The energy and power industry has a large amount of text data, which has become an important research of big data mining. In order to improve the retrieval efficiency of article topic information, this paper proposes an unsupervised keyword extraction method based on text semantic graph, which focuses on improving the text graph construction and word weight calculation. According to the semantic dependencies of words in sentences, we build the text semantic graph which consist of four type edges, namely concept connection, equivalent membership, function attribute and modification limitation. Thus the parameter setting of the window length required in the traditional graph generation method is omitted. Then a word weight is calculated by combining keyword position information, term frequency-inverse document frequency, concept level and connection strength. Finally, we sort the importance of words and the high-score nodes are selected to form the keyword set of the abstract text. Experiment shows that our algorithm achieves better results in keyword extraction than traditional algorithms.