Search

Radio map estimation (RME) is the task of predicting radio power at all locations in a two-dimensional area and at all frequencies in a given band. This thesis explores four deep learning approaches to RME: dual path autoencoders, skip connection autoencoders, diffusion, and joint learning with transmitter localization.

Abstract: Machine transliteration is important to machine translation and cross-lingual information retrieval. Previous works show that machine transliteration can benefit from supplemental phonetic transcriptions and transliterations from other languages through a re-ranking framework. In this thesis, I propose to leverage supplemental information by jointly considering it with the spelling of the source word during the transliteration generation process. This approach is shown to be more accurate and faster than the re-ranking approach. In addition, I propose to represent supplemental transliterations through a common phonetic interlingua. Experiments suggest that the interlingua representations can be as effective as the original orthography, and can even be obtained from languages that are not available during training.