Search

In recent years, word embeddings (WEs) have proven relevant for studying differences and similarities among job professions and skills required by the labour market across countries, providing valuable insights about the labour market dynamics to support policy and decision-making. In such a scenario, aligning WEs constructed across different countries and languages becomes key to allowing experts to reason on the labour market, catching technological and cultural shifts across borders. This paper proposes MEAL, an unsupervised method for aligning monolingual embeddings. Our approach selects a seed lexicon of anchors, i.e. words with the same meaning in both corpora that will be used as pivots in the alignment, without assuming a priori semantic similarities. Indeed, unlike previous literary works, to asses this relationship MEAL takes into account the semantic similarity between the neighbour of the two words in the WE space. Particularly, it chooses optimal anchors that are less susceptible to meaning shift. We deploy MEAL within the research framework of a European H-2020 Project that aims to use AI technologies to predict the future of the European labour market. Specifically, we apply it to the embeddings we train on 7+ millions of Online Job Advertisements (OJAs) collected in 2022. As a main outcome, MEAL allows stakeholders and policymakers (i) to estimate job similarities in Online Labour Markets across Europe, facilitating the assessment of how well these markets align with the taxonomy outlined by the official European Skills and Competences taxonomy, and (ii) to obtain indicators to support a data-driven policy design at a very fine-grained territorial level.

The utilization of word embeddings—powerful models computed through Neural Network architectures that encode words as vectors—has witnessed rapid growth across various Natural Language Processing applications, encompassing semantic analysis, information retrieval, dependency parsing, question answering, and machine translation. The efficacy of these tasks is strictly linked to the quality of the embeddings, underscoring the critical importance of evaluating and selecting optimal embedding models. While established procedures and benchmarks exist for intrinsic evaluation, the authors note a conspicuous absence of comprehensive evaluations of intrinsic embedding quality across multiple tasks. This paper introduces vec2best, a unified tool encompassing state-of-the-art intrinsic evaluation tasks across diverse benchmarks. vec2best furnishes the user with an extensive evaluation of word embedding models. It represents a framework for evaluating word embeddings trained using various methods and hyper-parameters on a range of tasks from the literature. The tool yields a holistic evaluation metric for each model called the PCE (Principal Component Evaluation). We conducted evaluations on 135 word embedding models, trained using GloVe, fastText, and word2vec, across four tasks integrated into vec2best (similarity, analogy, categorization, and outlier detection), along with their respective benchmarks. Additionally, we leveraged vec2best to optimize embedding hyper-parameter configurations in a real-world scenario. vec2best is conveniently accessible as a pip-installable Python package.

Explaining how two machine learning classification models differ in their behaviour is gaining significance in eXplainable AI, given the increasing diffusion of learning-based decision support systems. Human decision-makers deal with more than one machine learning model in several practical situations. Consequently, the importance of understanding how two machine learning models work beyond their prediction performances is key to understanding their behaviour, differences, and likeness. Some attempts have been made to address these problems, for instance, by explaining text classifiers in a time-contrastive fashion. In this paper, we present MERLIN, a novel eXplainable AI approach that provides contrastive explanations of two machine learning models, introducing the concept of model-contrastive explanations. We propose an encoding that allows MERLIN to work with both text and tabular data and with mixed continuous and discrete features. To show the effectiveness of our approach, we evaluate it on an extensive set of benchmark datasets. MERLIN is also implemented as a python-pip package.

Word embeddings have proven extremely useful across many NLP applications in recent years. Several key linguistic tasks, such as machine translation and transfer learning, require comparing distributed representations of words belonging to different vector spaces within or among different domains and languages to be aligned, known as embedding alignment. To this end, several existing methods exploit words that are supposed to have the same meaning in the two corpora, called seed lexicon or anchors, as reference points to map one embedding into the other. All those methods consider only the word that is supposed to have the same meaning in the two spaces to choose anchors, while its neighbours or similar words are neglected. We propose SeNSe, an unsupervised method for aligning monolingual embeddings, generating a bilingual dictionary composed of words with the most similar meaning among word vector spaces. Our approach selects a seed lexicon of words used in the same context in both corpora without assuming a priori semantic similarities. We compare our method with well-established benchmarks showing SeNSe outperforms state-of-the-art (SOTA) methods for embedding alignment on bilingual lexicon extraction in most cases.

A research area of NLP is known as keyphrases extraction, which aims to identify words and expressions in a text that comprehensively represent the content of the text itself. In this study, we introduce a new approach called KRAKEN (Keyphrease extRAction maKing use of EmbeddiNgs). Our method takes advantage of widely used NLP techniques to extract keyphrases from a text in an unsupervised manner and we compare the results with well-known benchmark datasets in the literature. The main contribution of this work is developing a novel approach for keyphrase extraction. Both natural language text preprocessing techniques and distributional semantics techniques, such as word embeddings, are used to obtain a vector representation of the texts that maintains their semantic meaning. Through KRAKEN, we propose and design a new method that exploits word embedding for identifying keyphrases, considering the relationship among words in the text. To evaluate KRAKEN, we employ benchmark datasets and compare our approach with state-of-the-art methods. Another contribution of this work is the introduction of a metric to rank the identified keyphrases, considering the relatedness of both the words within the phrases and all the extracted phrases from the same text.

With the increasing adoption of Artificial Intelligence (AI) for decision-making processes by companies, developing systems that behave fairly and do not discriminate against specific groups of people becomes crucial. Reaching this objective requires a multidisciplinary approach that includes domain experts, data scientists, philosophers, and legal experts, to ensure complete accountability for algorithmic decisions. In such a context, Explainable AI (XAI) plays a key role in enabling professionals from different backgrounds to comprehend the functioning of automatized decision-making processes and, consequently, being able to identify the presence of fairness issues. This paper presents FairX, an innovative approach that uses Group-Contrastive (G-contrast) explanations to estimate whether different decision criteria apply among distinct population subgroups. FairX provides actionable insights through a comprehensive explanation of the decision-making process, enabling businesses to: detect the presence of direct discrimination on the target variable and choose the most appropriate fairness framework.

The field of explainable artificial intelligence (XAI) is gaining increasing importance in recent years. As a consequence, several surveys have been published to explore the current state of the art on this topic. One aspect that seems to be overlooked by these works is the applied presentation methods and, specifically, the role of natural language in generating the final explanations. This survey reviews 70 XAI papers published between 2006 and 2021 and evaluates their readiness with respect to natural language explanations. Thus, together with a set of hierarchical criteria, we define a multi-criteria decision-making model. Finally, we conclude that only a handful of recent XAI works either considered natural language explanations to approach final users (see, e.g.,(Bennetot et al., 2021)) or implemented a method capable of generating such explanations.

In the era of rapid growth and transformation driven by artificial intelligence across various sectors, which is catalyzing the fourth industrial revolution, this research is directed toward harnessing its potential to enhance the efficiency of decision-making processes within organizations. When constructing machine learning-based decision models, a fundamental step involves the conversion of domain knowledge into causal-effect relationships that are represented in causal graphs. This process is also notably advantageous for constructing explanation models. We present a method for generating causal graphs that integrates the strengths of Large Language Models (LLMs) with traditional causal theory algorithms. Our method seeks to bridge the gap between AI’s theoretical potential and practical applications. In contrast to recent related works that seek to exclude the involvement of domain experts, our method places them at the forefront of the process. We present a novel pipeline that streamlines and enhances domain-expert validation by providing robust causal graph proposals. These proposals are enriched with transparent reports that blend foundational causal theory reasoning with explanations from LLMs.

The growing number of machine learning-based Cyber-Physical Systems (CPSs) and their ability to adapt and to learn is gaining research interest in several biomedical applications. The use of learning capabilities allows CPSs to interact and analyse their environment, learn from patterns, and perform highly complex prediction tasks. However, while on the one side the use of machine learning acts as a flywheel to the diffusion of those systems, on the other side exposes them to the problem of transparency and interpretability that affect any machine-learning-based systems. This, in critical fields like medicine, is just as important as models' performances, in order to understand their behaviour, errors, and to garner user trust. In this paper we investigate the role of state-of-the-art explainable AI techniques in the field of cyber-physical systems.

The recent growth of black-box machine-learning methods in data analysis has increased the demand for explanation methods and tools to understand their behaviour and assist human-ML model cooperation. In this paper, we demonstrate ContrXT, a novel approach that uses natural language explanations to help users to comprehend how a back-box model works. ContrXT provides time contrastive (t-contrast) explanations by computing the differences in the classification logic of two different trained models and then reasoning on their symbolic representations through Binary Decision Diagrams. ContrXT is publicly available at ContrXT.ai as a python pip package.

Ranking is required for many real applications, such as search, personalisation, recommendation, and filtering. Recent research has focused on developing reliable ranking algorithms that maintain fairness in their outcomes. However, only a few consider multiple protected groups since this extension introduces significant challenges. While useful in the research sector, considering only one binary sensitive feature for handling fairness is inappropriate when the algorithm must be deployed responsibly in real-world applications. Our work is built on top of Multinomial FA*IR, a Fair Top-k ranking with multiple protected groups, which we extend to provide users the option to balance fairness and utility, adapting the final ranking accordingly. Our experimental results show that alternative better solutions overlooked by Multinomial FA*IR may be found through our approach without violating fairness boundaries. The code of the implemented solution and the experiments are publicly available to the community as a GitHub repository.

Several studies have addressed the importance of context and users’ knowledge and experience in quantifying the usability and effectiveness of the explanations generated by explainable artificial intelligence (XAI) systems. However, to the best of our knowledge, no component-agnostic system that accounts for this need has yet been built. This paper describes an approach called ConvXAI, which can create a dialogical multimodal interface for any black-box explainer by considering the knowledge and experience of the user. First, we formally extend the state-of-the-art conversational explanation framework by introducing clarification dialogue as an additional dialogue type. We then implement our approach as an off-the-shelf Python tool. To evaluate our framework, we performed a user study including 45 participants divided into three groups based on their level of technology use and job function. Experimental results show that (i) different groups perceive explanations differently; (ii) all groups prefer textual explanations over graphical ones; and (iii) ConvXAI provides clarifications that enhance the usefulness of the original explanations.

Social media recommendation differs from traditional recommendation approaches as it needs considering not only the content information and users' similarities, but also users' social relationships and behavior within an online social network as well. In this article, a recommender system - designed for big data applications - is used for providing useful recommendations in online social networks. The proposed technique represents a collaborative and user-centered approach that exploits the interactions among users and generated multimedia contents in one or more social networks in a novel and effective way. The experiments performed on data collected from several online social networks show the feasibility of the approach towards the social media recommendation problem.

While word embeddings have been showing their effectiveness in capturing semantic and lexical similarities in a large number of domains, in case the corpus used to generate embeddings is associated with a taxonomy (i.e., classification tasks over standard de-jure taxonomies) the common intrinsic and extrinsic evaluation tasks cannot guarantee that the generated embeddings are consistent with the taxonomy. This, as a consequence, sharply limits the use of distributional semantics in those domains. To address this issue, we design and implement MEET, which proposes a new measure -HSS- that allows evaluating embeddings from a text corpus preserving the semantic similarity relations of the taxonomy.

We propose a recommender system that, starting from a set of users skills, identifies the most suitable jobs as they emerge from a large text of Online Job Vacancies (OJVs). To this aim, we process 2.5M+ OJVs posted in three different countries (United Kingdom, France and Germany), generating several embeddings and performing an intrinsic evaluation of their quality. Besides, we compute a measure of skill importance for each occupation in each country, the Revealed Comparative Advantage (rca). The best vector models, together with the rca, are used to feed a graph database, which will serve as the keystone for the recommender system. Finally, a user study of 10 validates the effectiveness of skills2job, both in terms of precision and nDGC.

Taxonomies provide a structured representation of semantic relations between lexical terms, acting as the backbone of many applications. The research proposed herein addresses the topic of taxonomy enrichment using an”human-in-the-loop” semi-supervised approach. I will be investigating possible ways to extend and enrich a taxonomy using corpora of unstructured text data. The objective is to develop a methodological framework potentially applicable to any domain.

Over the past decade, the development of Machine Learning (ML) algorithms to replace human decisions has raised concerns about potential bias issues. At the same time, significant advances have been made in the study of fairness in classification to prevent discrimination. However, only a few of these works have investigated how those techniques can have an impact on the society and if they remain reliable over time. This work aims (i) to shed light on traditional group fairness mitigation strategies that fail in real-time environments when financial data drifts affect only some classes of sensitive attributes; (ii) to investigate a strategy that encodes the human behaviour while retraining the model over time, favoring the convergence between individual and group fairness; (iii) to put the basis of strategies based on eXplainable AI (XAI), to monitor the evolution of financial gaps between different population subgroups, like gender or race, observing whether the mitigation strategy is bringing benefits to society. Preliminary results are provided, processing about 800k personal loan granting from 2016 to 2019 for Intesa Sanpaolo bank.

Explainable Artificial Intelligence (XAI) is gaining interests in both academia and industry, mainly thanks to the proliferation of darker more complex black-box solutions which are replacing their more transparent ancestors. Believing that the overall performance of an XAI system can be augmented by considering the end-user as a human being, we are studying the ways we can improve the explanations by making them more informative and easier to use from one hand, and interactive and customisable from the other hand.

The demand for transparency and fairness in AI-based decision-making systems is constantly growing. Organisations need to be assured that their applications, based on these technologies, behave fairly, without introducing negative social implications in relation to sensitive attributes such as gender or race. Since the notion of fairness is context dependent and not uniquely defined, studies in the literature have proposed various formalisation. In this work, we propose a novel, flexible, discrimination-aware decision-tree that allows the user to employ different fairness criteria depending on the application domain. Our approach enhances decision-tree classifiers to provide transparent and fair rules to final users.

In the last few years, we have been witnessing the increasing deployment of machine learning-based systems, which act as black boxes whose behaviour is hidden to end-users. As a side-effect, this contributes to increasing the need for explainable methods and tools to support the coordination between humans and ML models towards collaborative decision-making. In this paper, we demonstrate ContrXT, a novel tool that computes the differences in the classification logic of two distinct trained models, reasoning on their symbolic representation through Binary Decision Diagrams. ContrXT is available as a pip package and API.

Lexical taxonomies are widely used to foster information retrieval and exchange in several domains and applications. When there are multiple taxonomies, heterogeneity among them is a severe problem for efficient collaboration processes. In this paper, we propose WETA, a domain-independent, knowledge-poor method for automatic taxonomy alignment via word embeddings. WETA associates all the leaf terms of the origin taxonomy to one or many concepts in the destination taxonomy, employing a scoring function, which merges the score of a hierarchical method based on cosine similarity and the score of a classification task. WETA is developed in the context of an EU Grant aiming at bridging the national taxonomies of EU countries towards the European Skills, Competences, Qualifications and Occupations taxonomy (ESCO) using AI Algorithms. The results, validated within the EU project activities for bridging the Italian occupation taxonomy CP and ESCO, confirm the usefulness of WETA in supporting the automatic alignment of national labor taxonomies. WETA reaches a 0.8 accuracy on recommending top-5 occupations and a wMRR of 0.72. WETA reduces the human effort needed for building a mapping from scratch: it would allow domain experts to concentrate on the validation task and decrease the incoherence due to multiple judgments. It would also make the approach reproducible and transparent to policymakers.

Lexical taxonomies and distributional representations are largely used to support a wide range of NLP applications, including semantic similarity measurements. Recently, several scholars have proposed new approaches to combine those resources into unified representation preserving distributional and knowledge-based lexical features. In this paper, we propose and implement TaxoVec, a novel approach to selecting word embeddings based on their ability to preserve taxonomic similarity. In TaxoVec, we first compute the pairwise semantic similarity between taxonomic words through a new measure we previously developed, the Hierarchical Semantic Similarity (HSS), which we show outperforms previous measures on several benchmark tasks. Then, we train several embedding models on a text corpus and select the best model, that is, the model that maximizes the correlation between the HSS and the cosine similarity of the pair of words that are in both the taxonomy and the corpus. To evaluate TaxoVec, we repeat the embedding selection process using three other semantic similarity benchmark measures. We use the vectors of the four selected embeddings as machine learning model features to perform several NLP tasks. The performances of those tasks constitute an extrinsic evaluation of the criteria for the selection of the best embedding (i.e. the adopted semantic similarity measure). Experimental results show that (i) HSS outperforms state-of-the-art measures for measuring semantic similarity in taxonomy on a benchmark intrinsic evaluation and (ii) the embedding selected through TaxoVec achieves a clear victory against embeddings selected by the competing measures on benchmark NLP tasks. We implemented the HSS, together with other benchmark measures of semantic similarity, as a full-fledged Python package called TaxoSS, whose documentation is available at https://pypi.org/project/TaxoSS.

The need for explanations of ML systems is growing as new models outperform their predecessors while becoming more complex and less comprehensible for their end-users. Though several XAI methods have been proposed in recent years, not enough attention was paid to explaining how models change their behaviour in contrast with previous ones (e.g., due to retraining). In such cases, an XAI system should explain why the model changes its predictions concerning past outcomes. Capturing and understanding such differences is crucial, as the need for trust is key in any application to support human-AI decision-making processes. This is the idea of ContrXT, a novel approach that (i) traces the decision criteria of a black box text classifier by encoding the changes in the decision logic through Binary Decision Diagrams. Then (ii) it provides global, model-agnostic, Time-Contrastive (T-contrast) explanations in natural language, estimating why – and to what extent – the model has modified its behaviour over time. We implemented and evaluated ContrXT over several supervised ML models trained on a benchmark dataset and a real-life application, showing it is effective in catching majorly changed classes and in explaining their variation through a user study. The approach has been implemented, and it is available to the community both as a python package and through REST API, providing contrastive explanations as a service.

Labor Market Intelligence (LMI) is an emerging field of study that has been gaining interest as it allows employing Artificial Intelligence (AI) algorithms on labor market information. The goal of LMI is to support decision and policy making activities (e.g., real-time monitoring of Online Job Vacancies (OJV) across countries, forecast skill requested within vacancies, compare similar labor markets across borders, etc.). The European project in which this work is framed can be placed in this field, as it aims at collecting and classifying millions of OJVs from 28 EU Countries, handling 32 languages, and also extracting the requested skills. The result is a huge amount of information useful for understanding labor market dynamics and trends. The goal of this work is to realize a system - namely GraphLMI - that organizes such Labor Market information as a graph, enabling the representation of occupation/skill relevance and similarity over the European Labor Market; another goal is to enrich the European standard taxonomy of occupations and skills (ESCO) to better fit the labor market expectations. We formalize and design the GraphLMI data model, then we implement it as a graph-database, generated by processing 5.3+ million OJVs composed by free text and collected between 2018 and 2019 for France, Germany, and the United Kingdom. Finally, we show how the resulting knowledge can be queried through a declarative query language to understand, compare and evaluate country-based labor market dynamics for supporting policy and decision making activities at European level.

We propose a recommender system that, starting from a set of users’ skills, identifies the most suitable jobs as they emerge from a large dataset of Online Job Vacancies (OJVs). To this aim, we process 2.5M+ OJVs posted in three different countries (United Kingdom, France, and Germany), training several embeddings and performing an intrinsic evaluation of their quality. Besides, we compute a measure of skill importance for each occupation in each country, the Revealed Comparative Advantage (rca). The best vector model, one for each country, together with the rca, is used to feed a graph database, which will serve as the keystone for the recommender system. Results are evaluated through a user study of 10 labor market experts, using P@3 and nDCG as scores. Results show a high precision for the recommendations provided by skills2job, and the high values of nDCG (0.985 and 0.984 in a [0,1] range) indicate a strong correlation between the experts’ scores and the rankings generated by skills2job.

Taxonomies are the mainstay of the semantic web as they aim at organising knowledge in concepts linked by IS-A relationships. However, keeping such hierarchies updated and able to represent the domain from which they have been drawn is still a time-consuming, costly and error prone activity. Here, word embeddings have proven to be effective in catching lexicon and semantic similarities to enrich taxonomies from text data. This, in turn, would require to evaluate the generated embeddings to estimate the extent to which they encode the semantic similarity derived from the hierarchy itself. In this paper, we propose and implement MEET-LM, a methodology that aims at generating and evaluating embeddings from a text corpus preserving the co-hyponymy relations synthesised from a domain-specific taxonomy. We apply MEET-LM to a real-life dataset of 2M+ vacancies related to ICT-jobs, framed within the research activities of an EU project that collects millions of Online Job Vacancies and classifies them within the European standard hierarchy ESCO. To show MEET-LM is useful in practice, we also trained a neural network to classify co-hyponym relations using the selected embeddings as features. Our experiments reach 99.4% of accuracy and 86.5% of f1-score.

The widespread use of Online Social Networks has also involved the scientific field in which researchers interact each other by publishing or citing a given paper. The huge amount of information about scientific research documents has been described through the term Big Scholarly Data. In this paper we propose a framework, namely Discovery Information using COmmunity detection (DICO), for identifying overlapped communities of authors from Big Scholarly Data by modeling authors' interactions through a novel graph-based data model combining jointly document metadata with semantic information. In particular, DICO presents three distinctive characteristics:i) the co-authorship network has been built from publication records using a novel approach for estimating relationships weight between users;ii) a new community detection algorithm based on Node Location Analysis has been developed to identify overlapped communities;iii) some built-in queries are provided to browse the generated network, though any graph-traversal query can be implemented through the Cypher query language. The experimental evaluation has been carried out to evaluate the efficacy of the proposed community detection algorithm on benchmark networks.Finally, DICO has been tested on a real-world Big Scholarly Dataset to show its usefulness working on the DBLP+AMiner dataset, that contains 1.7M+ distinct authors, 3M+ papers, handling 25M+ citation relationships.

Taxonomies provide a structured representation of semantic relations between lexical terms. In the case of standard official taxonomies, the refinement task consists of maintaining them updated over time, while preserving their original structure. To date, most of the approaches for automated taxonomy refinement rely on word vector models. However, none of them considers to what extent those models encode the taxonomic similarity between words. Motivated by this, we propose and implement TaxoRef, a methodology that (i) synthesises the semantic similarity between taxonomic elements through a new metric, namely HSS, (ii) evaluates to what extent the embeddings generated from a text corpus preserve those similarity relations and (iii) uses the best embedding resulted from this evaluation to perform taxonomy refinement. TaxoRef is a part of the research activity of a 4-year EU project that collects and classifies millions of Online Job Ads for the 27+1 EU countries. It has been tested over 2M ICT job ads classified over ESCO, the European standard occupation and skill taxonomy. Experimental results confirm (i) the HSS outperforms previous metrics for semantic similarity in taxonomies, and (ii) TaxoRef accurately encodes similarities among occupations, suggesting a refinement strategy.

In this paper, we present skills2graph, a tool that, starting from a set of users' professional skills, identifies the most suitable jobs as they emerge from a large corpus of 2.5M+ Online Job Vacancies (OJVs) posted in three different countries (the United Kingdom, France, and Germany). To this aim, we rely both on co-occurrence statistics - computing a count-based measure of skill-relevance named Revealed Comparative Advantage (rca) - and distributional semantics - generating several embeddings on the OJVs corpus and performing an intrinsic evaluation of their quality. Results, evaluated through a user study of 10 labor market experts, show a high P@3 for the recommendations provided by skills2graph, and a high nDCG (0.985 and 0.984 in a [0,1] range), that indicates a strong correlation between the experts' scores and the rankings generated by skills2graph.

This paper studies the relationship between a job vacancy population obtained from an international project based on web scraping of online job vacancies and job vacancies derived by the Eurostat job vacancy rate referred to Italian national statistics survey. We compare the time series properties of both time series between 2013 and 2018, globally and in each specific sector of economic activity. Using time series decomposition and cointegration analyses, we find that, apart some specific sector, the web and national statistics office vacancies data present similar time series properties, suggesting that both time series represent the same underlying phenomenon, namely the real number of new vacancies in the Italian economy. The study confirms promising frontiers to measure in real time aggregate demand in the labour market based on web data.

The rise of Big Data era calls for more efficient and effective Data Exploration and analysis tools. In this respect, the need to support advanced analytics on Big Data is driving data scientist’ interest toward massively parallel distributed systems and software platforms, such as Map-Reduce and Spark, that make possible their scalable utilization. However, when complex data mining algorithms are required, their fully scalable deployment on such platforms faces a number of technical challenges that grow with the complexity of the algorithms involved. Thus algorithms, that were originally designed for a sequential nature, must often be redesigned in order to effectively use the distributed computational resources. In this paper, we explore these problems, and then propose a solution which has proven to be very effective on the complex hierarchical clustering algorithm CLUBS+. By using four stages of successive refinements, CLUBS+ delivers high-quality clusters of data grouped around their centroids, working in a totally unsupervised fashion. Experimental results confirm the accuracy and scalability of CLUBS+ on platforms tailored for Big Data management.

Taxonomies are the mainstay of a wide range of applications in the semantic web and provide formal support for syntactic and semantic exchanges. Hence, they are pivotal for machine understanding in natural language processing and business-oriented tasks. However, keeping such hierarchies updated and able to represent the domain from which they have been drawn, is still a time-consuming and error-prone activity. Here, word embeddings have proven to be effective in catching lexicon and semantic similarities to enrich taxonomies from text data. This, in turn, would require to evaluate the generated embeddings to estimate the extent to which they encode the semantic similarity derived from the hierarchy itself. In this paper, we propose and implement MEET a methodology that aims at generating and evaluating embeddings from a text corpus preserving the semantic similarity relations synthesised from a taxonomy. To this aim, we develop a new measure - namely Hierarchical Semantic Similarity (HSS) - to compute the semantic similarity between taxonomic elements. Experimental results show that (i) HSS outperforms state-of-the-art measures for measuring semantic similarity in taxonomy and (ii) the embedding selected through MEET obtains a clear victory against embeddings selected by the literature on benchmark tasks. We made available an open-source repository 1 1 gitlab anonymised project with all the material employed in this research, including the HSS for 35,000 word pairs.

Discharge letters (DiL) are used within any hospital Information Systems to track diseases of patients during their hospitalisation. Such records are commonly classified over the standard taxonomy made by the World Health Organization, that is the International Statistical Classification of Diseases and Related Health Problems (ICD-10). Particularly, classifying DiLs on the right code is crucial to allow hospitals to be refunded by Public Administrations on the basis of the health service provided. In many practical cases the classification task is carried out by hospital operators, that often have to cope under pressure, making this task an error-prone and time-consuming activity. This process might be improved by applying machine learning techniques to empower the clinical staff. In this paper, we present a system, namely eXDiL, that uses a two-stage Machine Learning and XAI-based approach for classifying DiL data on the ICD-10 taxonomy. To skim the common cases, we first classify automatically the most frequent codes. The codes that are not automatically discovered will be classified into the appropriate chapter and given to an operator to assess the correct code, in addition to an extensive explanation to help the evaluation, comprising of an explainable local surrogate model and a word similarity task. We also show how our approach will be beneficial to healthcare operators, and in particular how it will speed up the process and potentially reduce human errors.

We demonstrate GraphDBLP, a tool to allow researchers for querying the DBLP bibliography as a graph. The DBLP source data were enriched with semantic similarity relationships computed using wordembeddings. A user can interact with the system either via a Web-based GUI or using a shell-interface, both provided with three parametric and pre-defined queries. GraphDBLP would represent a first graph-database instance of the computer scientist network, that can be improved through new relationships and properties on nodes at any time, and this is the main purpose of the tool, that is freely available on Github. To date, GraphDBLP contains 5+ million nodes and 24+ million relationship.