I don't think it makes sense to select keyphrases based solely on word vectors. But one could inform/rerank the extractions from other methods.
Basic keyphrase extraction mostly works based on TF-IDF -- words/phrases that appear frequently in the document and not so frequently in other documents should be good descriptive keyphrases.
TF-IDF, however, has no idea about word similarity, e.g. could extract keyphrases like "Dijkstra algorithm", "Dijkstra method", "Dijkstra methodology" from a paper on graph algorithms. If you can provide only a short list of keyphrases, the list might be clogged with these redundant keyphrases and you miss some important ones.
Word Embeddings could inform your ranking on whether they are similar or not: E.g>if a new keyphrase's embedding is too similar in terms of cosine distance to existing keyphrases, you should probably not add it to the list.
You may also want to consider the use of Term Proximity with Term Frequency-Inverse Document Frequency or even Dimension Independent Matrix Square using MapReduce