The topic of this thesis is statistical models for the analysis of textual data, emphasizing contexts in which text samples are grouped. When dealing with text data, the first issue is to process it, making it computationally and methodologically compatible with the existing mathematical and statistical methods produced and continually developed by the scientific community. Therefore, the thesis firstly reviews existing methods for analytically representing and processing textual datasets, including Vector Space Models, distributed representations of words and documents, and contextualized embeddings. It realizes this review by standardizing a notation that, even within the same representation approach, appears highly heterogeneous in the literature. Then, two domains of application are explored: social media and cultural tourism. About the former, a study is proposed about self-presentation among diverse groups of individuals on the StockTwits platform, where finance and stock markets are the dominant topics. The methodology proposed integrated various types of data, including textual and categorical data. This study revealed insights into how people present themselves online and found recurring patterns within groups of users. About the latter, the thesis delves into a study conducted as part of the "Data Science for Brescia - Arts and Cultural Places" Project, where a language model was trained to classify Italian-written online reviews into four distinct semantic areas related to cultural attractions in the Italian city of Brescia. The model proposed allows for the identification of attractions in text documents, even when not explicitly mentioned in document metadata, thus opening possibilities for expanding the database related to these cultural attractions with new sources, such as social media platforms, forums, and other online spaces. Lastly, the thesis presents a methodological study examining the group-specificity of words, analyzing various group-specificity estimators proposed in the literature. The study considered grouped text documents with both outcome and group variables. Its contribution consists of the proposal of modeling the corpus of documents as a multivariate distribution, enabling the simulation of corpora of text documents with predefined characteristics. The simulation provided valuable insights into the relationship between groups of documents and words. Furthermore, all its results can be freely explored through a web application, whose components are also described in this manuscript. In conclusion, this thesis has been conceived as a collection of papers. It aimed to contribute to the field with both applications and methodological proposals, and each study presented here suggests paths for future research to address the challenges in the analysis of grouped textual data.
L'argomento di questa tesi sono i modelli statistici per l'analisi dei dati testuali, con particolare attenzione ai contesti in cui i campioni di testo sono raggruppati. Quando si ha a che fare con dati testuali, il primo problema è quello di elaborarli, per renderli compatibili dal punto di vista computazionale e metodologico con i metodi matematici e statistici prodotti e continuamente sviluppati dalla comunità scientifica. Per questo motivo, la tesi passa in rassegna i metodi esistenti per la rappresentazione analitica e l'elaborazione di campioni di dati testuali, compresi i "Vector Space Models", le "rappresentazioni distribuite" di parole e documenti e i "contextualized embeddings". Questa rassegna comporta la standardizzazione di una notazione che, anche all'interno dello stesso approccio di rappresentazione, appare molto eterogenea in letteratura. Vengono poi esplorati due domini di applicazione: i social media e il turismo culturale. Per quanto riguarda il primo, viene proposto uno studio sull'autodescrizione di gruppi diversi di individui sulla piattaforma StockTwits, dove i mercati finanziari sono gli argomenti dominanti. La metodologia proposta ha integrato diversi tipi di dati, sia testuali che variabili categoriche. Questo studio ha agevolato la comprensione sul modo in cui le persone si presentano online e ha trovato stutture di comportamento ricorrenti all'interno di gruppi di utenti. Per quanto riguarda il turismo culturale, la tesi approfondisce uno studio condotto nell'ambito del progetto "Data Science for Brescia - Arts and Cultural Places", in cui è stato addestrato un modello linguistico per classificare le recensioni online scritte in italiano in quattro aree semantiche distinte relative alle attrazioni culturali della città di Brescia. Il modello proposto permette di identificare le attrazioni nei documenti di testo, anche quando non sono esplicitamente menzionate nei metadati del documento, aprendo così la possibilità di espandere il database relativo a queste attrazioni culturali con nuove fonti, come piattaforme di social media, forum e altri spazi online. Infine, la tesi presenta uno studio metodologico che esamina la specificità di gruppo delle parole, analizzando diversi stimatori di specificità di gruppo proposti in letteratura. Lo studio ha preso in considerazione documenti testuali raggruppati con variabile di "outcome" e variabile di gruppo. Il suo contributo consiste nella proposta di modellare il corpus di documenti come una distribuzione multivariata, consentendo la simulazione di corpora di documenti di testo con caratteristiche predefinite. La simulazione ha fornito preziose indicazioni sulla relazione tra gruppi di documenti e parole. Inoltre, tutti i risultati possono essere liberamente esplorati attraverso un'applicazione web, i cui componenti sono altresì descritti in questo manoscritto. In conclusione, questa tesi è stata concepita come una raccolta di studi, ognuno dei quali suggerisce percorsi di ricerca futuri per affrontare le sfide dell'analisi dei dati testuali raggruppati.
Statistical analysis of grouped text documents / Ricciardi, Riccardo. - (2024 Jan 23).
Statistical analysis of grouped text documents
Ricciardi, Riccardo
2024-01-23
Abstract
The topic of this thesis is statistical models for the analysis of textual data, emphasizing contexts in which text samples are grouped. When dealing with text data, the first issue is to process it, making it computationally and methodologically compatible with the existing mathematical and statistical methods produced and continually developed by the scientific community. Therefore, the thesis firstly reviews existing methods for analytically representing and processing textual datasets, including Vector Space Models, distributed representations of words and documents, and contextualized embeddings. It realizes this review by standardizing a notation that, even within the same representation approach, appears highly heterogeneous in the literature. Then, two domains of application are explored: social media and cultural tourism. About the former, a study is proposed about self-presentation among diverse groups of individuals on the StockTwits platform, where finance and stock markets are the dominant topics. The methodology proposed integrated various types of data, including textual and categorical data. This study revealed insights into how people present themselves online and found recurring patterns within groups of users. About the latter, the thesis delves into a study conducted as part of the "Data Science for Brescia - Arts and Cultural Places" Project, where a language model was trained to classify Italian-written online reviews into four distinct semantic areas related to cultural attractions in the Italian city of Brescia. The model proposed allows for the identification of attractions in text documents, even when not explicitly mentioned in document metadata, thus opening possibilities for expanding the database related to these cultural attractions with new sources, such as social media platforms, forums, and other online spaces. Lastly, the thesis presents a methodological study examining the group-specificity of words, analyzing various group-specificity estimators proposed in the literature. The study considered grouped text documents with both outcome and group variables. Its contribution consists of the proposal of modeling the corpus of documents as a multivariate distribution, enabling the simulation of corpora of text documents with predefined characteristics. The simulation provided valuable insights into the relationship between groups of documents and words. Furthermore, all its results can be freely explored through a web application, whose components are also described in this manuscript. In conclusion, this thesis has been conceived as a collection of papers. It aimed to contribute to the field with both applications and methodological proposals, and each study presented here suggests paths for future research to address the challenges in the analysis of grouped textual data.File | Dimensione | Formato | |
---|---|---|---|
Pdf_A PhD Thesis Riccardo Ricciardi 23_12_23.pdf
accesso aperto
Descrizione: Tesi
Tipologia:
Tesi di dottorato
Dimensione
3.03 MB
Formato
Adobe PDF
|
3.03 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.