Sirvi Autor "Kuulmets, Hele-Andra, juhendaja" järgi

Nüüd näidatakse 1 - 2 2

Suhtlusvõimekuse arendamine sotsiaalsele humanoidrobotile SemuBot
(Tartu Ülikool, 2024) Unn, Albert; Kruusamäe, Karl, juhendaja; Luhtaru, Agnes, juhendaja; Kuulmets, Hele-Andra, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Social robots have been developed for decades, but creating the ability to have natural conversations with humans without strict rules has been a significant challenge. Approaches for communication between humans and robots have relied on pre-programmed dialogue options, which limits interaction and forces users to follow strictly defined rules. The rapid advancement of large language models offers a promising solution to this problem, enabling significant progress to be made in the quality of social robots. SemuBot is a student project to develop the first Estonian-speaking social humanoid robot, and this work focuses on exploring various solutions to achieve its ability to have conversations with people using a large language model. The study explores the use of three key components: speech recognition, large language models, and speech synthesis. The goal is to find optimal solutions for these components in the context of social robotics. As a result, a hardware and software solution for the social humanoid robot SemuBot is developed, enabling the robot to engage in natural conversations with people in Estonian.
Suurte keelemudelite võrdlev analüüs Eesti bioloogiaolümpiaadide küsimuste põhjal
(Tartu Ülikool, 2025) Kiil, Ahto; Purason, Taido, juhendaja; Kuulmets, Hele-Andra, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Several types of tests are used to evaluate large language models – translation, text comprehension, image recognition, answering questions etc. Typically, evaluation datasets are translated from English, and there is a lack of test sets that consider specific local context and are originally composed in Estonian. As part of this BA thesis, a multiple-choice dataset consisting of 1,031 questions was compiled using tasks from Estonian biology olympiads between 2005 and 2024. In the second phase, five OpenAI models, 13 Estonian-trained models from the Hugging Face platform and nine of the most recent closed commercial models accessed via websites were evaluated. The best model's accuracy (85.35%) is comparable to the average result (87.16%) of pupils who placed in the top three in Estonian olympiads.