Mineração de opiniões comparativas em português

• Autor(a) principal: Daniel Pimentel Kansaon
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFMG
• Texto Completo: http://hdl.handle.net/1843/52746
• topic: Mineração de opinião; Análise de sentimentos; Processamento de linguagem natural; Opinião comparativa; Detecção de preferência; Computação – Teses; Mineração de opinião – Teses; Análise de sentimentos – Teses; Processamento de linguagem natural (Computacao) – Teses

The constant expansion of e-commerce, recently boosted due to the coronavirus pandemic, has led to a huge increase in online shopping, made by increasingly demanding customers, who seek comments and reviews on the Web to assist in decision-making regarding the purchase of products. In these reviews, part of the opinions found are comparisons, which contrast aspects expressing a preference for an object over others, allowing, for example, companies to know how customers compare their products to their competitors. However, this information is neglected by traditional sentiment analysis techniques, and it is not applicable for comparisons, since they do not directly express a positive or negative sentiment. In this context, despite efforts in the English language, almost no studies have been done to develop appropriate solutions that allow the analysis of comparisons in the Portuguese language. This work presents one of the first studies on comparative opinion in Portuguese. In general, this work contains two main contributions. First, a hierarchical approach for detecting comparisons was proposed, which consists of an initial binary step, which subdivides the regular opinions of the comparatives, to further categorize the comparatives into the five groups of opinions: (1) Non-Comparative; (2) Non-Equal Gradable; (3) Equative, (4) Superlative; and (5) Non-Gradable. The results obtained are promising, reaching 87% of Macro-F1 and 0.94 of AUC for the binary step, and 61% of Macro-F1 for classification in multiple classes. Finally, in the second contribution, an algorithm was proposed to detect the entity expressed as preferred in comparative sentences, reaching 94% of Macro-F1 for Superlative and almost 84% for Non-Equal Gradable opinions.