Formal communications such as public calls, announcements, or regulations are supposed to exhibit respect for diversity in terms of gender, race, age, and disability. However, human writers often lack adequate inclusive writing skills. For instance, they tend to overuse the masculine as a neutral form, mainly because they are self-trained on biased text examples. To overcome this issue, we propose to leverage Generative Artificial Intelligence to support inclusive language writing. Focusing on formal Italian communications, we have designed and developed an AI-assisted tool for non-inclusive text detection and reformulation. Thanks to the joint work with a team of linguistic experts, we first define a set of linguistic criteria necessary to model inclusive writing forms in Italian. Based on these criteria, we collect and annotate a dataset of Italian administrative documents enriched with fine-grained inclusive annotations. Finally, we train deep learning models on the collected data for non-inclusive language detection and inclusive language reformulation tasks. We perform quantitative and human-driven evaluations on the trained models. The best detection model correctly classifies 89% of the sentences, whereas the best reformulation model produces 73% fully correct reformulations. Both models have been integrated into a writing assistance tool acting as a text proofreader and self-learning tool for non-expert writers, namely Inclusively . Once a non-inclusive piece of text is detected, the proposed approach suggests inclusive reformulations. The tool also provides explanations of the models’ outputs to increase system transparency. Furthermore, it allows expert end-users to provide further annotations for system fine-tuning. The trained models and the writing assistance tool are publicly available for research purposes.

Concept drift is the phenomenon in which the underlying data distributions and statistical properties of a target domain change over time, leading to a degradation in model performance. Consequently, production models require continuous drift detection monitoring. Most drift detection methods to date are supervised, relying on ground-truth labels. However, they are inapplicable in many real-world scenarios, as true labels are often unavailable. Although recent efforts have proposed unsupervised drift detectors, many lack the accuracy required for reliable detection or are too computationally intensive for real-time use in high-dimensional, large-scale production environments. Moreover, they often fail to characterize or explain drift effectively. To address these limitations, we proposeDRIFTLENS, an unsupervised framework for real-time concept drift detection and characterization. Designed for deep learning classifiers handling unstructured data,DRIFTLENSleverages distribution distances in deep learning representations to enable efficient and accurate detection. Additionally, it characterizes drift by analyzing and explaining its impact on each label. Our evaluation across classifiers and data-types demonstrates thatDRIFTLENS(i) outperforms previous methods in detecting drift in 15/17 use cases; (ii) runs at least 5 times faster; (iii) produces drift curves that align closely with actual drift (correlation$\geq 0.85$); (iv) effectively identifies representative drift samples as explanations.

Media headlines shape our initial interpretation of news, framing narratives that influence societal engagement with political and social issues. Yet, they often rely on sensationalism and bias to capture readers’ attention.In this paper, we aim to uncover distinct patterns in Italian headline composition, examining how language and framing vary across political leanings. We analyze a dataset of daily Italian newspaper articles from two outlets with opposing political perspectives, anonymized as Newspaper A and Newspaper B. Our study encompasses the entire set of news and a subset of topics (n = 8) likely to contain stereotypes or clickbait headlines identified using a Large Language Model. Our methodology combines (1) a lexicometric analysis to identify characteristic words of each newspaper, and (2) the training of an accurate deep learning classifier (F 1 = 0.84) to learn specific patterns for categorizing headlines into these two perspectives and leveraging explainability techniques to extract and interpret these patterns.Our analysis reveals distinct tonal differences between the two newspapers: Newspaper A generally adopts a more balanced and nuanced approach, while Newspaper B often favors a more direct and sometimes provocative style, especially regarding topics like immigration and social justice. Additionally, Newspaper B’s headlines tend to be brief and punchy, in contrast to the longer, more detailed ones from Newspaper A. Despite these tonal differences, both outlets exhibit similar stereotypical patterns in their coverage, such as consistently emphasizing nationality and group distinctions in ways that can reinforce social stereotypes. This shared tendency suggests that, although their narrative strategies differ, both outlets could contribute to a broader pattern of stereotype reinforcement.

Abstract The accuracy and flexibility of Deep Convolutional Neural Networks (DCNNs) have been highly validated over the past years. However, their intrinsic opaqueness is still affecting their reliability and limiting their application in critical production systems, where the black-box behavior is difficult to be accepted. This work proposes EBAnO, an innovative explanation framework able to analyze the decision-making process of DCNNs in image classification by providing prediction-local and class-based model-wise explanations through the unsupervised mining of knowledge contained in multiple convolutional layers. EBAnO provides detailed visual and numerical explanations thanks to two specific indexes that measure the features’ influence and their influence precision in the decision-making process. The framework has been experimentally evaluated, both quantitatively and qualitatively, by (i) analyzing its explanations with four state-of-the-art DCNN architectures, (ii) comparing its results with three state-of-the-art explanation strategies and (iii) assessing its effectiveness and easiness of understanding through human judgment, by means of an online survey. EBAnO has been released as open-source code and it is freely available online.

AI regulations are expected to prohibit machine learning models from using sensitive attributes during training. However, the latest Natural Language Processing (NLP) classifiers, which rely on deep learning, operate as black-box systems, complicating the detection and remediation of such misuse. Traditional bias mitigation methods in NLP aim for comparable performance across different groups based on attributes like gender or race but fail to address the underlying issue of reliance on protected attributes. To partly fix that, we introduce NLPGuard, a framework for mitigating the reliance on protected attributes in NLP classifiers. NLPGuard takes an unlabeled dataset, an existing NLP classifier, and its training data as input, producing a modified training dataset that significantly reduces dependence on protected attributes without compromising accuracy. NLPGuard is applied to three classification tasks: identifying toxic language, sentiment analysis, and occupation classification. Our evaluation shows that current NLP classifiers heavily depend on protected attributes, with up to 23% of the most predictive words associated with these attributes. However, NLPGuard effectively reduces this reliance by up to 79%, while slightly improving accuracy.

Assessing the technical efficiency of a set of observations requires that the associated data composed of inputs and outputs are perfectly known. If this is not the case, then biased estimates will likely be obtained. Data Envelopment Analysis (DEA) is one of the most extensively used mathematical models to estimate efficiency. It constructs a piecewise linear frontier against which all observations are compared. Since the frontier is empirically defined, any deviation resulting from low data quality (imperfect knowledge of data or IKD) may lead to efficiency under/overestimation. In this study, we model IKD and, then, apply the so-called Hit & Run procedure to randomly generate admissible observations, following some prespecified probability density functions. Sets used to model IKD limit the domain of data associated with each observation. Any point belonging to that domain is a candidate to figure out as the observation for efficiency assessment. Hence, this sampling procedure must run a sizable number of times (infinite, in theory) in such a way that it populates the whole sets. The DEA technique is used during the execution of each iteration to estimate bootstrapped efficiency scores for each observation. We use some scenarios to show that the proposed routine can outperform some of the available alternatives. We also explain how efficiency estimations can be used for statistical inference. An empirical case study based on the Portuguese public hospitals database (2013–2016) was addressed using the proposed method.

The ability to correctly identify areas damaged by forest wildfires is essential to plan and monitor the restoration process and estimate the environmental damages after such catastrophic events. The wide availability of satellite data, combined with the recent development of machine learning and deep learning methodologies applied to the computer vision field, makes it extremely interesting to apply the aforementioned techniques to the field of automatic burned area detection. One of the main issues in such a context is the limited amount of labeled data, especially in the context of semantic segmentation. In this paper, we introduce a publicly available dataset for the burned area detection problem for semantic segmentation. The dataset contains 73 satellite images of different forests damaged by wildfires across Europe with a resolution of up to 10m per pixel. Data were collected from the Sentinel-2 L2A satellite mission and the target labels were generated from the Copernicus Emergency Management Service (EMS) annotations, with five different severity levels, ranging from undamaged to completely destroyed. Finally, we report the benchmark values obtained by applying a Convolutional Neural Network on the proposed dataset to address the burned area identification problem.

Abstract Despite the high accuracy offered by state-of-the-art deep natural-language models (e.g., LSTM, BERT), their application in real-life settings is still widely limited, as they behave like a black-box to the end-user. Hence, explainability is rapidly becoming a fundamental requirement of future-generation data-driven systems based on deep-learning approaches. Several attempts to fulfill the existing gap between accuracy and interpretability have been made. However, robust and specialized eXplainable Artificial Intelligence solutions, tailored to deep natural-language models, are still missing. We propose a new framework, named T-EBAnO, which provides innovative prediction-local and class-based model-global explanation strategies tailored to deep learning natural-language models. Given a deep NLP model and the textual input data, T-EBAnO provides an objective, human-readable, domain-specific assessment of the reasons behind the automatic decision-making process. Specifically, the framework extracts sets of interpretable features mining the inner knowledge of the model. Then, it quantifies the influence of each feature during the prediction process by exploiting the normalized Perturbation Influence Relation index at the local level and the novel Global Absolute Influence and Global Relative Influence indexes at the global level. The effectiveness and the quality of the local and global explanations obtained with T-EBAnO are proved on an extensive set of experiments addressing different tasks, such as a sentiment-analysis task performed by a fine-tuned BERT model and a toxic-comment classification task performed by an LSTM model. The quality of the explanations proposed by T-EBAnO, and, specifically, the correlation between the influence index and human judgment, has been evaluated by humans in a survey with more than 4000 judgments. To prove the generality of T-EBAnO and its model/task-independent methodology, experiments with other models (ALBERT, ULMFit) on popular public datasets (Ag News and Cola) are also discussed in detail.

Preserving diversity and inclusion is becoming a compelling need in both industry and academia. The ability to use appropriate forms of writing, speaking, and gestures is not widespread even in formal communications such as public calls, public announcements, official reports, and legal documents. The improper use of linguistic expressions can foment unacceptable forms of exclusion, stereotypes as well as forms of verbal violence against minorities, including women. Furthermore, existing machine translation tools are not designed to generate inclusive content.The present paper investigates a joint effort of the research communities of linguistics and Deep Learning Natural Language Understanding in fighting against non-inclusive, prejudiced language forms. It presents a methodology aimed at tackling the improper use of language in formal communication, with a particular attention paid to Romanic languages (Italian, in particular). State-of-the-art Deep Language Modeling architectures are exploited to automatically identify non-inclusive text snippets, suggest alternative forms, and produce inclusive text rephrasing. A preliminary evaluation conducted on a benchmark dataset shows promising results, i.e., 85% accuracy in predicting inclusive/non-inclusive communications.

Abstract Measuring the level of sustainability taking into account many contributing aspects is a challenge. In this paper, we apply a multiple criteria decision aiding framework, namely, the hierarchical-SMAA-PROMETHEE method, to assess the environmental, social, and economic sustainability of 20 European cities in the period going from 2012 to 2015. The application of the method is innovative for the following reasons: (i) it permits to study the sustainability of the mentioned cities not only comprehensively but also considering separately particular macro-criteria, providing in this way more specific information on their weak and strong points; (ii) the use of PROMETHEE and, in particular, of PROMETHEE II, avoids the compensation between different and heterogeneous criteria, that is arbitrarily assumed in value function aggregation models; finally, (iii) thanks to the application of the Stochastic Multicriteria Acceptability Analysis, the method provides more robust recommendations than a method based on a single instance of the considered preference model compatible with few preference information items provided by the Decision Maker.

We consider a problem of multi-decision sorting subject to multiple criteria. In the newly formulated decision problem, besides performances on multiple criteria, alternatives get evaluations on multiple interrelated decision attributes involving preference-ordered classes. We propose a dedicated method for dealing with such a problem, incorporating a threshold-based value-driven sorting procedure. The Decision Maker (DM) is expected to holistically evaluate a subset of reference alternatives by indicating the quality or risk level on a pre-defined scale of each decision attribute. Based on these evaluations, we construct a set of interrelated preference models, one for each decision attribute, compatible with intra- and inter-decision constraints imposed by such indirect preference information. We also formulate a new way of dealing with potentially non-monotonic criteria by discovering local monotonicity changes in different performance scale regions. The marginal value functions for criteria with unknown monotonicity are represented as a sum of two value functions assuming opposing preference directions, one non-decreasing and the other non-increasing. This permits to obtain an aggregated marginal value function with an arbitrary non-monotonic shape. The practical usefulness of the approach is demonstrated on a case study concerning risk management related to handling (i.e., production, use, manipulation, and processing) nanomaterials in different conditions. We analyze the expert judgments and discuss the inferred preference models, which can be applied to support health and safety managers in reducing the possible risk associated with the respective exposure scenario.

We propose a novel approach to multiple criteria sorting incorporating a threshold-based value-driven procedure. The parameters deciding upon the shape of marginal value functions and separating class thresholds are inferred through preference disaggregation from the Decision Maker's incomplete assignment examples and partial requirements on the type of (non-)monotonicity for each marginal value function. These types include standard monotonic shapes, level-monotonic functions, A- and V-types combining increasing and decreasing value trends, and unknown monotonicity constraints. A representative instance of the sorting model compatible with the preference information is constructed by solving a dedicated Mixed-Integer Linear Programming problem. Its complexity is controlled by minimizing the number of changes in monotonicity between all subsequent sub-intervals of marginal value functions. The assignments derived using the constructed representative model are validated against the outcomes of robustness analysis. The proposed method is applied to a real-world problem of exposure management of engineered nanomaterials. We develop a model for predicting precaution level while handling nanomaterials in certain conditions using a respirator. The model captures interrelations between ten accounted evaluation criteria, including both monotonic and non-monotonic criteria, and the recommended class assignment. This makes it suitable for the management of exposure scenarios, which have not been directly judged by the experts.

In this paper we present a new approach to rough set approximations that permits to distinguish between two kinds of “imperfect” knowledge in a joint framework: on one hand, vagueness, due to imprecise knowledge and uncertainty typical of fuzzy sets, and on the other hand, ambiguity, due to granularity of knowledge originating from the coarseness typical of rough sets. The basic idea of our approach is that each concept is represented by an orthopair, that is, a pair of disjoint sets in the universe of knowledge. The first set in the pair contains all the objects that are considered as surely belonging to the concept, while the second set contains all the objects that surely do not belong to the concept. In this context, following some previous research conducted by us on the algebra of rough sets, we propose to define as rough approximation of the orthopair representing the considered concept another orthopair composed of lower approximations of the two sets in the first orthopair. We shall apply this idea to the classical rough set approach based on indiscernibility, as well as to the dominance-based rough set approach. We discuss also a variable precision rough approximation, and a fuzzy rough approximation of the orthopairs. Some didactic examples illustrate the proposed methodology.

This paper proposes an interactive multiobjective evolutionary algorithm (MOEA) that attempts to learn a value function capturing the users' true preferences. At regular intervals, the user is asked to rank a single pair of solutions. This information is used to update the algorithm's internal value function model, and the model is used in subsequent generations to rank solutions incomparable according to dominance. This speeds up evolution toward the region of the Pareto front that is most desirable to the user. We take into account the most general additive value function as a preference model and we empirically compare different ways to identify the value function that seems to be the most representative with respect to the given preference information, different types of user preferences, and different ways to use the learned value function in the MOEA. Results on a number of different scenarios suggest that the proposed algorithm works well over a range of benchmark problems and types of user preferences.

The paper focuses on Bayesian confirmation measures used for evaluation of rules induced from data. To distinguish between many confirmation measures, their properties are analyzed. The article considers a group of symmetry properties. We demonstrate that the symmetry properties proposed in the literature focus on extreme cases corresponding to entailment or refutation of the rule's conclusion by its premise, forgetting intermediate cases. We conduct a thorough analysis of the symmetries regarding that the confirmation should express how much more probable the rule's hypothesis is when the premise is present rather than when the negation of the premise is present. As a result we point out which symmetries are desired for Bayesian confirmation measures. Next, we analyze a set of popular confirmation measures with respect to the symmetry properties and other valuable properties, being monotonicity M, Ex 1 and weak Ex 1 , logicality L and weak L. Our work points out two measures to be the most meaningful ones regarding the considered properties.

Multiple Criteria Decision Aiding (MCDA) offers a diversity of approaches designed for providing the decision maker (DM) with a recommendation concerning a set of alternatives (items, actions) evaluated from multiple points of view, called criteria. This paper aims at drawing attention of the Machine Learning (ML) community upon recent advances in a representative MCDA methodology, called Robust Ordinal Regression (ROR). ROR learns by examples in order to rank a set of alternatives, thus considering a similar problem as Preference Learning (ML-PL) does. However, ROR implements the interactive preference construction paradigm, which should be perceived as a mutual learning of the model and the DM. The paper clarifies the specific interpretation of the concept of preference learning adopted in ROR and MCDA, comparing it to the usual concept of preference learning considered within ML. This comparison concerns a structure of the considered problem, types of admitted preference information, a character of the employed preference models, ways of exploiting them, and techniques to arrive at a final ranking.

We introduce the bipolar complemented de Morgan Brouwer-Zadeh distributive lattice in order to give an algebraic model the to Dominance-based Rough Set Approach. We present also the concept of bipolar approximation space and we show how it can be ind

We present basic ideas related to application of Dominance-based Rough Set Approach (DRSA) in interactive Evolutionary Multiobjective Optimization (EMO). In the proposed methodology, the preference information elicited by the decision maker in successive iterations consists in sorting some solutions in the current population into “relatively good” and “others”, or in comparing some pairs of solutions with respect to preference. The “if ..., then ...” decision rules are then induced from this preference information using Dominance-based Rough Set Approach (DRSA). These rules are used within EMO in order to focus on populations of solutions satisfying the preferences of the decision maker, speeding up convergence to the most preferred region of the Pareto-front. The resulting interactive schemes, corresponding to the two types of preference information, are called DRSA-EMO and DRSA-EMO-PCT, respectively. The proposed methodology permits also to take into account robust concerns in multiobjective optimization.

Multicriteria decision analysis studies decision problems in which the alternatives are evaluated on several dimensions or viewpoints. In the problems we consider in this article, the scales used for assessing the alternatives with respect to a viewpoint are bipolar and univariate or unipolar and bivariate. In the former case, the scale is divided in two zones by a neutral point; a positive feeling is associated to the zone above the neutral point and a negative feeling to the zone below this point. On unipolar bivariate scales, an alternative can receive both a positive evaluation and a negative evaluation, reflecting contradictory feelings or stimuli. The article discusses procedures and models that have been proposed to aggregate multicriteria evaluations when the scale of each criterion is of one of these two types. We present both a constructive view and a descriptive view on this question; the descriptive approach is concerned with characterizations of models of preference, whereas the constructive approach aims at building preferences by questioning the decision maker. We show that these views are complementary. © 2008 Wiley Periodicals, Inc.

Decision rules induced from a data set represent knowledge patterns relating premises and decisions in ‘if … , then …’ statements. Premise is a conjunction of elementary conditions relative to independent variables and decision is a conclusion relative to dependent variables. Given a set of decision rules induced from a data set, it is useful to estimate possible effects on the dependent variables caused by an intervention on some independent variables. The authors introduce a methodology for quantifying the impact of a strategy of intervention based on a decision rule induced from data. While the usual interestingness measures of decision rules are taking into account only characteristics of universe U where they come from, the measures of efficiency of intervention depend also on characteristics of universe U′ where intervention takes place. The authors are considering the intervention on a single independent variable and on a combination of these variables.

Given a ranking of actions evaluated by a set of evaluation criteria, we construct a rough approximation of the preference relation known from this ranking. The rough approximation of the preference relation is a starting point for mining " if... then" decision rules constituting a symbolic preference model. The set of rules is induced such as to be compatible with a concordance-discordance preference model used in well-known multicriteria decision aiding methods. An application of the set of decision rules to a new set of actions gives a fuzzy outranking graph. Positive and negative flows are calculated for each action in the graph, giving arguments about its strength and weakness. Aggregation of both arguments leads to a final ranking, either partial or complete. The approach can be applied to support a multicriteria choice and ranking of actions when the input information is a ranking of some reference actions.

In this article we are considering a multicriteria classification that differs from usual classification problems since it takes into account preference orders in the description of objects by condition and decision attributes. To deal with multicriteria classification we propose to use a dominance-based rough set approach (DRSA). This approach is different from the classic rough set approach (CRSA) because it takes into account preference orders in the domains of attributes and in the set of decision classes. Given a set of objects partitioned into pre-defined and preference-ordered classes, the new rough set approach is able to approximate this partition by means of dominance relations (instead of indiscernibility relations used in the CRSA). The rough approximation of this partition is a starting point for induction of if-then decision rules. The syntax of these rules is adapted to represent preference orders. The DRSA keeps the best properties of the CRSA: it analyses only facts present in data, and possible inconsistencies are not corrected. Moreover, the new approach does not need any prior discretization of continuous-valued attributes. In this article we characterize the DRSA as well as decision rules induced from these approximations. The usefulness of the DRSA and its advantages over the CRSA are presented in a real study of evaluation of the risk of business failure. © 2002 John Wiley & Sons, Inc.