Loading...
Publication
Utilização de técnicas de Processamento de Linguagem Natural para construção de um sistema de recomendação baseado na similaridade de vetores de representação textual
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 4.89 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Num contexto global de produção massiva de informação, os dados textuais não estruturados — frequentemente desprezados ou subutilizados — apresentam-se como uma oportunidade para organizações que pretendem personalizar a experiência do utilizador sem, contudo, investir montantes elevados em infraestruturas de dados dispendiosas. Uma vez que os sistemas de recomendação tradicionais, especialmente aqueles baseados em filtragem colaborativa, exigem grandes volumes de dados comportamentais, tornando-se excessivamente onerosos para organizações de menor dimensão, este estudo propõe como alternativa acessível e igualmente eficaz o desenvolvimento de sistemas de recomendação inteiramente baseados em conteúdo textual. Para tal, aplicaram-se técnicas de processamento de linguagem natural (PLN) para analisar a similaridade semântica entre vetores de representação textual, com baixo recurso a capacidade computacional. Foram desenvolvidos dois protótipos — um baseado em Term Frequency - Inverse Document Frequency (abordagem de ponderação lexical) e outro em Sentence BERT (arquitetura de transformers e embeddings densos) — avaliando a sua capacidade de gerar recomendações em dois cenários: o catálogo de oitocentos e sete cursos da Escola Virtual de Governo (EVG) e uma coleção de quase trinta mil livros da Google Books API. Os resultados evidenciaram que ambos os modelos produzem sugestões relevantes, sendo o baseado em Sentence BERT aquele que se destacou pela precisão semântica na captura de nuances contextuais (por exemplo, a distinção de polissemias e relações temáticas implícitas), enquanto o Term Frequency - Inverse Document Frequency mostrou-se preciso na identificação de correspondências lexicais exatas entre palavras chave. A aplicação prática nestes domínios demonstrou a viabilidade de implementação, com recomendações
adaptáveis a diferentes necessidades. Este trabalho demonstra que técnicas modernas de processamento de linguagem natural podem democratizar o acesso a sistemas de recomendação e oferecer soluções eficientes às mais diversas organizações, sem comprometer a qualidade das sugestões.
In a global context of massive information production, unstructured textual data — often overlooked or underutilized — represents an opportunity for organizations seeking to personalize user experience without investing heavily in costly data infrastructures. Since traditional recommendation systems, particularly those based on collaborative filtering, require large volumes of behavioral data and are therefore prohibitively expensive for smaller organizations, this study proposes an accessible and equally effective alternative: the development of recommendation systems entirely based on textual content. To achieve this, Natural Language Processing (NLP) techniques were applied to analyze semantic similarity between textual representation vectors, with minimal computational resource requirements. Two prototypes were developed — one based on Term Frequency - Inverse Document Frequency (a lexical weighting approach), and another based on Sentence BERT (a transformers architecture using dense embeddings) — and their recommendation capabilities were evaluated in two scenarios: the catalog of eight hundred seven courses from the Escola Virtual de Governo (EVG), and a collection of almost thirty thousand books from the Google Books API. The results showed that both models generated relevant suggestions, with the Sentence BERT model standing out for its semantic precision in capturing contextual nuances (such as distinguishing polysemies and implicit thematic relationships), while the TF-IDF model proved accurate in identifying exact lexical matches between keywords. The practical application in these domains demonstrated the feasibility of implementation, with recommendations adaptable to different needs. This work demonstrates that modern NLP techniques can democratize access to recommendation systems, providing efficient solutions for a wide range of organizations without compromising suggestion quality. Keywords:
In a global context of massive information production, unstructured textual data — often overlooked or underutilized — represents an opportunity for organizations seeking to personalize user experience without investing heavily in costly data infrastructures. Since traditional recommendation systems, particularly those based on collaborative filtering, require large volumes of behavioral data and are therefore prohibitively expensive for smaller organizations, this study proposes an accessible and equally effective alternative: the development of recommendation systems entirely based on textual content. To achieve this, Natural Language Processing (NLP) techniques were applied to analyze semantic similarity between textual representation vectors, with minimal computational resource requirements. Two prototypes were developed — one based on Term Frequency - Inverse Document Frequency (a lexical weighting approach), and another based on Sentence BERT (a transformers architecture using dense embeddings) — and their recommendation capabilities were evaluated in two scenarios: the catalog of eight hundred seven courses from the Escola Virtual de Governo (EVG), and a collection of almost thirty thousand books from the Google Books API. The results showed that both models generated relevant suggestions, with the Sentence BERT model standing out for its semantic precision in capturing contextual nuances (such as distinguishing polysemies and implicit thematic relationships), while the TF-IDF model proved accurate in identifying exact lexical matches between keywords. The practical application in these domains demonstrated the feasibility of implementation, with recommendations adaptable to different needs. This work demonstrates that modern NLP techniques can democratize access to recommendation systems, providing efficient solutions for a wide range of organizations without compromising suggestion quality. Keywords:
Description
Keywords
Processamento de Linguagem Natural Sistemas de Recomendação Embeddings Similaridade Semântica Natural Language Processing Recommender Systems Semantic Similarity