Search

The significant progress in the development of Large Language Models has contributed to blurring the distinction between human and AI-generated text. The increasing pervasiveness of AI-generated text and the difficulty in detecting it poses new challenges for our society. In this paper, we tackle the problem of detecting and attributing AI-generated text by proposing WhosAI, a triplet-network contrastive learning framework designed to predict whether a given input text has been generated by humans or AI and to unveil the authorship of the text. Unlike most existing approaches, our proposed framework is conceived to learn semantic similarity representations from multiple generators at once, thus equally handling both detection and attribution tasks. Furthermore, WhosAI is model-agnostic and scalable to the release of new AI text-generation models by incorporating their generated instances into the embedding space learned by our framework. Experimental results on the TuringBench benchmark of 200K news articles show that our proposed framework achieves outstanding results in both the Turing Test and Authorship Attribution tasks, outperforming all the methods listed in the TuringBench benchmark leaderboards., Comment: Accepted for publication at the 27th European Conference on Artificial Intelligence (ECAI-2024)

Verbs form the backbone of language, providing the structure and meaning to sentences. Yet, their intricate semantic nuances pose a longstanding challenge. Understanding verb relations through the concept of lexical entailment is crucial for comprehending sentence meanings and grasping verb dynamics. This work investigates the capabilities of eight Large Language Models in recognizing lexical entailment relations among verbs through differently devised prompting strategies and zero-/few-shot settings over verb pairs from two lexical databases, namely WordNet and HyperLex. Our findings unveil that the models can tackle the lexical entailment recognition task with moderately good performance, although at varying degree of effectiveness and under different conditions. Also, utilizing few-shot prompting can enhance the models' performance. However, perfectly solving the task arises as an unmet challenge for all examined LLMs, which raises an emergence for further research developments on this topic.

Constitutions are foundational legal documents that underpin the governmental and societal structures. As such, they are a reflection of a nation's cultural and social uniqueness, but also contribute to establish topics of universal importance, like citizens' rights and duties (RD). In this work, using the renowned GPT-3.5, we leverage generative large language models to understand constitutional passages that transcend national boundaries. A key contribution of our study is the introduction of a novel application of abstractive summarization on a multi-source collection of constitutional texts, with a focus on European countries' constitution passages related to RD topics. Our results show the meaningfulness of GPT-3.5 to produce informative, coherent and faithful summaries capturing RD topics across European countries.

The emergence of unveiling human-like behaviors in Large Language Models (LLMs) has led to a closer connection between NLP and human psychology. Scholars have been studying the inherent personalities exhibited by LLMs and attempting to incorporate human traits and behaviors into them. However, these efforts have primarily focused on commercially-licensed LLMs, neglecting the widespread use and notable advancements seen in Open LLMs. This work aims to address this gap by employing a set of 12 LLM Agents based on the most representative Open models and subject them to a series of assessments concerning the Myers-Briggs Type Indicator (MBTI) test and the Big Five Inventory (BFI) test. Our approach involves evaluating the intrinsic personality traits of Open LLM agents and determining the extent to which these agents can mimic human personalities when conditioned by specific personalities and roles. Our findings unveil that $(i)$ each Open LLM agent showcases distinct human personalities; $(ii)$ personality-conditioned prompting produces varying effects on the agents, with only few successfully mirroring the imposed personality, while most of them being ``closed-minded'' (i.e., they retain their intrinsic traits); and $(iii)$ combining role and personality conditioning can enhance the agents' ability to mimic human personalities. Our work represents a step up in understanding the dense relationship between NLP and human psychology through the lens of Open LLMs., Comment: Enhanced methodology and evaluation based on BFI in addition to MBTI, with expanded set of LLM agents. Author list changed w.r.t. the previous version (v1), see Acknowledgements

Centralized social media platforms are currently experiencing a shift in user engagement, drawing attention to alternative paradigms like Decentralized Online Social Networks (DOSNs). The rising popularity of DOSNs finds its root in the accessibility of open-source software, enabling anyone to create a new instance (i.e., server) and participate in a decentralized network known as Fediverse. Despite this growing momentum, there has been a lack of studies addressing the effect of positive and negative interactions among instances within DOSNs. This work aims to fill this gap by presenting a preliminary examination of instances' polarization in DOSNs, focusing on Mastodon -- the most widely recognized decentralized social media platform, boasting over 10M users and nearly 20K instances to date. Our results suggest that polarization in the Fediverse emerges in unique ways, influenced by the desire to foster a federated environment between instances, also facilitating the isolation of instances that may pose potential risks to the Fediverse.

Transformer-based language models (TLMs) have widely been recognized to be a cutting-edge technology for the successful development of deep-learning-based solutions to problems and applications that require natural language processing and understanding. Like for other textual domains, TLMs have indeed pushed the state-of-the-art of AI approaches for many tasks of interest in the legal domain. Despite the first Transformer model being proposed about six years ago, there has been a rapid progress of this technology at an unprecedented rate, whereby BERT and related models represent a major reference, also in the legal domain. This article provides the first systematic overview of TLM-based methods for AI-driven problems and tasks in the legal sphere. A major goal is to highlight research advances in this field so as to understand, on the one hand, how the Transformers have contributed to the success of AI in supporting legal processes, and on the other hand, what are the current limitations and opportunities for further research development., Comment: Please refer to the published version: Greco, C.M., Tagarelli, A. (2023) Bringing order into the realm of Transformer-based language models for artificial intelligence and law. Artif Intell Law, Springer Nature. November 2023. https://doi.org/10.1007/s10506-023-09374-7

Two-dimensional flat-band systems have recently attracted considerable interest due to the rich physics unveiled by emergent phenomena and correlated electronic states at van Hove singularities. However, the difficulties in electrically detecting the flat band position in field-effect structures are slowing down the investigation of their properties. In this work, we employ Indium Selenide (InSe) as a flat-band system due to a van Hove singularity at the valence band edge in a few-layer form of the material without the requirement of a twist angle. We investigate tunneling photocurrents in gated few-layer InSe structures and relate them to ambipolar transport and photoluminescence measurements. We observe an appearance of a sharp change in tunneling mechanisms due to the presence of the van Hove singularity at the flat band. We further corroborate our findings by studying tunneling currents as a reliable probe for the flat-band position up to room temperature. Our results create an alternative approach to studying flat-band systems in heterostructures of 2D materials.

The migration of Twitter users to Mastodon following Elon Musk's acquisition presents a unique opportunity to study collective behavior and gain insights into the drivers of coordinated behavior in online media. We analyzed the social network and the public conversations of about 75,000 migrated users and observed that the temporal trace of their migrations is compatible with a phenomenon of social influence, as described by a compartmental epidemic model of information diffusion. Drawing from prior research on behavioral change, we delved into the factors that account for variations across different Twitter communities in the effectiveness of the spreading of the influence to migrate. Communities in which the influence process unfolded more rapidly exhibit lower density of social connections, higher levels of signaled commitment to migrating, and more emphasis on shared identity and exchange of factual knowledge in the community discussion. These factors account collectively for 57% of the variance in the observed data. Our results highlight the joint importance of network structure, commitment, and psycho-linguistic aspects of social interactions in describing grassroots collective action, and contribute to deepen our understanding of the mechanisms driving processes of behavior change of online groups.

Transition-metal dichalcogenide bilayers exhibit a rich exciton landscape including layer-hybridized excitons, i.e. excitons which are of partly intra- and interlayer nature. In this work, we study hybrid exciton-exciton interactions in naturally stacked WSe$_2$ homobilayers. In these materials, the exciton landscape is electrically tunable such that the low-energy states can be rendered more or less interlayer-like depending on the strength of the external electric field. Based on a microscopic and material-specific many-particle theory, we reveal two intriguing interaction regimes: a low-dipole regime at small electric fields and a high-dipole regime at larger fields, involving interactions between hybrid excitons with a substantially different intra- and interlayer composition in the two regimes. While the low-dipole regime is characterized by weak inter-excitonic interactions between intralayer-like excitons, the high-dipole regime involves mostly interlayer-like excitons which display a strong dipole-dipole repulsion and give rise to large spectral blue-shifts and a highly anomalous diffusion. Overall, our microscopic study sheds light on the remarkable electrical tunability of hybrid exciton-exciton interactions in atomically thin semiconductors and can guide future experimental studies in this growing field of research., Comment: 8+8 pages, 4+2 figures

The fervor for Non-Fungible Tokens (NFTs) attracted countless creators, leading to a Big Bang of digital assets driven by latent or explicit forms of inspiration, as in many creative processes. This work exploits Vision Transformers and graph-based modeling to delve into visual inspiration phenomena between NFTs over the years. Our goals include unveiling the main structural traits that shape visual inspiration networks, exploring the interrelation between visual inspiration and asset performances, investigating crypto influence on inspiration processes, and explaining the inspiration relationships among NFTs. Our findings unveil how the pervasiveness of inspiration led to a temporary saturation of the visual feature space, the impact of the dichotomy between inspiring and inspired NFTs on their financial performance, and an intrinsic self-regulatory mechanism between markets and inspiration waves. Our work can serve as a starting point for gaining a broader view of the evolution of Web3., Comment: Under Review

Interactions between out-of-plane dipoles in bosonic gases enable the long-range propagation of excitons. The lack of direct control over collective dipolar properties has hitherto limited the degrees of tunability and the microscopic understanding of exciton transport. In this work, we modulate the layer hybridization and interplay between many-body interactions of excitons in a van der Waals heterostructure with an applied vertical electric field. By performing spatiotemporally resolved measurements supported by microscopic theory, we uncover the dipole-dependent properties and transport of excitons with different degrees of hybridization. Moreover, we find constant emission quantum yields of the transporting species as a function of excitation power with dominating radiative decay mechanisms over nonradiative ones, a fundamental requirement for efficient excitonic devices. Our findings provide a complete picture of the many-body effects in the transport of dilute exciton gases and have crucial implications for the study of emerging states of matter, such as Bose-Einstein condensation, as well as for optoelectronic applications based on exciton propagation.

Non-Fungible Tokens (NFTs) represent deeds of ownership, based on blockchain technologies and smart contracts, of unique crypto assets on digital art forms (e.g., artworks or collectibles). In the spotlight after skyrocketing in 2021, NFTs have attracted the attention of crypto enthusiasts and investors intent on placing promising investments in this profitable market. However, the NFT financial performance prediction has not been widely explored to date. In this work, we address the above problem based on the hypothesis that NFT images and their textual descriptions are essential proxies to predict the NFT selling prices. To this purpose, we propose MERLIN, a novel multimodal deep learning framework designed to train Transformer-based language and visual models, along with graph neural network models, on collections of NFTs' images and texts. A key aspect in MERLIN is its independence on financial features, as it exploits only the primary data a user interested in NFT trading would like to deal with, i.e., NFT images and textual descriptions. By learning dense representations of such data, a price-category classification task is performed by MERLIN models, which can also be tuned according to user preferences in the inference phase to mimic different risk-return investment profiles. Experimental evaluation on a publicly available dataset has shown that MERLIN models achieve significant performances according to several financial assessment criteria, fostering profitable investments, and also beating baseline machine-learning classifiers based on financial features., Comment: Accepted paper at The ACM Web Conference 2023, April 30--May 04, 2023, Austin, Texas, USA

Interactions among a collection of particles generate many-body effects in solids resulting in striking modifications of material properties. The heavy carrier mass that yields strong interactions and gate control of carrier density over a wide range, make two-dimensional semiconductors an exciting playground to explore many-body physics. The family of III-VI metal monochalcogenides emerges as a new platform for this purpose due to its excellent optical properties and the flat valence band dispersion with a Mexican-hat-like inversion. In this work, we present a complete study of charge-tunable excitons in few-layer InSe by photoluminescence spectroscopy. From the optical spectra, we establish that free excitons in InSe are more likely to be captured by ionized donors due to the large exciton Bohr radius, leading to the formation of bound exciton complexes. Surprisingly, a pronounced redshift of the exciton energy accompanied by a decrease of the exciton binding energy upon hole-doping reveals a significant band gap renormalization and dynamical screening induced by the presence of the Fermi reservoir. Our findings establish InSe as a reproducible and potentially manufacturable platform to explore electron correlation phenomena without the need for twist-angle engineering.

Decentralized Online Social Networks (DOSNs) represent a growing trend in the social media landscape, as opposed to the well-known centralized peers, which are often in the spotlight due to privacy concerns and a vision typically focused on monetization through user relationships. By exploiting open-source software, DOSNs allow users to create their own servers, or instances, thus favoring the proliferation of platforms that are independent yet interconnected with each other in a transparent way. Nonetheless, the resulting cooperation model, commonly known as the Fediverse, still represents a world to be fully discovered, since existing studies have mainly focused on a limited number of structural aspects of interest in DOSNs. In this work, we aim to fill a lack of study on user relations and roles in DOSNs, by taking two main actions: understanding the impact of decentralization on how users relate to each other within their membership instance and/or across different instances, and unveiling user roles that can explain two interrelated axes of social behavioral phenomena, namely information consumption and boundary spanning. To this purpose, we build our analysis on user networks from Mastodon, since it represents the most widely used DOSN platform. We believe that the findings drawn from our study on Mastodon users' roles and information flow can pave a way for further development of fascinating research on DOSNs., Comment: Preprint of article published with Online Social Networks and Media, vol. 30:100220, June 2022. Elsevier

Modeling law search and retrieval as prediction problems has recently emerged as a predominant approach in law intelligence. Focusing on the law article retrieval task, we present a deep learning framework named LamBERTa, which is designed for civil-law codes, and specifically trained on the Italian civil code. To our knowledge, this is the first study proposing an advanced approach to law article prediction for the Italian legal system based on a BERT (Bidirectional Encoder Representations from Transformers) learning framework, which has recently attracted increased attention among deep learning approaches, showing outstanding effectiveness in several natural language processing and learning tasks. We define LamBERTa models by fine-tuning an Italian pre-trained BERT on the Italian civil code or its portions, for law article retrieval as a classification task. One key aspect of our LamBERTa framework is that we conceived it to address an extreme classification scenario, which is characterized by a high number of classes, the few-shot learning problem, and the lack of test query benchmarks for Italian legal prediction tasks. To solve such issues, we define different methods for the unsupervised labeling of the law articles, which can in principle be applied to any law article code system. We provide insights into the explainability and interpretability of our LamBERTa models, and we present an extensive experimental analysis over query sets of different type, for single-label as well as multi-label evaluation tasks. Empirical evidence has shown the effectiveness of LamBERTa, and also its superiority against widely used deep-learning text classifiers and a few-shot learner conceived for an attribute-aware prediction task.

Dipolar bosonic gases are currently the focus of intensive research due to their interesting many-body physics in the quantum regime. Their experimental embodiments range from Rydberg atoms to GaAs double quantum wells and van der Waals heterostructures built from transition metal dichalcogenides. Although quantum gases are very dilute, mutual interactions between particles could lead to exotic many-body phenomena such as Bose-Einstein condensation and high-temperature superfluidity. Here, we report the effect of repulsive dipolar interactions on the dynamics of interlayer excitons in the dilute regime. By using spatial and time-resolved photoluminescence imaging, we observe the dynamics of exciton transport, enabling a direct estimation of the exciton mobility. The presence of interactions significantly modifies the diffusive transport of excitons, effectively acting as a source of drift force and enhancing the diffusion coefficient by one order of magnitude. The repulsive dipolar interactions combined with the electrical control of interlayer excitons opens up appealing new perspectives for excitonic devices.

Open-source, Decentralized Online Social Networks (DOSNs) are emerging as alternatives to the popular yet centralized and profit-driven platforms like Facebook or Twitter. In DOSNs, users can set up their own server, or instance, while they can actually interact with users of other instances. Moreover, by adopting the same communication protocol, DOSNs become part of a massive social network, namely the Fediverse. Mastodon is the most relevant platform in the Fediverse to date, and also the one that has attracted attention from the research community. Existing studies are however limited to an analysis of a relatively outdated sample of Mastodon focusing on few aspects at a user level, while several open questions have not been answered yet, especially at the instance level. In this work, we aim at pushing forward our understanding of the Fediverse by leveraging the primary role of Mastodon therein. Our first contribution is the building of an up-to-date and highly representative dataset of Mastodon. Upon this new data, we have defined a network model over Mastodon instances and exploited it to investigate three major aspects: the structural features of the Mastodon network of instances from a macroscopic as well as a mesoscopic perspective, to unveil the distinguishing traits of the underlying federative mechanism; the backbone of the network, to discover the essential interrelations between the instances; and the growth of Mastodon, to understand how the shape of the instance network has evolved during the last few years, also when broading the scope to account for instances belonging to other platforms. Our extensive analysis of the above aspects has provided a number of findings that reveal distinguishing features of Mastodon and that can be used as a starting point for the discovery of all the DOSN Fediverse., Comment: Accepted for publication with Applied Network Science, Springer Open, June 12, 2021

Multilayer networks have been widely used to represent and analyze systems of interconnected entities where both the entities and their connections can be of different types. However, real multilayer networks can be difficult to analyze because of irrelevant information, such as layers not related to the objective of the analysis, because of their size, or because traditional methods defined to analyze simple networks do not have a straightforward extension able to handle multiple layers. Therefore, a number of methods have been devised in the literature to simplify multilayer networks with the objective of improving our ability to analyze them. In this article we provide a unified and practical taxonomy of existing simplification approaches, and we identify categories of multilayer network simplification methods that are still underdeveloped, as well as emerging trends., Comment: Accepted for publication in Computer Science Review, Elsevier

Embedding diversity into knowledge discovery tasks is of crucial importance to enhance the meaningfulness of the mined patterns with high-impact aspects related to novelty, serendipity, and ethics. Surprisingly, in the classic problem of influence maximization in social networks, relatively little study has been devoted to diversity and its integration into the objective function of an influence maximization method. In this work, we propose the integration of a side-information-based notion of seed diversity into the objective function of a targeted influence maximization problem. Starting from the assumption that side-information is available at node level in the general form of categorical attribute values, we design a class of monotone submodular functions specifically conceived for determining the diversity within a set of categorical profiles associated with the seeds to be discovered. This allows us to develop an efficient scalable approximate method, with a constant-factor guarantee of optimality. More precisely, we formulate the attribute-based diversity-sensitive targeted influence maximization problem under the state-of-the-art reverse influence sampling framework, and we develop a method, dubbed ADITUM, that ensures a (1-1/e-\epsilon)-approximate solution under the general triggering diffusion model. We experimentally evaluated ADITUM on five real-world networks, including comparison with methods that exploit numerical-attribute-based diversity and topology-driven diversity in influence maximization., Comment: Initially conceived: October 2018. First article-version: February 1, 2019. Last update: September 11, 2019

A multiplex network models different modes of interaction among same-type entities. In this article we provide a taxonomy of community detection algorithms in multiplex networks. We characterize the different algorithms based on various properties and we discuss the type of communities detected by each method. We then provide an extensive experimental evaluation of the reviewed methods to answer three main questions: to what extent the evaluated methods are able to detect ground-truth communities, to what extent different methods produce similar community structures and to what extent the evaluated methods are scalable. One goal of this survey is to help scholars and practitioners to choose the right methods for the data and the task at hand, while also emphasizing when such choice is problematic., Comment: 55 pages. Accepted for publication on ACM Computing Surveys in a shorter version