AbstractsBusiness Management & Administration

Before buying a product or service, consumers often search for expert reviews, but increasingly also for opinions of other consumers. Many useful opinions are expressed in plain texts (e.g., in discussion fora or social media). It is then desirable to extract aspects (e.g., screen, battery) from the texts that discuss a particular target entity (e.g., a smartphone), i.e., figure out what is being discussed about the target entity, and also estimate aspect scores, i.e., how positive or negative the average sentiment for each aspect is. These two goals are jointly known as Aspect Based Sentiment Analysis (ABSA). Several ABSA systems have been developed, but there is no established framework to evaluate them and no consensus task decomposition of ABSA. This thesis proposes a decomposition of ABSA into three subtasks: aspect term extraction, aspect aggregation, and aspect sentiment estimation. Aspect term extraction identifies single- and multi-word terms that name aspects of the target entity (e.g., "battery", "hard disk"). Aspect aggregation groups similar aspect terms (e.g., "price" and "cost", but possibly also "design" and "color") into fewer or more clusters, depending on user preferences and other constraints (e.g., display size, when ABSA results are to be reported on mobile devices). Aspect sentiment estimation determines the average sentiment for each aspect term or cluster of aspect terms of the target entity. For each one of these subtasks, appropriate benchmark datasets from multiple domains (e.g., restaurant and laptop reviews) were constructed during the work of this thesis, which were also adopted (with some changes) in the SemEval 2014 and 2015 ABSA competitions. The thesis also proposes new evaluation measures for each subtask, arguing that they are more appropriate than previous, more commonly used measures. New methods or modifications of previous methods are also proposed for each subtask, and they are evaluated using the benchmark datasets and proposed evaluation measures, with experimental results showing that they are competitive to state of the art methods. Τα συστήματα Ανάλυσης Συναισθήματος Βασισμένης σε Χαρακτηριστικά (Aspect Based Sentiment Analysis, ABSA) δέχονται ως είσοδο ένα σύνολο κειμένων (π.χ. κριτικές προϊόντων ή μηνύματα κοινωνικών δικτύων) που σχολιάζουν μια συγκεκριμένη οντότητα (π.χ. ένα συγκεκριμένο μοντέλου κινητού τηλεφώνου). Επιχειρούν να εντοπίσουν τα κυριότερα (π.χ. τα πιο συχνά σχολιαζόμενα) χαρακτηριστικά (aspects) της οντότητας (π.χ. μπαταρία, οθόνη), καθώς και να εκτιμήσουν το μέσο συναίσθημα (π.χ. πόσο θετικό ή αρνητικό είναι) που εκφράζουν τα κείμενα για κάθε χαρακτηριστικό της οντότητας. Αν και έχουν αναπτυχθεί αρκετά συστήματα αυτού του είδους, δεν υπάρχει καθιερωμένος τρόπος αποδόμησης και αξιολόγησης των εργασιών (tasks) που απαιτούν.Σε αυτή τη διατριβή προτείνεται μια αποδόμηση των εργασιών των συστημάτων ABSA που περιλαμβάνει τρεις υπο-εργασίες: εξαγωγή χαρακτηριστικών όρων (aspect term extraction), ομαδοποίηση χαρακτηριστικών (aspect aggregation) και εκτίμηση…