diplomska naloga
Andraž Pelicon (Author), Petra Kralj Novak (Mentor), Biljana Mileva-Boshkoska (Co-mentor)

Abstract

Diplomska naloga se ukvarja z analizo sentimenta v novicah. To področje v zadnjem času pridobiva na priljubljenosti, predvsem v okviru napovedovanja gibanja finančnih trgov, vendar je za slovenski jezik še dokaj slabo raziskano. Za slovenščino sicer obstajajo modeli, osnovani na metodi podpornih vektorjev, vendar ti niso dostopni za javno uporabo. V okviru te raziskave smo zasnovali arhitekturo na osnovi nevronskih mrež, ki za klasifikacijo uporablja kombinacijo samodejno generiranih značilk in TF-IDF obtežitev. Modeli, ki uporabljajo omenjeno arhitekturo, dosegajo primerljive rezultate z že obstoječimi modeli in so sposobni učinkovitega učenja na korpusih v velikosti okrog 10.000 dokumentov. Najuspešnejši model iz raziskave je na voljo kot spletna storitev na naslovu classify.ijs.si.

Keywords

analiza sentimenta;novice;slovenščina;nevronske mreže;globoko učenje;

Data

Language: Slovenian
Year of publishing:
Typology: 2.11 - Undergraduate Thesis
Organization: FIŠ - Faculty of Information Studies
Publisher: [A. Pelicon]
UDC: 004.032.26(043.2)
COBISS: 2048611347 Link will open in a new window
Views: 356
Downloads: 16
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary abstract: The present thesis deals with the sentiment analysis of news. This field has recently gained in popularity, especially as a supporting method for stock market prediction, but not much research has yet been done on the news in the Slovenian language. Models based on support vector machines do exist but are not available for public use. We developed a neural network architecture that leverages both automatically generated features and TF-IDF weights for classification of Slovenian news. Models based on this architecture achieve comparable results with existing models and can be successfully trained on datasets of approximately 10,000 documents. Our best performing model is available for public use in the form of a web service on the URL classify.ijs.si.
Secondary keywords: sentiment analysis;news;Slovene;neural networks;deep learning;
Type (COBISS): Bachelor thesis/paper
Thesis comment: Fakulteta za informacijske študije v Novem mestu
Source comment: Na ov.: Diplomska naloga : visokošolskega strokovnega študijskega programa prve stopnje;
Pages: XI, 51 str.
ID: 11237998