IDENTIFYING SEMANTICALLY SIMILAR ARABIC WORDS USING A LARGE VOCABULARY SPEECH RECOGNITION SYSTEM

摘要

Users search digital libraries for book references using one or more attributes such as keywords, subject and author name. Some book titles might contain the keyword that the user specified and thus these titles will directly qualify as candidate results. On the other hand there are other titles that are relevant but do not contain the same exact search keyword. A user expects to retrieve all titles that are relevant to a specified keyword. Similarly when searching for an author name, the system should be able to retrieve the different forms of the name. The library science community developed a mechanism called authority control that allows the user to do a comprehensive search and retrieve all the records that are relevant to the query keyword. In this paper we propose an approach that allows the user to query an Arabic audio library using voice. We use a combination of class-based language models and robust interpretation to recognize and identify the spoken keywords. The mechanism uses a Large Vocabulary Recognition System (LVCSR) to implement the functionality of the authority control system. A series of experiments were performed to assess the accuracy and the robustness of the proposed approach: restricted grammar recognition with semantic interpretation, class-based statistical language models (CB-SLM) with robust interpretation, and generalized CB-SLM. The results have shown that the combination of CB-SLM and robust interpretation provides better accuracy and robustness than the traditional grammar-based parsing.