The aim of this work is to get the best semantic representation of words, using sentence completion as a benchmark for its evaluation. We develop a semantic language model integrating a tensor to represent word co-occurrence statistics, an algorithm to transform dependency and skipgram word-word relations into a sentence relevance vector and we finally apply machine learning classification to improve the accuracy of the model. This model clearly highlights the difference of performances between statistical semantic models of the litterature, including those using hyperparameters inspired by neural network models. It shows the complementarity of dependency and linear context relations and gives a precise measurement of their respective contribution. A fully explicit model yields a performance of 77% of correct word in the sentence, enhanced up to 78.6% by machine learning classification, to be compared with a chance level of 20% (one out of five). Moreover we observe the presence of a bias in a well known sentence completion test set, and we propose a new set of 6000 sentences to the research community for further studies.