Chatbots and virtual assistants are becoming more and more present in our lives. They allow users to communicate in spoken or written natural language, through various communication channels. Slovenian language is poorly supported by globally used smart assistants, due to the small number of people that use it for communication.
We have developed a Slovenian virtual assistant for smart home management. The assistant understands natural language and uses a language model to classify the purpose and entities in the user's message by taking the context of the entire conversation into account. It uses a conversational model to determine the name of the action responsible for generating the response. We used actions to implement various skills. They enable users to obtain various information and perform different tasks.
In our thesis, we have developed and evaluated several different models for intent classification and entity extraction. The highest performance in the intent classification was achieved by using word embeddings from the SloBERTa language model (F1 score = 0,900). In the extraction of entities, the highest performance was achieved by using word embeddings from the fastText model (F1 score = 0.924).
|
