Most existing methodologies for text classification represent text as vectors of words, to be specific "bag-of-words." This content portrayal results in a high dimensionality of feature space and much of the time experiences surface jumbling. When it comes to short texts, these become even more serious because of their shortness and sparsity and with the bilingual similarity of text it gets more difficult. This paper proposes an approach to deal with both sparsity and computational complexity of bilingual similarity of short text. English short text is mapped with Probase and Hindi short text is mapped with BabelNet a knowledge base with coverage of words and concepts for 248 languages. A semantic network is created to manipulate the word to word and concept to concept correlation. Unlike the earlier approaches of embedding, words and concepts from both English and Hindi short texts are treated separately to yield word embedding (Word2Vec) and concept embedding (Concept2Vec) respectively. The similarity between bilingual short texts is computed using the skip-gram based word embedding and concept embedding. When evaluated with Pilot and STSS 131 short text benchmark datasets, the proposed optimized bilingual short text embedding gives better similarity score