ABSAuDA - Análise de Sentimentos baseado em Aspectos utilizando Análise de Dependência

Abstract

Aspect-Based Sentiment Analysis consists of examining people’s opinions, feel ings, evaluations, assessments, attitudes and emotions towards entities and their as pects, expressed in different formats, such as text, image, audio or video. Entities can be, for example, products and services, while aspects can be, for example, price and durability of a product or service. However, to perform Aspect-Based Sentiment Anal ysis it is necessary to use resources called Linguistic Resources, which are used to support research and applications, such as Datasets and Linguistic Knowledge Bases. However, in Low Resources Languages, such as Brazilian Portuguese, there is a lack of resources. Due to insufficient resources there is a shortage of reliable methods for Low Resource Languages. This work proposes an Aspect-Based Sentiment Anal ysis system, employing domain ontologies and syntactic dependency analysis. The HOntology ontology was used, responsible for storing concepts, relationships and in stances in the accommodation domain, to identify aspects related to the hotel system. Syntactic dependency analysis is used to identify opinionated terms related to as pects. Furthermore, the study incorporated the use of several sentiment lexicons (e.g., AffectPT-br, EmoLex, LeIA, LIWC2007pt, Onto.PT, OpLexicon, Reli-Lex, SentiLex-PT, SentiWordNet-PT-BR, UNILEX, WordNetAffectBR), which are responsible for storing the polarity of words, and the use of EnsembleLex, an ensemble of the results of the four best lexicons. The tests were carried out to verify the feasibility of using spelling correction (manual and through the autocorrect, pyspellchecker and LanguageTool li braries) and the use of the fashion of joining adjectives that were related to an aspect. The results demonstrated that EnsembleLex surpassed the results obtained by other lexicons, reaching a maximum of 0.461 for f-measure (macro) and 0.536 for f-measure (micro), using reviews with neutral annotations regarding the sentiment directed at as pects, the fashion of adjective union and the pyspellchecker spelling correction library.