Neural network-based semantic search of educational programmes fitting labor market requirements

With the growth of open educational content, growing demand for professional education from the labor market, and the development of the concept of lifelong learning, the task of updating the content of educational programs today is extremely important. The article discusses the semantic search method to retrieval and ranking of educational content for the specified requirements of the labor market, determined by professional standards. In contrast to traditional approaches of matching and analyzing the content of educational programs based on ontological models and rules, we propose the usage of word embedding and well-known neural network language models word2vec and fastText. The initial requests are specific requirements for knowledge, skills and descriptions of labor activities and labor functions extracted from professional standards. The search results are the descriptions of academic disciplines and online courses, including goals and objectives, learning outcomes, the structure and content of the main topics. We include the results of the expert evaluation of the ranking quality for the semantic search by metrics NDCG (Normalized Discounted Cumulative Gain) and MAP (Mean Average Precision) on the representative corpus of IT disciplines programmes of universities and massive open online courses (MOOC). The best results for the search are shown by the word2vec and fastText models, which are trained without supervision on large specially prepared corpuses of curriculum programs and descriptions of online courses. To move from word vectors to document vectors various combinations of neural network models with the TF-IDF weighting scheme are investigated.