This three-year thesis aims to analyze the functioning of Sentiment Analysis tools, one of the main research areas in the field of Natural Language Processing (NLP). The work develops starting from an overview of the components and main objectives of NLP, examining the reasons for its rapid diffusion in contemporary society and addressing the main ethical issues related to its use. Next, the focus is on Sentiment Analysis, a field constantly expanding thanks to the enormous amount of opinions generated daily on web platforms such as blogs and social media. In a context where the consultation of structured and summarized information becomes increasingly valuable, users' trust in analysis models is closely linked to their transparency and interpretability. To demonstrate the importance of these aspects, the thesis analyzes a dataset of film reviews, implementing a neural network that carries out a binary classification of sentiment. Downstream of the predictive process, two advanced analyzes are developed using the SHAP (SHapley Additive exPlanations) and DiCE (Diverse Counterfactual Explanations) frameworks, in order to understand the decision-making process of the created model. The analyzes allow us to identify, with the help of dedicated metrics, the most influential terms for the model's predictions, highlighting the preferences, biases and possible weaknesses of the approach adopted and suggesting potential applications in different contexts.
Il presente elaborato di tesi triennale si propone di analizzare il funzionamento degli strumenti di Sentiment Analysis, una delle principali aree di ricerca nell’ambito dell’Elaborazione del Linguaggio Naturale (Natural Language Processing, NLP). Il lavoro si sviluppa partendo da una panoramica dei componenti e degli obiettivi principali dell’NLP, esaminando le ragioni della sua rapida diffusione nella società contemporanea e affrontando le principali questioni etiche connesse al suo utilizzo. Successivamente, l’attenzione si concentra sulla Sentiment Analysis, un campo in continua espansione grazie all’enorme quantità di opinioni generate quotidianamente su piattaforme Web come blog e social media. In un contesto dove la consultazione di informazioni strutturate e riassuntive diventa sempre più preziosa, la fiducia degli utenti nei confronti dei modelli di analisi risulta strettamente legata alla loro trasparenza e interpretabilità. Per dimostrare l’importanza di questi aspetti, l’elaborato analizza un dataset di recensioni cinematografiche, implementando una rete neurale che svolga una classificazione binaria del sentiment. A valle del processo predittivo, vengono sviluppate due analisi avanzate mediante i framework SHAP (SHapley Additive exPlanations) e DiCE (Diverse Counterfactual Explanations), al fine di comprendere il processo decisionale del modello realizzato. Le analisi permettono di individuare, con l’ausilio di metriche dedicate, i termini più influenti per le predizioni del modello, evidenziando le preferenze, i bias e le possibili debolezze dell’approccio adottato e suggerendone potenziali applicazioni in contesti diversi.
DETTAGLIARE LA SENTIMENT ANALYSIS: IDENTIFICAZIONE DI PAROLE CHIAVE MEDIANTE GENERAZIONE DI CONTROFATTUALI
de CANDIA, MARCANTONIO
2023/2024
Abstract
This three-year thesis aims to analyze the functioning of Sentiment Analysis tools, one of the main research areas in the field of Natural Language Processing (NLP). The work develops starting from an overview of the components and main objectives of NLP, examining the reasons for its rapid diffusion in contemporary society and addressing the main ethical issues related to its use. Next, the focus is on Sentiment Analysis, a field constantly expanding thanks to the enormous amount of opinions generated daily on web platforms such as blogs and social media. In a context where the consultation of structured and summarized information becomes increasingly valuable, users' trust in analysis models is closely linked to their transparency and interpretability. To demonstrate the importance of these aspects, the thesis analyzes a dataset of film reviews, implementing a neural network that carries out a binary classification of sentiment. Downstream of the predictive process, two advanced analyzes are developed using the SHAP (SHapley Additive exPlanations) and DiCE (Diverse Counterfactual Explanations) frameworks, in order to understand the decision-making process of the created model. The analyzes allow us to identify, with the help of dedicated metrics, the most influential terms for the model's predictions, highlighting the preferences, biases and possible weaknesses of the approach adopted and suggesting potential applications in different contexts.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi definitiva esse3 de Candia Marcantonio.pdf
File consultabile da postazioni biblioteca
Descrizione: Elaborato Tesi in formato PDF
Dimensione
2.03 MB
Formato
Adobe PDF
|
2.03 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14253/66900