magistrsko delo
Pavle Kerkez (Author), Damjan Škulj (Mentor), Gregor Petrič (Co-mentor)

Abstract

Naraščajoč pomen spletnega komuniciranja v sodobni družbi je poudaril potrebo po razumevanju in prepoznavanju samomorilnih misli v teh spletnih prostorih. Spletne skupnosti, še posebej tiste, osredotočene na duševno zdravje, pogosto vključujejo komunikacije, ki so tesno prepletene z izrazi samomorilnih misli. Medtem ko je zaznavanje teh izrazov pomembno za raziskave, je tudi ključnega pomena za proaktivno moderiranje in preventivne strategije na teh platformah. Tradicionalne metode strojnega učenja so pokazale obetajoče rezultate pri prepoznavanju samomorilnih nagnjenj v besedilnih podatkih. Vendar pa pojav velikih jezikovnih modelov (LLM), kot je GPT-4, zasnovanih na sofisticiranih arhitekturah globokega učenja, ponuja potencial za globlje in bolj niansirano zaznavanje subtilnih namigov, povezanih s samomorilnimi mislimi, ki so pogosto prepleteni z drugimi temami in jih je težko ločiti. Osrednja tema te raziskave je preučiti sposobnost LLM pri zaznavanju samomorilne vsebine v spletnih vsebinah. Cilji vključujejo 1) vdelavo besedil in njihovo združevanje na podlagi podobnosti vsebine ter 2) fino nastavitev modelov za razlikovanje in kategorizacijo dokumentov glede na prisotnost pristnih samomorilnih misli nasproti splošnim razpravam o duševnem zdravju. Rezultati potrjujejo učinkovitost LLM v obeh nalogah, saj uspešno združujejo objave na podlagi njihove podobnosti vsebine, da ustvarijo oznake razredov, poleg tega pa imajo visoko natančnost in obnovitev pri razlikovanju samomorilnih misli od splošnih pripovedi o duševnem zdravju.

Keywords

suicide detection;machine learning;large language models;document embedding;clustering;

Data

Language: English
Year of publishing:
Typology: 2.09 - Master's Thesis
Organization: UL FDV - Faculty of Social Sciences
Publisher: [P. Kerkez]
UDC: 316.624(043.2)
COBISS: 167950083 Link will open in a new window
Views: 32
Downloads: 6
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: Slovenian
Secondary title: Suicidal ideation detection in online posts using large language models
Secondary abstract: The increasing significance of online communication in contemporary society has underlined the need to understand and identify suicidal ideation within these online spaces. Online communities, especially those centered on mental health, frequently feature communications deeply interwoven with expressions of suicidal ideation. While detecting these expressions is important for research, it is also fundamental for proactive moderation and prevention strategies within these platforms. Traditional machine learning methodologies have shown promise in recognizing suicidal tendencies in textual data. However, the emergence of large language models (LLM’s) like GPT-4, built on sophisticated deep learning architectures, offers potential for a deeper and more nuanced detection of subtle cues linked with suicidal ideation that are often mingled with other themes and difficult to isolate. The core focus of this research is to examine the capability of LLM's in detecting suicidal content in online content. The objectives include 1) embedding the texts and clustering them based on content similarity, and 2) fine-tuning the models to distinguish and categorize documents based on the presence of genuine suicidal ideation versus general mental health discussions. The results validate the efficacy of LLMs in both tasks, achieving successful clustering of posts based on their content similarities to generate class labels, as well as having high precision and recall in differentiating suicidal ideation from general mental health narratives.
Secondary keywords: zaznavanje samomora;strojno učenje;veliki jezikovni modeli;vložitve dokumentov;hierarhično združevanje;Socialna psihologija;Svetovni splet;Samomorilnost;Univerzitetna in visokošolska dela;
Type (COBISS): Master's thesis/paper
Study programme: 0
Embargo end date (OpenAIRE): 1970-01-01
Thesis comment: Univ. v Ljubljani, Fak. za družbene vede
Pages: 66 str.
ID: 20034058