1,720,981 research outputs found
Lexicon- and Learning-based Techniques for Emotion Recognition in Social Contents
No full textIn tempi recenti, la diffusione massiva dei social network ha reso disponibili grandi quantità di contenuti generati dagli utenti, i quali spesso contengono informazioni autentiche in merito alle emozioni ed ai pensieri delle persone. L'analisi di tali contenuti attraverso tecniche di emotion recognition offre informazioni preziose in merito alla percezione di prodotti, servizi ed eventi, permettendo di estendere i tradizionali processi di Business Intelligence. A tal fine, nella presente tesi sono proposte tecniche innovative, basate sia sull'uso di risorse lessicali (lexicon-based) che di algoritmi di machine learning (learning-based), per l'emotion recognition, in particolare con applicazioni a contenuti social. Per quanto riguarda gli approcci lexicon-based, vengono estese le tecniche classiche introducendo due algoritmi, rispettivamente per la disambiguazione delle parole polisemiche e l'analisi delle frasi contenenti negazioni. Il primo algoritmo individua la variante semantica di una parola polisemica più adatta al contesto cercando il percorso più breve, all'interno di una risorsa lessicale, fra la parola polisemica e le parole vicine. Il secondo, invece, individua lo scope della negazione mediante analisi dell'albero sintattico.
La tesi presenta inoltre la progettazione e l'implementazione di una piattaforma basata su approcci lexicon-based per l'analisi delle opinioni espresse dagli utenti in vari social network.
Per quanto concerne gli approcci learning-based, è stata definita una metodologia per la creazione autoamtica di corpora annotati attraverso l'analisi delle espressioni facciali in video sottotitolati. La metodologia propone l'utilizzo di numerose tecniche di video preprocessing, per il filtraggio dei frame non rilevanti, e di un classificatore di espressioni facciali, implementabile mediante due approcci differenti.
Le tecniche proposte sono state valutate sperimentalmente attraverso numerosi dataset e i risultati sono promettenti.In recent years, the massive diffusion of social networks has made available large amounts of user-generated content, which often contains authentic information about people's emotions and thoughts.
The analysis of such content through emotion recognition provides valuable insights into people's feeling about products, services and events, and allows to extend traditional processes of Business Intelligence.
To this purpose, in the present work we propose novel techniques for lexicon- and learning-based emotion recognition, in particular for the analysis of social content. For what concerns lexicon-based approaches, the present work extends traditional techniques by introducing two algorithms for the disambiguation of polysemous words and the correct analysis of negated sentences. The former algorithm detects the most suitable semantic variant of a polysemous word with respect of its context, by searching for the shortest path in a lexical resource from the polysemous word to its nearby words. The latter detects the right scope of negation through the analysis of parse trees.
Moreover, the paper describes the design and implementation of an application of the lexicon-based approach, that is a full-fledged platform for information discovery from multiple social networks, which allows for the analysis of users' opinions and characteristics and is based on Exploratory Data Analysis.
For what concerns learning-based approaches, a methodology has been defined for the automatic creation of annotated corpora through the analysis of facial expressions in subtitled videos. The methodology is composed of several video preprocessing techniques, with the purpose of filtering out irrelevant frames, and a facial expression classifier, which can be implemented using two different approaches.
The proposed techniques have been experimentally evaluated using several real-world datasets and the results are promising
Semantic Disambiguation in a Social Information Discovery System
No full textSentiment Analysis of microblog content calls for specific tools able to cope with the dynamic nature of information published in social networks, and the intrinsic complexity and ambiguity of human language. In this work we introduce a Word Sense Disambiguation (WSD) algorithm for polysemous word disambiguation which uses a dictionary-based approach to determine the most fitting meaning of a term, basing on nearby words in the sentence. The work is a part of a Business Intelligence system for the integration and discovery of social information from multiple social networks, namely Facebook and Twitter. In this work we also extend the number of sources taking into account LinkedIn, as it is typically used by professionals, and discussions thereof provide added benefits when a non-generic evaluation of the topic to be analyzed is required
Extraction of User Daily Behavior from Home Sensors through Process Discovery
Get PDFIn the last years, the wide availability on the market of low-cost smart devices paved the way for the development of smart environments, which offer an unprecedented opportunity to recognize patterns of activities from the large amount of collected data, with the ultimate aim of monitoring user behavior. In this paper, we propose a methodology which relies on Process Discovery techniques to analyze sensor data in terms of activation sequences and to discover process models representing user’s behavioral patterns. The extraction of such models is valuable not only in the perspective of gaining a better insight on how a certain task is performed, but also in supporting novel smart services. In order to evaluate the effectiveness of the approach, in this work we also consider a real-world case study set in an ambient assisted living environment
An emotion-aware search engine for multimedia content based on deep learning algorithms
Get PDFNowadays, large amounts of unstructured data are available online. Such data often contain users’ emotions and feelings about a variety of topics but their retrieval and selection on the basis of an emotional perspective are usually unfeasible through traditional search engines, which only rank web content according to its relevance with respect to a given search keyword. For this reason, in the present work we introduce the architecture of a novel emotion-aware search engine that can return search results ranked on the basis of seven human emotions. Using this system, users can benefit from a more advanced semantic search that also takes into account emotions. The system uses emotion recognition algorithms based on deep learning to extract emotion vectors from texts, images and videos and then populates an emotional index to allow users to visualise results related to given emotions. We also discuss and evaluate different deep learning models for building emotional indexes from texts, images and videos
Semantic Representation of Key Performance Indicators Categories for Prioritization
Get PDFKey Performance Indicators (KPIs) are crucial tools that are remarkably used to evaluate business performance. Recently, the management of KPIs has fascinated the focus of both academic and business professionals, and that lead to the development of research on various methods dealing with issues such as modeling, maintenance, and expressiveness of KPIs. As a need for organizations and processes to adapt to continuously changing demands, the KPIs used to measure their effectiveness evolve too. In order to make KPI management easier, this research aims to define the best sequence of KPIs evaluation based on semantic relations. After an extensive analysis of the literature on KPIs ontologies, it proposes the idea of KPIs prioritization on the basis of relations among different categories of kpis established by a KPIs ontology. Our approach can be used independently from the particular KPI’s management strategy being employed
Predicting students’ academic performance based on early career behaviors:a process-aware approach
Get PDFEducational Process Mining aims at supporting educational processes by leveraging historical data of students’ behaviors. In this work, we show how to leverage behaviors characterizing students’ first year of university to predict whether they will graduate on time. We transform the sequence of exams a student has taken into an Instance Graph modeling both sequential and concurrent behaviors. We then compare two prediction approaches, i.e., i) providing the graphs as input to a graph neural network, and ii) extracting relevant subprocesses used as features for traditional machine learning approaches. The results show that graph neural network performs best, achieving a 93.56% accuracy against the 82.43% achieved by machine learning approaches using subprocesses. Nevertheless, we advocate that these subprocesses have value both to provide descriptive insights on students’ early careers and to explain the results of the graph neural network
Metrics of Parallel Complexity of Operational Business Processes
No full textThis paper addresses the problem of quantifying the parallelism in a business process. Having a synthetic metric to quantify the parallelism of a process may provide an assessment of the complexity of the process and guide certain design choice. In the present paper we discuss the advantages and disadvantages of two metrics presented in the literature, as well of two novel metrics that leverage on the notion of Instance Graph. Analysis is performed by means of use cases that are representative of operational business processes. The proposed metrics show to provide a sensible way to evaluate the overall parallel complexity of a process model
Comparing data-driven meta-heuristics for the bi-objective Component Repairing Problem
No full textDue to both the increasing use of automation in production processes and the budget devoted for purchasing equipment, maintenance plays a key role in making a company competitive in the marketplace. Moreover, the use of data analysis techniques and the advent of Internet of Things make the IoT-based predictive maintenance possible. In addition, since all the resources (e.g., budget and human) involved in the maintenance activities are usually limited, a company is also interested in defining optimized maintenance plans. In this paper, the integration of IoT-based predictive maintenance with optimization techniques is investigated by developing a data-driven Greedy Randomized Adaptive Search Procedure (GRASP) meta-heuristic aimed at efficiently defining maintenance plans. In particular, we address the bi-objective component repairing problem (b-CRP), aimed at determining the set of components of a production system to repair that are more likely to fail. Having the breakage probability of each component, derived from historical data, the system reliability is maximized and the maximum time required to repair one component among those selected in the solution is minimized, under constraints on both budget and time for performing the maintenance activities. Then, we compare the solutions of GRASP with those of an already existing bi-objective Large Neighborhood Search meta-heuristic
Process-aware IIoT Knowledge Graph: A semantic model for Industrial IoT integration and analytics
Get PDFThe integration of the huge data streams produced by the Industrial Internet of Things (IIoT) can provide invaluable knowledge in the context of Industry 4.0, and is also an open research issue. The present paper proposes a semantic approach to this issue, centered around the notion of process as the backbone. We build an ontology describing the fundamental elements involved in IIoT and their relations, and discuss the construction of the Process-aware IIoT Knowledge Graph, where raw sensor data are enriched with information about process activities and the physical production environment. We also propose a framework for querying the Knowledge Graph, and we demonstrate its capabilities by considering the production of metal accessories as case study
A Negation Handling Technique for Sentiment Analysis
No full textTraditional lexicon-based approaches for sentiment analysis are usually not able to model negation, as they do not provide proper techniques to identify the right negation window. In this work we address the problem of the automatic determination of the scope of negation and we present a negation handling algorithm based on dependency-based parse trees. The proposal is based on the use of grammatical relations among words to model a sentence, and hence to determine words that are affected by negation. The proposed algorithm has been coupled with a semantic disambiguation technique to identify the sentiment of a sentence. Experiments on different datasets have proven that our proposal improves the accuracy of the sentiment analysis. The proposed algorithm has been implemented as part of a Social Information Discovery system, which allows for an integrated near-real-time analysis of discussions from multiple social networks
- …
