Amharic is the official language of Ethiopia, currently having a population of over 118 million. Developing effective information retrieval (IR) system for Amharic has been a challenging task due to limited resources coupled with complex morphology of the language. This paper presents the development of Amharic semantic IR system using query expansion based on deep neural learning model and WordNet. In order to optimize the retrieval result, we propose Amharic text representation using root forms of words applied for stopword identification, indexing, term matching and query expansion. Comparisons are made with the conventional stem-based text representation for information retrieval, and we show that using the root forms of words is better for both resource construction and system development. The effectiveness of the proposed Amharic semantic IR system is evaluated on Amharic Adhoc Information Retrieval Test Collection (2AIRTC).