast. Transcriptions and annotations of some selected high quality recordings have been aligned to the acoustic signal stream. The corpus is available in audio and text format.
La dislocació del verb en català
(Publicacions de l'Abadia de Montserrat, 2001) Vallduví, Enric
La dislocació (a l'esquerra) d'arguments i adjunts es un fenomen força conegut en la sintaxi del català (e.g. Solà 1991, Vallduví 1992a) i altres llengües romàniques. Tanmateix, la dislocació del verb és gairebé del tot desconeguda a nivell acadèmic, tot i ésser un fenomen ben viu a tots els parlars catalans. Aquest paper presenta exemples de dislocació del verb i mostra la seva equivalència formal a la dislocació dels arguments, fent servir un seguit de tests (recció d'un element coreferenciat dins de l'oració, ubicació de partícules perifèriques) i n'esmenta les diferències (represa amb forma verbal i no pronominal, desviació morfològica cap a la forma d'infinitiu). També distingeix entre la dislocació del verb i l'avantposició del sintagma verbal, un fenomen estructural diferent. Des del punt de vista formal, és interessant la interacció entre la dislocació del verb i els pronoms febles, ja que els pronoms febles poden ser copiats a l'infinitiu dislocat de manera optativa. Finalment, s'analitza la càrrega semanticopragmàtica associada a aquesta configuració, que comparteix amb les altres dislocacions a l'esquerra.
Exploring undergraduates’ attitudes towards ChatGPT: is AI resistance constraining the acceptance of chatbot technology?
(Springer, 2024) Sánchez Reina, Jesús Roberto; Theophilou, Emily; Hernández Leo, Davinia; Ognibene, Dimitri
The advent of Artificial Intelligence (AI) has revolutionized multiple sectors including education. The popularization of tools such as ChatGPT has sparked the debate concerning the impact of AI on traditional education and the nature of learning. This paper explores undergraduate students’ attitudes towards AI and ChatGPT acceptance. A descriptive cross-sectional study with 72 Public Relations students (M age = 19.2 years old) took place in Barcelona (Spain) during the first semester of 2023. The study implemented a mixed method approach with two validated questionnaires and an open text question to gather comprehensive insights. Findings reveal positive attitudes towards artificial intelligence and ChatGPT acceptance. The assessment of negative perceptions show concerns regarding artificial intelligence and the use of ChatGPT among participants. The correlational analysis of scales showed an intricate relationship between AI attitudes and ChatGPT acceptance while the qualitative analysis highlighted three major attitudes among students: openness, awareness, and alertness. The present study contributes to the ongoing discourse surrounding the use of ChatGPT in educational settings, emphasizing the importance of exploring students’ attitudes and concerns. As artificial intelligence continues to permeate various aspects in our daily life, it becomes crucial to explore its impact on education, particularly in higher education. By understanding students’ attitudes, both educators and institutions can enhance their proficiency in integrating artificial intelligence in a more efficient manner, ensuring a well-balanced approach that maximizes benefits while mitigating potential drawbacks of adopting AI technology.
Generative AI chatbot in PyramidApp: students’ behaviors and design principles
(Springer, 2024) Gutiérrez-Ferré, Aldric; Hernández Leo, Davinia; Sánchez Reina, Jesús Roberto
Generative Artificial Intelligence (GenAI) offers new opportunities to implement useful features within Computer Supported Collaborative Learning (CSCL) environments. Despite these growing prospects, there is still limited research concerning the application of GenAI in learning environments. This work in progress aims to evaluate the mediation of a masked GenAI chatbot in the setting of the CSCL web application PyramidApp. A quasi-experimental within-subjects study was designed to assess the effect of GenAI chatbot intervention within the environment of PyramidApp. In the setting of 9 online activities, we evaluated the effect of the GenAI chatbot activity in 105 conversational chat rooms. The findings revealed that the GenAI chatbot provides useful feedback as students rate the chatbot’s answers higher than their peers’ answers (MChatbot = 4.11, MStudents = 3.91). The presence of the chatbot has an effect on group communication with the length of messages increased in chat rooms where the chatbot was present. Moreover, chatbot behavior to rate the students’ answers was correlated with the students’ behavior. The present study offers valuable insights into the optimal strategies for integrating a GenAI Large Language Model into educational tools and computer supported learning.
IA, contenido visual y periodismo: fundamentos para un debate
(Asociación Española de Investigación de la Comunicación (AE-IC), 2024) Freixa Font, Pere; Redondo i Arolas, Mar
La irrupción de la inteligencia artificial (IA) en los medios de comunicación ha provocado un amplio debate académico sobre su potencial transformador. Sin embargo, resulta sorprendente que el fotoperiodismo no se haya implicado en esa polémica. Ante el reto de la IA, a través de declaraciones y manifiestos, los actores involucrados en la creación de fotografías de prensa (fotógrafos, agencias, organizaciones profesionales y medios) se han posicionado prohibiendo su presencia en los medios. La negativa a cuestionar y debatir el impacto de la IA en el periodismo fotográfico invisibiliza el trabajo de autores que están explorando su potencialidad discursiva. Esta investigación, mediante el estudio de caso, tiene por objetivo localizar los principales aspectos tanto formales como narrativos y procedimentales que caracterizan la incorporación de IA en la creación fotoperiodística. La prácticamente nula existencia de estudios académicos sobre estos aspectos (Ritchin, 2023; Darrach, 2023) evidencia la necesidad de su abordaje, dada la relevancia de la comunicación visual en la elaboración de relatos periodísticos.
IA y fotoperiodismo: un escenario por definir
(2023) Freixa Font, Pere; Redondo i Arolas, Mar
La irrupción de la inteligencia artificial en los medios de comunicación está generando un intenso debate en torno a sus capacidades y repercusiones. Se percibe como una posibilidad o una amenaza según se valore su capacidad de transformación. En relación con los contenidos visuales, la IA está transformando radicalmente la producción y creación de imágenes ilustrativas. Sin embargo, el debate no parece haber llegado al fotoperiodismo. Actualmente aún no existen trabajos que analicen en profundidad el alcance que puede llegar a tener la incorporación de la IA en la creación de fotografía periodística. En esta investigación, por medio del estudio de caso, se localizan y definen aquellos aspectos que formarán parte del debate sobre los efectos de la incorporación de la IA en la construcción de relatos fotoperiodísticos. El uso incipiente de la IA por parte de fotógrafos de renombre ha demostrado su capacidad para desdibujar los límites del fotoperiodismo.
SemEval-2016 task 5: aspect based sentiment analysis
(ACL (Association for Computational Linguistics), 2016) Pontiki, Maria; Galanis, Dimitris; Papageorgiou, Haris; Androutsopoulos, Ion; Manandhar, Suresh; Al-Smadi, Mohammad; Al-Ayyoub, Mahmoud; Zhao, Yanyan; Qin, Bing; De Clercq, Orphée; Hoste, Véronique; Apidianaki, Marianna; Tannier, Xavier; Loukachevitch, Natalia; Kotelnikov, Evgeniy; Bel, Nuria; Bel Rafecas, Núria; Jiménez-Zafra, Salud María; Eryiğit, Gülşen
This paper describes the SemEval 2016 shared task on Aspect Based Sentiment Analysis (ABSA), a continuation of the respective tasks of 2014 and 2015. In its third year, the task provided 19 training and 20 testing datasets for 8 languages and 7 domains, as well as a common evaluation procedure. From these datasets, 25 were for sentence-level and 14 for text-level ABSA; the latter was introduced for the first time as a subtask in SemEval. The task attracted 245 submissions from 29 teams.
MemoryPrompt: a light wrapper to improve context tracking in pre-trained language models
(ELRA (European Language Resources Association), 2024) Carraz Rakotonirina, Nathanaël; Baroni, Marco
Transformer-based language models (LMs) track contextual information through large, hard-coded input windows. We introduce MemoryPrompt, a leaner approach in which the LM is complemented by a small auxiliary recurrent network that passes information to the LM by prefixing its regular input with a sequence of vectors, akin to soft prompts, without requiring LM finetuning. Tested on a task designed to probe a LM’s ability to keep track of multiple fact updates, a MemoryPrompt-augmented LM outperforms much larger LMs that have access to the full input history. We also test MemoryPrompt on a long-distance dialogue dataset, where its performance is comparable to that of a model conditioned on the entire conversation history. In both experiments we also observe that, unlike full-finetuning approaches, MemoryPrompt does not suffer from catastrophic forgetting when adapted to new tasks, thus not disrupting the generalist capabilities of the underlying LM.
Rhetorically-based scalar-additivity: the view from Italian addirittura
(Cornell University. Department of Linguistics, 2022) Pistoia-Reda, Salvatore; McNally, Louise, 1965-
Even-like particles have widely been analyzed as inducing scalar and additive presuppositions (cf. Horn 1969; Karttunen & Peters 1979; Rooth 1992; Gast & van der Auwera 2011). However, the additivity of even has been controversial since at least Rullmann 1997 and increasingly called into question (see Greenberg & Umbach 2021 for references); Greenberg specifically argues that scalar even-like particles can vary in additivity. This claim is surprising in light of the typological study in Gast & van der Auwera 2011, which subsumes even and similar expressions under a larger class of additive particles. Against this background, we present an analysis of Italian addirittura, which with perfino has been described as scalaradditive (Visconti 2005) – but only optionally so – and is chosen preferentially over perfino precisely in those contexts that Greenberg takes to challenge the additivity of even. We argue, drawing on observations in Atayan 2017, that addirittura contrasts with perfino in deriving its scalar alternatives from rhetorical structure rather than focus structure. Once this is recognized we can view addirittura as additive, after all, in a rhetorical sense we describe below.
Reference bias in monolingual machine translation evaluation
(ACL (Association for Computational Linguistics), 2016) Fomicheva, Marina; Specia, Lucia
In the translation industry, human translations are assessed by comparison with the source texts. In the Machine Translation (MT) research community, however, it is a common practice to perform quality assessment using a reference translation instead of the source text. In this paper we show that this practice has a serious issue – annotators are strongly biased by the reference translation provided, and this can have a negative impact on the assessment of MT quality.
F-to-c-structure mapping: accounting for inflectional morphology and periphrasis
(CSLI Publications, 2023) Alsina i Keith, Àlex
The treatment of inflectional periphrasis is problematic in LFG, apparently because of the lexicalist nature of the framework. A close inspection of what is usually understood by lexicalism reveals two distinct, but related, notions: lexicalism and lexical encapsulation. Complex inflectional systems show that one can preserve lexicalism (the idea that words and phrases are different in terms of units and rules of composition), but that it is necessary to reject lexical encapsulation (the idea that words are formed without input from syntax). An adequate theory of inflectional morphology needs a framework that is not constrained by lexical encapsulation. With such a framework, it is then possible to give a correct account of inflectional periphrasis. The paper develops the analysis of two periphrastic constructions, one in Latin and one in Catalan, within a non-encapsulated version of LFG.
Partitivity in romance and the syntax-morphology connection
(CSLI Publications, 2022) Alsina i Keith, Àlex
This paper claims that the relationship between morphology and syntax is multidirectional. It argues against the generally accepted position in LFG that word formation feeds the syntax and that syntax cannot feed word formation. The proposal is that the rules of inflectional morphology take f-structure information, together with other information, as their input. The main argument for this claim is provided by the comparative analysis of two Romance languages, one with the partitive affix and one without it. The observation that languages without the partitive affix have null indefinite objects, whereas languages with this affix seemingly do not, follows straightforwardly only if we assume that syntax feeds word formation.
Unnatural language processing: how do language models handle machine-generated prompts?
(ACL (Association for Computational Linguistics), 2023) Kervadec, Corentin; Franzon, Francesca; Baroni, Marco
Language model prompt optimization research has shown that semantically and grammatically well-formed manually crafted prompts are routinely outperformed by automatically generated token sequences with no apparent meaning or syntactic structure, including sequences of vectors from a model’s embedding space. We use machine-generated prompts to probe how models respond to input that is not composed of natural language expressions. We study the behavior of models of different sizes in multiple semantic tasks in response to both continuous and discrete machine-generated prompts, and compare it to the behavior in response to humangenerated natural-language prompts. Even when producing a similar output, machinegenerated and human prompts trigger different response patterns through the network processing pathways, including different perplexities, different attention and output entropy distributions, and different unit activation profiles. We provide preliminary insight into the nature of the units activated by different prompt types, suggesting that only natural language prompts recruit a genuinely linguistic circuit.

Navegar

Examinar

Enviaments recents