The Hidden Markov Models (HMMs) have been shown to achieve good performance when applied to information extraction tasks. This paper describes the training aspect of exploring HMMs for the task of metadata extraction from tagged bibliographic references. This paper is the improvement of the technique proposed by earlier researchers for smoothing emission probabilities in order to avoid the occurrence of zero values. The results show the effectiveness of the proposed method. The main advantage of HMMs in language modeling is the fact that they are well suited for the modeling of sequential data, such as spoken or written language.