Although the machine translation (MT) quality has been significantly improved, it still fails to meet the practical use requirements, especially in the narrow fields of expertise and under-resource languages. To solve this problem, most studies have been focusing on improving algorithms, translation models and corpora. However, very few studies could address a very important aspect that greatly affects the translation quality, which is semantic-oriented translation. In this article, we propose a solution of building a context-based semantic-oriented MT system by improving the neural network translation model in combination with a big semantic-enriched corpus. The neural network translation approach allows understanding the semantics of the whole sentence based on context vector and phrase translation memory. Moreover, automatic translation results are pre-processed by enriching well-defined meanings to entities for creating the final translated text showing to users. This solution has been used to build an English-Vietnamese semantic-oriented machine translation system dedicated in the tourism field. The result shows that this solution gives good translations that are very helpful and useful to users.