This provides embeddings that follow graph relationships to improve correlations between entities when looking through text. For example, if you use word2vec you will see that terms in the same context appear in similarly trained vectors...but those vectors are lacking formal relationships and just go off where they are likely to appear in text. A good use case for this might be Entity Linking / Disambiguation, when performing Named Entity Recognition on unstructured text...such as knowing that William Shatner and Bill Shatner are the same person, and 'She drank a Manhattan' is referring to the cocktail and not the city.
BTW, part 2: I know that Facebook and Google get a lot of criticism over privacy issues but the flip side of the privacy coin is how generous they are in sharing results, pre-trained models, etc. (even given that publishing great results, data, and code is a great way to attract more potential employees).
