As the number of scholarly articles we can access is increasing, it becomes possible to read them freely. However, it is difficult to understand scholarly articles since they are basically written for experts. Our big goal is, developing methods to extract essential elements of articles, to facilitate open innovation. To this end, this paper is devoted to considering automatic identification of dataset names in articles. Because a dictionary of datasets is necessary for evaluation, existing methods have focused on some specific discipline. To achieve applicability to any disciplines, we adopt a machine learning approach with a huge amount of scholarly papers. Because we treat papers in multi-disciplines, it is challenging how to evaluate experimental results. To solve it, we quantitatively evaluate experimental results with precision@N, which does not require to know all the dataset names in the papers we use, and qualitatively check if candidate tokens are dataset names or not using a GUI tool we have developed. While about 1/3 tokens of the top 20 output by our method were dataset names, the other ones are names of methods, models, or organizations. So it is important future work to remove such noise results, using additive compositionality of word vectors.