Otkrivanje prisutnosti stresa u objavama na Redditu pomoću obrade prirodnog jezika

undergraduate thesis

• Title (english): Detecting Presence of Stress in Reddit Posts Using Natural Language Processing
• Author: Luka Pavlović
• Mentor: Jan Šnajder (mentor)
• Committee member: Jan Šnajder (predsjednik povjerenstva)
• Committee member: Ivan Crnomarković (član povjerenstva)
• Committee member: Zoran Medić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2021-07-05, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Cilj ovog rada bio je razvoj modela strojnog učenja koji će uspješno otkrivati prisutnost
stresa u objavama na Redditu. Skup podataka korišten u ovom radu stvorili su istraživači sa
Sveučilišta Columbia. Skup podataka stvoren je web-scrapingom društvene mreže Reddit.
Podaci su podijeljeni po domenama koje sadrže jedan subreddit ili više njih. Iako skup
podataka sadrži veći broj značajki, samo tri su korištene u ovom radu, a to su: subreddit, tekst
objave i oznaka koja označava prisutnost stresa. U pokušaju stvaranja modela koji će imati
zadovoljavajuće performanse izgrađena su četiri modela. Kao osnovni modeli stvoreni su:
klasifikator većinske klase te stroj potpornih vektora (engl. Support Vector Machine, SVM),
koji je koristio Term Frequency-Inverse Document Frequency vektore. Također, stvorene
su dvije verzije modela dubokog učenja: Long Short-Term Memory (LSTM) s globalnim
vektorima za reprezentaciju riječi (engl. Global Vectors for Word Representation, GloVe) i
LSTM s vektorizatorom FastText. Model SVM i model LSTM s GloVe optimizirani su na
najbolje hiperparametre te su evaluirani na skupu za testiranje. Unatoč visokim očekivanjima
od modela dubokog učenja, najbolje rezultate dao je model SVM.

Abstract (english)

The aim of this thesis was to develop a machine learning model that will successfully
detect the presence of stress in Reddit posts. The dataset used in this thesis was created by
researchers from Columbia University. The dataset was created by scraping of social network
Reddit. Data is divided into domains that contain one or more subreddits. Although
the dataset contains a number of features, only three were used for purposes of this thesis,
namely: subreddit, users’ post text, and a label indicating the presence of stress. Four
models were built in an attempt to create a model that would have satisfactory performance.
As baseline models, the following have been created: majority class baseline and Support
Vector Machine (SVM) which uses TF-IDF vectors. Also, two versions of the deep learning
model were created: Long Short-Term Memory (LSTM) with GloVe vectors and LSTM with
FastText vectorizer. SVM model and LSTM with GloVe were optimized for the best hyperparameters
and evaluated on the test set. Despite high expectations from the deep learning
model, the best results were obtained using the SVM model.

• Keywords: obrada prirodnog jezika; strojno učenje; Long Short-Term Memory; binarna klasifikacija; stroj potpornih vektora; stres; anksioznost; FastText
• Keywords (english): natural language processing; machine learning; Long Short-Term Memory; binary classification; Support Vector Machine; stress; anxiety; FastText
• Language: english
• URN:NBN: urn:nbn:hr:168:720918
• Study programme: Title: Computing; Study programme type: university; Study level: undergraduate; Academic / professional title: sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva (sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-14 21:23:40