Discussion summarization using natural language processing techniquesStropnik, Vid (Avtor) Bosnić, Zoran (Mentor) Osipov, Evgeny (Komentor) bstractive summarizationOnline discussion summarizationLatent Dirchlet AllocationWord embeddingsHyperdimensional computingDimensionality reductionTopic LabelingSentiment AnalysisDiscussions held on online forums differ from traditional text documents in several ways. In addition to individual utterances usually being very short, they also have multiple messengers, each of whom may exhibit their own form of non-natural punctuation and undocumented internet lingo use. Consequently, the current state-ofthe-art methods for summarizing text and providing a clear, coherent picture of the topics discussed in a comments section cannot be easily applied to these sorts of corpora. This thesis discusses the techniques that can. In this work, we analyse the field of online discussion summarization. We pay the most attention to the topic modeling step of the current state-of-the-art method for this task, providing detailed theoretical explanations for each technique used therein. We externally examine three topical-clustering methods, concluding that Latent Dirichlet Allocation, word embeddings and dimensionally-reduced hyperdimensional computing can be considered comparable for this use-case. Additionally, a novel abstractive summarizer framework is proposed and compared to the current state-ofthe-art output, shedding a light on the potential direction of future work in the field. The experimental results show that the distinct systems used in our summarizer synergize well to produce legible and coherent conversation abstractions.20202020-09-09 09:10:00Diplomsko delo119504sl