Loading...
Publication
Implementação de um sistema de recomendação de restaurantes, para empresários, na área metropolitana de Lisboa
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 11.43 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
O presente trabalho tem como objetivo a implementação de um sistema de recomendação de restaurantes para empresários na Área Metropolitana de Lisboa(AML), utilizando Large Language Models (LLMs) para a análise de sentimentos contidos em comentários de restaurantes, provenientes da plataforma online DIG-IN (ex-Zomato). Para sustentar a abordagem metodológica, foi realizada uma revisão da literatura sobre a utilização de LLMs e a análise de sentimentos no setor da restauração. A revisão da literatura realizada centrou-se na exploração de diferentes abordagens de sistemas de recomendação e análise de sentimentos.
Foram analisados modelos tradicionais, como o DINESERV, e a sua adaptação para plataformas digitais, bem como técnicas mais recentes, como a Análise de Sentimento Baseada em Aspetos (ABSA). Esta revisão destaca também a crescente relevância dos LLMs como ferramentas robustas para interpretar grandes volumes de dados textuais, especialmente em setores como a restauração e o turismo. A metodologia adotada segue uma estrutura organizada em três fases principais. Na primeira fase, efetuou-se o pré-processamento dos dados, compilando e organizando conjuntos de dados (datasets) relativos às características
dos restaurantes na AML, como o tipo de cozinha, localização, número de comentários e avaliações. Na segunda fase, foram definidos os atributos e dimensões relevantes para a satisfação dos clientes, sendo aplicada uma extração de palavras-chave dos
comentários disponíveis na plataforma. Esta extração foi conduzida com o apoio dos LLMs, que permitiram identificar termos e expressões associados a aspetos críticos da experiência gastronómica. Na terceira fase, foram desenvolvidos e aplicados prompts aos modelos GPT-4 (versão de 2024-05-13), GPT-3.5 Turbo, e Mistral 7B Instruct para a classificação dos atributos contidos nos comentários. Estes modelos analisaram a polaridade dos sentimentos relacionados com aspetos como a qualidade da comida,
o serviço, o ambiente e outros fatores que influenciam a experiência dos clientes. A principal aplicação prática deste estudo foi a construção de um protótipo de uma ferramenta que permite a potenciais investidores avaliar o cenário competitivo na AML. Esta ferramenta não apenas analisa o potencial de implementação de determinado tipo de restaurante em diferentes localizações, mas também oferece uma visão detalhada dos concorrentes, com base na análise de sentimentos extraída dos comentários dos consumidores. Além disso, o estudo abordou a adaptação do modelo DINESERV para a análise automática de comentários em plataformas digitais, simplificando os seus 29 critérios originais em 10 novos critérios, mais adequados à natureza concisa das
avaliações online. A viabilidade da utilização de grandes modelos de linguagem (LLMs) para a extração, identificação e classificação de atributos relevantes na análise de sentimento foi demonstrada com sucesso, abrindo portas para futuras investigações que explorem o seu potencial em outras áreas da restauração e do turismo.
The present work aims to implement a restaurant recommendation system for entrepreneurs in the Lisbon Metropolitan Area (AML), using Large Language Models (LLMs) for sentiment analysis of restaurant reviews sourced from the online platform DIG-IN (formerly Zomato). To support the methodological approach, a literature review was conducted on the use of LLMs and sentiment analysis in the restaurant industry. The literature review focused on exploring different approaches to recommendation systems and sentiment analysis. Traditional models, such as DINESERV, and their adaptation to digital platforms were analyzed, as well as more recent techniques like Aspect-Based Sentiment Analysis (ABSA). This review also highlights the growing relevance of LLMs as robust tools for interpreting large volumes of textual data, especially in sectors such as restaurants and tourism. The methodology adopted follows a structure organized into three main phases. In the first phase, data preprocessing was carried out, compiling and organizing datasets related to restaurant characteristics in the AML, such as type of cuisine, location, number of reviews, and ratings. In the second phase, relevant attributes and dimensions for customer satisfaction were defined, and a keyword extraction from the reviews available on the platform was conducted. This extraction was supported by LLMs, which helped identify terms and expressions associated with critical aspects of the dining experience. In the third phase, prompts were developed and applied to the GPT-4 (2024-05-13 version), GPT-3.5 Turbo, and Mistral 7B Instruct models to classify the attributes found in the reviews. These models analyzed the sentiment polarity related to aspects such as food quality, service, ambiance, and other factors that influence the customer experience. The main practical application of this study was the construction of a prototype tool that allows potential investors to assess the competitive landscape in the AML. This tool not only analyzes the potential for implementing specific types of restaurants in different locations but also provides a detailed overview of competitors based on sentiment analysis extracted from consumer reviews. Additionally, the study addressed the adaptation of the DINESERV model for the automatic analysis of reviews on digital platforms, simplifying its 29 original criteria into 10 new criteria, more suited to the concise nature of online reviews. The viability of using Large Language Models (LLMs) to extract, identify, and classify relevant attributes in sentiment analysis was successfully demonstrated, opening the door for future research to explore their potential in other areas of the restaurant and tourism industries.
The present work aims to implement a restaurant recommendation system for entrepreneurs in the Lisbon Metropolitan Area (AML), using Large Language Models (LLMs) for sentiment analysis of restaurant reviews sourced from the online platform DIG-IN (formerly Zomato). To support the methodological approach, a literature review was conducted on the use of LLMs and sentiment analysis in the restaurant industry. The literature review focused on exploring different approaches to recommendation systems and sentiment analysis. Traditional models, such as DINESERV, and their adaptation to digital platforms were analyzed, as well as more recent techniques like Aspect-Based Sentiment Analysis (ABSA). This review also highlights the growing relevance of LLMs as robust tools for interpreting large volumes of textual data, especially in sectors such as restaurants and tourism. The methodology adopted follows a structure organized into three main phases. In the first phase, data preprocessing was carried out, compiling and organizing datasets related to restaurant characteristics in the AML, such as type of cuisine, location, number of reviews, and ratings. In the second phase, relevant attributes and dimensions for customer satisfaction were defined, and a keyword extraction from the reviews available on the platform was conducted. This extraction was supported by LLMs, which helped identify terms and expressions associated with critical aspects of the dining experience. In the third phase, prompts were developed and applied to the GPT-4 (2024-05-13 version), GPT-3.5 Turbo, and Mistral 7B Instruct models to classify the attributes found in the reviews. These models analyzed the sentiment polarity related to aspects such as food quality, service, ambiance, and other factors that influence the customer experience. The main practical application of this study was the construction of a prototype tool that allows potential investors to assess the competitive landscape in the AML. This tool not only analyzes the potential for implementing specific types of restaurants in different locations but also provides a detailed overview of competitors based on sentiment analysis extracted from consumer reviews. Additionally, the study addressed the adaptation of the DINESERV model for the automatic analysis of reviews on digital platforms, simplifying its 29 original criteria into 10 new criteria, more suited to the concise nature of online reviews. The viability of using Large Language Models (LLMs) to extract, identify, and classify relevant attributes in sentiment analysis was successfully demonstrated, opening the door for future research to explore their potential in other areas of the restaurant and tourism industries.
Description
Keywords
Grandes modelos de linguagem Avaliações online Restauração Engenharia de prompts