Loading...
Publication
Análise semântica das avaliações do TripAdvisor: um sistema de monitorização e apoio à decisão de restaurantes no Algarve
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 2.46 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
O presente trabalho tem como objetivo investigar a aplicação de métodos de Análise Semântica e Processamento de Linguagem Natural (PLN) para analisar os comentários e sentimentos expressos em avaliações de restaurantes na região do Algarve, disponíveis numa plataforma online (TripAdvisor). O estudo abordou esta tarefa através de três propostas metodológicas fundamentadas em técnicas de Análise Semântica e PLN: - A primeira abordagem utilizou modelos de embeddings pré-treinados, especificamente BERT (Bidirectional Encoder Representations from Transformers), USE (Universal Sentence Encoder) e OpenAI, para extrair cinco atributos fundamentais da experiência gastronómica e os sentimentos associados a cada avaliação. Os resultados mostraram métricas F1 promissoras, especialmente para o modelo OpenAI. No entanto, este último apresenta custos associados, tornando o modelo USE uma opção mais equilibrada em termos de custo-benefício. - A segunda abordagem focou-se na extração de onze subatributos, provenientes dos cinco atributos principais anteriormente identificados. Foram explorados modelos alternativos de Inferência em Linguagem Natural (Natural Language Inference - NLI), nomeadamente BART (Bidirectional and Auto-Regressive Transformer) e DeBERTa (Decoding-enhanced BERT with Disentangled Attention). Os resultados indicam uma superioridade do modelo BART em termos de desempenho. - A terceira e última abordagem envolveu o uso do modelo ChatGPT para a extração de subatributos e seus respetivos sentimentos. Este modelo mostrou um nível de concordância próximo ao humano, mas apresentou limitações relacionadas ao custo computacional, especialmente quando aplicado a um grande volume de dados. Como aplicação prática deste estudo, foi desenvolvido um protótipo de plataforma web designado RestMon, que utiliza o modelo BART para analisar mais de 700 mil avaliações do TripAdvisor sobre restaurantes no Algarve. A plataforma oferece análises detalhadas sobre onze subatributos cruciais da experiência gastronómica, incluindo representações visuais e painéis comparativos. Em suma, este trabalho permite avaliar a viabilidade da aplicação de técnicas avançadas de PLN na análise de avaliações online de restaurantes, mas também oferece contribuições práticas que têm o potencial de impactar significativamente a gestão estratégica e operacional em estabelecimentos de restauração.
The present work aims to investigate the application of methods of Semantic Analysis and Natural Language Processing (NLP) for analysing comments and sentiments expressed in restaurant reviews in the Algarve region, available on an online platform (TripAdvisor). The study approached this task through three methodological proposals grounded in techniques of Semantic Analysis and NLP: - The first approach used pre-trained embedding models, specifically BERT (Bidirectional Encoder Representations from Transformers), USE (Universal Sentence Encoder), and OpenAI, to extract five fundamental attributes of the gastronomic experience and the sentiments associated with each review. The results showed promising F1 metrics, especially for the OpenAI model. However, the latter has associated costs, making the USE model a more balanced option in terms of cost-benefit. - The second approach focused on the extraction of eleven sub-attributes, derived from the five main attributes previously identified. Alternative models of Natural Language Inference (NLI) were explored, namely BART (Bidirectional and Auto-Regressive Transformer) and DeBERTa (Decoding-enhanced BERT with Disentangled Attention). The results indicate a superiority of the BART model in terms of performance. - The third and final approach involved the use of the ChatGPT model for the extraction of sub-attributes and their respective sentiments. This model showed a level of agreement close to human, but presented limitations related to computational cost, especially when applied to a large volume of data. As a practical application of this study, a web platform prototype called RestMon was developed, which uses the BART model to analyse more than 700,000 TripAdvisor reviews about restaurants in the Algarve. The platform offers detailed analyses on eleven crucial sub-attributes of the gastronomic experience, including visual representations and comparative panels. In summary, this work allows for assessing the viability of applying advanced NLP techniques in the analysis of online restaurant reviews, but also offers practical contributions that have the potential to significantly impact the strategic and operational management in catering establishments.
The present work aims to investigate the application of methods of Semantic Analysis and Natural Language Processing (NLP) for analysing comments and sentiments expressed in restaurant reviews in the Algarve region, available on an online platform (TripAdvisor). The study approached this task through three methodological proposals grounded in techniques of Semantic Analysis and NLP: - The first approach used pre-trained embedding models, specifically BERT (Bidirectional Encoder Representations from Transformers), USE (Universal Sentence Encoder), and OpenAI, to extract five fundamental attributes of the gastronomic experience and the sentiments associated with each review. The results showed promising F1 metrics, especially for the OpenAI model. However, the latter has associated costs, making the USE model a more balanced option in terms of cost-benefit. - The second approach focused on the extraction of eleven sub-attributes, derived from the five main attributes previously identified. Alternative models of Natural Language Inference (NLI) were explored, namely BART (Bidirectional and Auto-Regressive Transformer) and DeBERTa (Decoding-enhanced BERT with Disentangled Attention). The results indicate a superiority of the BART model in terms of performance. - The third and final approach involved the use of the ChatGPT model for the extraction of sub-attributes and their respective sentiments. This model showed a level of agreement close to human, but presented limitations related to computational cost, especially when applied to a large volume of data. As a practical application of this study, a web platform prototype called RestMon was developed, which uses the BART model to analyse more than 700,000 TripAdvisor reviews about restaurants in the Algarve. The platform offers detailed analyses on eleven crucial sub-attributes of the gastronomic experience, including visual representations and comparative panels. In summary, this work allows for assessing the viability of applying advanced NLP techniques in the analysis of online restaurant reviews, but also offers practical contributions that have the potential to significantly impact the strategic and operational management in catering establishments.
Description
Keywords
Processamento de linguagem natural Avaliações online Gastronomia e turismo Análise semântica