Intelligenza artificiale spiegabile per applicazioni testuali: sfide aperte e opportunità per Entity Matching e Fact-checking.
Year: 2025
Type: dissertation
Abstract: In the era of Artificial intelligence, advancements in machine and deep learning enabled the creation of models that effectively tackle a wide and rapidly expanding set of tasks.While these models frequently achieve extraordinary performance, sometimes surpassing human abilities, their complex architectures, which include millions of parameters, frequently obscure their internal reasoning.Furthermore, rather than using logic that people would find reasonable or reliable, models may achieve high accuracy by taking advantage of data artifacts.This lack of transparency becomes especially concerning in sensitive domains where AI decisions can significantly impact people's lives.AI systems can behave unfairly for various reasons: societal biases may be reflected in the training data, in the decisions made during the development, or in the systems themselves.The field of Explainable Artificial Intelligence (XAI) has emerged to address these issues, focusing on both interpreting existing models and designing inherently transparent AI architectures.This thesis explores and advances the frontiers of XAI through a systematic investigation across two critical domains: Entity-Matching (EM) and computational fact-checking.In EM, this research first explores novel techniques to explain black-box models, addressing unique challenges in this field.Traditional post-hoc XAI techniques, when directly applied to EM, do not account for its characteristics, i.e. comparing pairs of descriptions originating from independent sources.The research advances the state of the art by developing adaptations and enhancements of explanation techniques that specifically address these challenges and provide "what-if" scenarios that enable users to understand how change in input data would affect matching outcomes.These innovations improve the fidelity and trustworthiness of generated explanations.The thesis then conducts an in-depth analysis of BERT's internal architecture when applied to EM, revealing crucial insights about how this state-of-the-art model processes and leverages information in entity descriptions.Experimental evaluations reveal that: fine-tuning on EM task mainly affects the last layers of BERT; this model recognizes the unique structure of EM datasets (records consisting of entity descriptions pairs); and BERT-based EM models do not heavily rely on pair-wise semantic similarity between tokens.List of tables 6.4 Performance comparison across different models and datasets in terms of claim verification (left), explanations from the intrinsic explainers (middle), and explanations from the post-hoc explainers (right).Best explanation result for every dataset and observed model pair in bold. . . . . . . . . . . . . . .92 6.5 Performance comparison across different models and datasets in terms time per explanations.mean (std) . . . . . . . . . . . . . . . . . . . . . . . . .99 6.6 Performance metrics of LLaMa 3.1 70B versus 8B (scores in parenthesis) for Claim labelling and Evidence labelling across datasets. . . . . . . . . .103 7.1 The datasets used in the experiments.Size is the dimension of the dataset, n is the number of data points (#rows), d is the number of features (#columns), S and |S| are the name and the cardinality of the sensitive attribute. . . . . .119 Chapter 1 Introduction Overview and Scope of the ResearchDigitalization is pervading the world in various aspects, from industry to personal life.In this scenario, advancements in Artificial Intelligence (AI) enabled the creation of models that effectively address an already wide yet rapidly expanding set of tasks, from healthcare and finance to education, law, and beyond.Increased automation in numerous applications is made possible by these models, which perform on par with or even better than humans.However, the rapid advancement of this technology is surpassing our capacity to comprehend it completely.These systems exploit complex machine-learning architectures that reach billions of parameters.Therefore, when looking into their internal architectures, their operation becomes obscure and incomprehensible.As a result, these models are employed in "black-box" mode, which presents issues on several levels since they are used without considering how they leverage data to make predictions.It is not possible to audit and validate their decision-making processes.It's impractical to know with certainty how these models are using the data and whether they are leveraging subtle patterns within data, i.e. patterns that can sometimes deceive reliable, logical reasoning and instead reflect biases or artifacts in the data itself.The only way to assess their quality is to check the predictions made, regardless of the internal reasoning.This "black-box" nature erodes trust and makes it more difficult to figure out how to improve them.For sensitive applications, such as healthcare, criminal justice, or financial lending, AI decisions could significantly affect people.Some form of explanation and justification of decisions helps to build trust and may be necessary for applying these models; regulatory standards (e.g., GDPR, AI Act) also increasingly require transparent and interpretable AI.Additionally, societal biases present in the training data, in the decisions made during development, or biases in the systems themselves tend to make AI act unfairly, causing Automated Fact-Checking Fake news is thought to have an annual direct cost to the economy of about $78 billion across a variety of sectors.These costs span financial misinformation, reputation management, public health misinformation, online platform safety, brand safety, and political spending.Fake news is further diminishing public trust in specific sectors, trust in the news media has dropped from 55% in 2015 to 32% in 2019.The risks
Authors: Guerra, Francesco, Rovati, Luigi, Baraldi, Andrea
Institution: University of Modena and Reggio Emilia
Language: Italian
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6991958994
