When being trained on API documentation and tutorials, Word2Vec
produces vector representations to estimate the relevance between
texts and API elements. However, existing Word2Vec-based approaches to
measure document similarities aggregate Word2Vec vectors of individual
words or APIs to build the representation of a document as if the
words are independent. Thus, the semantics of API descriptions or code
fragments are not well represented. In this work, we conjecture that
we need a new model that fits with API documentation better than
Word2Vec. We present D2Vec, a neural network model that considers two
complementary contexts to better capture the semantics of API
documentation. First, we connect the global context of the current API
topic under description to all the text phrases within the description
of that API. Second, the local orders of words and API elements in the
text phrases are maintained in computing the vector representations
for the APIs. We conducted an experiment to verify two intrinsic
properties of D2Vec's vectors: 1) similar words and relevant API
elements are projected into nearby locations; and 2) some vector
operations carry semantics. We demonstrate the usefulness and good
performance of D2Vec in three applications: API code search
(text-to-code retrieval), API tutorial fragment search (code-to-text
retrieval), and mining API mappings between software libraries
(code-to-code retrieval). Finally, we provide actionable insights and
implications for researchers in using our model in other
applications with other types of documents.